ArticlePDF Available

Abstract and Figures

In recent years, image databases are growing at exponential rates, making their management, indexing, and retrieval, very challenging. Typical image retrieval systems rely on sample images as queries. However, in the absence of sample query images, hand-drawn sketches are also used. The recent adoption of touch screen input devices makes it very convenient to quickly draw shaded sketches of objects to be used for querying image databases. This paper presents a mechanism to provide access to visual information based on users’ hand-drawn partially colored sketches using touch screen devices. A key challenge for sketch-based image retrieval systems is to cope with the inherent ambiguity in sketches due to the lack of colors, textures, shading, and drawing imperfections. To cope with these issues, we propose to fine-tune a deep convolutional neural network (CNN) using augmented dataset to extract features from partially colored hand-drawn sketches for query specification in a sketch-based image retrieval framework. The large augmented dataset contains natural images, edge maps, hand-drawn sketches, de-colorized, and de-texturized images which allow CNN to effectively model visual contents presented to it in a variety of forms. The deep features extracted from CNN allow retrieval of images using both sketches and full color images as queries. We also evaluated the role of partial coloring or shading in sketches to improve the retrieval performance. The proposed method is tested on two large datasets for sketch recognition and sketch-based image retrieval and achieved better classification and retrieval performance than many existing methods.
Content may be subject to copyright.
Data augmentation-assisted deep learning of
hand-drawn partially colored sketches for
visual search
Jamil Ahmad, Khan Muhammad, Sung Wook Baik*
Department of Software, College of Software and Convergence Technology, Sejong University, Seoul,
Republic of Korea
In recent years, image databases are growing at exponential rates, making their manage-
ment, indexing, and retrieval, very challenging. Typical image retrieval systems rely on sam-
ple images as queries. However, in the absence of sample query images, hand-drawn
sketches are also used. The recent adoption of touch screen input devices makes it very
convenient to quickly draw shaded sketches of objects to be used for querying image data-
bases. This paper presents a mechanism to provide access to visual information based on
users’ hand-drawn partially colored sketches using touch screen devices. A key challenge
for sketch-based image retrieval systems is to cope with the inherent ambiguity in sketches
due to the lack of colors, textures, shading, and drawing imperfections. To cope with these
issues, we propose to fine-tune a deep convolutional neural network (CNN) using aug-
mented dataset to extract features from partially colored hand-drawn sketches for query
specification in a sketch-based image retrieval framework. The large augmented dataset
contains natural images, edge maps, hand-drawn sketches, de-colorized, and de-texturized
images which allow CNN to effectively model visual contents presented to it in a variety of
forms. The deep features extracted from CNN allow retrieval of images using both sketches
and full color images as queries. We also evaluated the role of partial coloring or shading in
sketches to improve the retrieval performance. The proposed method is tested on two large
datasets for sketch recognition and sketch-based image retrieval and achieved better classi-
fication and retrieval performance than many existing methods.
With the widespread use and adaptation of portable smart devices like phones and tablets in
our day-to-day computing activities, sketch-based image retrieval (SBIR) has shown promising
potential as an intuitive means to retrieve multimedia contents. The touch screen interface of
these devices allow users to quickly and conveniently draw rough sketches of objects or scenes
with their fingers and retrieve similar images from the collection of images contained in the
repositories [1]. Hand drawn sketches are abstract representations of objects and scenes with
PLOS ONE | August 31, 2017 1 / 19
Citation: Ahmad J, Muhammad K, Baik SW (2017)
Data augmentation-assisted deep learning of hand-
drawn partially colored sketches for visual search.
PLoS ONE 12(8): e0183838.
Editor: Zhihan Lv, University College London,
Received: May 8, 2017
Accepted: August 11, 2017
Published: August 31, 2017
Copyright: ©2017 Ahmad et al. This is an open
access article distributed under the terms of the
Creative Commons Attribution License, which
permits unrestricted use, distribution, and
reproduction in any medium, provided the original
author and source are credited.
Data Availability Statement: The data used in this
study was not collected by the authors. However,
both datasets used in this research can be obtained
from the respective institutions. Visit http://www. for multi-view
objects dataset, and
de/eitz/projects/classifysketch/ to obtain sketches
dataset used in the paper.
Funding: This work was supported by the
National Research Foundation of Korea (NRF)
grant funded by the Korea government (MSIP)
(No.2016R1A2B4011712). The funder had no
imperfections in contours and noise. They differ greatly from black and white or full-color
images, and poses several challenges in distinctive and robust representation for sketch based
image retrieval applications [2]. In case the features of hand-drawn sketches are extracted and
represented appropriately, they can serve as an effective means to specify queries when exam-
ple images are unavailable.
Hand-drawn sketches are merely rough descriptions of scenes and objects and do not need
to be artistic [3,4]. They are mainly composed of simple lines and strokes without any fill col-
ors or details. These contours are considered as highly informative according to human per-
spective, and usually suffice for recognition by humans. In traditional sketch based retrieval
systems, users need to fill sketches with colors to make them visually similar to full-color
images. Though such techniques were considered as cumbersome for the users, the modern
touch screen interfaces can make the process far more convenient than the traditional key-
board and mouse interfaces of the past.
Image retrieval systems must understand users’ intent while processing their queries. This
is a difficult task which becomes even more severe in case of sketch-based queries due to inher-
ent ambiguity caused by the absence of semantic information, textures, colors, and luminance.
This ambiguity has been resolved previously by posing SBIR as model fitting approach
attempted to align sketches with image data. However, such approaches carried with it huge
computational costs. Other approaches attempted to extract local or global features from
sketches and compared them with the features extracted from the edge maps of database
images. In these approaches, content matching of images and sketches is accomplished using
contour matching. In majority of these systems, the authors have used hand-engineered fea-
tures like the variants of histogram of oriented gradients (HoG) [5], bag-of-visual words
(BoVW) [6,7], and various local and global feature similarity methods. However, both local
and global matching approaches have their shortcomings. For instance, the global contour
matching approaches have to take into account the imprecise nature of hand-drawn sketches,
thereby requiring some degree of tolerance. This approach in matching images with sketches
often does not reflect content similarity. Though this problem has been reduced using local
approaches, they are computationally very expensive. Several researchers attempted to address
this issue by introducing efficient methods for reducing computational cost by sacrificing
retrieval performance such as Wang et al. [2] who introduced an edgel index structure to effi-
ciently solve sketch retrieval problem. However, their method heavily relied on local features,
and the matching process was not very robust. Qian et al. [8] proposed a re-ranking and rele-
vance feedback approach to address this issue which attempts to refine search results using
relevance feedback mechanisms. Although such methods improve retrieval results based on
local features, their retrieval performance still depends on the hand-engineered features which
inherently lack capability to describe high level semantics in images.
To overcome these issues posed by hand-crafted local features based matching schemes in
image retrieval systems, researchers have also used the recent powerful deep learning based
approaches which are capable of modeling high level characteristics in images. The major
advantage of these methods is that they can automatically learn features without requiring us
to design algorithms for them. The recent advancements in image recognition due to these
methods have motivated researchers to design powerful models to perform a variety of tasks.
The authors in [9] proposed a sketch based image retrieval method using Siamese convolu-
tional neural network (CNN). Their main idea was to derive similar features for image-sketch
pairs that are marked relevant and derive dissimilar features for irrelevant ones. It was
achieved by tuning two identical CNNs linked by one loss function. One of the CNN was
tuned on the edge maps derived from full color images and the other on corresponding
sketches. The joint output generated by the two linked models correspond to the degree of
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 2 / 19
role in study design, data collection and analysis,
decision to publish, or preparation of the
Competing interests: The authors have declared
that no competing interests exist.
similarity between the two inputs. This way, images were matched with their corresponding
sketches during retrieval phase by propagating sketch through one CNN and the image
through the other. Though CNNs are known to be capable of learning high level representa-
tions in images and even edge maps, ignoring the color and texture aspects of images affect the
overall representation process. Instead of eliminating essential aspect of visual media, i.e. color
and texture, from the image matching process, we propose to optimize the inputs to allow
learning of better representations with discriminative and deep CNN architectures. In this
regard, we experimented with different data augmentation methods to allow effective repre-
sentations of images that will also facilitate accurate matching with partially colored sketches.
Though in the past, fully colored sketches were regarded as burdensome for the users, the cur-
rent work targets portable smart devices where users can easily sketch and apply partial colors
to various regions using onscreen tools on the touch screen devices. The objective of our work
is to assess the suitability of deep CNNs for representing various facets of images including
edge maps, de-texturized, edge enhanced, and de-colorized versions, so that optimal sketch
based image retrieval system can be designed. We will also attempt to assess how the discrimi-
native capabilities offered by the powerful deep CNNs are enhanced when these different rep-
resentations of the same images are presented to them. Major contributions in this work are as
1. Assess the representation capability of deep CNNs for hand-drawn sketches
2. Attempt to enhance sketch recognition using fine-tuning with augmented data
3. Evaluate the effects of data augmentation for sketch recognition and SBIR
4. Propose an optimal method for drawing partially colored sketches on portable smart
5. Determine optimal features from the fine-tuned CNN for sketch based retrieval
The rest of the paper is organized as: Section 2 presents a brief survey of state-of-the-art
SBIR methods based on traditional hand-engineered features and deep features. The proposed
method is illustrated in Section 3 and evaluated on two large datasets in Section 4. The paper
concludes in Section 5 with strengths and weaknesses of the proposed method along with
future directions.
Related work
With the popularity of portable touch screen devices, people prefer to draw or write on touch
screens instead of using pen. Touch screen devices make it very convenient to draw sketches
and transform them into colorful drawings very quickly using onscreen controls. In the con-
text of visual search in educational environments, sketch based image retrieval can make it
more convenient to specify the query instead of textual query or looking for a sample image.
Users can quickly draw a rough sketch of what they need and the retrieval engine will attempt
to find relevant images or sketches from the dataset. Previous works on SBIR can be grouped
into two categories based on the type of features they used to represent sketches.
2.1 Traditional approaches
A general workflow of a traditional SBIR system works by extracting edges from natural
images in order to make them look like sketches and then extract hand-engineered features
from the edge maps of images. The features of hand-drawn sketches are then matched with the
features of edge maps to determine their similarity. Such methods are generally categorized
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 3 / 19
into local and global approaches depending on how these features are extracted. In [10],
authors extracted holistic features from sketches using edge pixels similar to shape context
representation. Shao et al. [11] used similar features of sampled strokes to account for tolerate
differences between sketches. Similarly, Cao et al. [12] developed an edge descriptor for facili-
tating sketch based image search. The main problem with these global representation schemes
is that they are less effective in matching complex sketches. On the other hand local methods
are more robust in representation. Elitz et al. [7] leveraged scale invariant features transform
(SIFT) [13] to formulate bag-of-visual-words (BoVW) for SBIR. A similar approach based on
the BoVW framework using histograms of gradients was presented for SBIR by Hu et al. in
[5]. Both of these methods used k-means to build their corresponding codebooks. Xiao et al.
[6] develop a method to extract shape words from sketches, followed by matching through
Chamfer matching technique to perform shape matching. Shape words is a small segment of
the sketch containing a group of connected edge pixels forming line segments and arcs. Each
shape word have their own properties like direction, location, and size. Zhang et al. [14] fur-
ther improved the shape words method by first discovering discriminative patches for various
sketch categories. The shape patches are extracted from multiple scales, followed by construc-
tion of pyramid histogram. The discovery of discriminative patches is accomplished through
an iterative procedure involving discriminative ranking and cluster merging. Major problems
with these techniques is inherent complexity in matching boundaries of real images to roughly
drawn sketches due to ambiguity, and imperfection. Furthermore, the semantic gap in hand-
crafted image features causes SBIR method to significantly underperform in large datasets.
2.2 Deep learning based approaches
Deep CNNs have exhibited strong performance in a variety of computer vision tasks including
image retrieval [15,16] and classification [17]. These methods have significantly outperformed
traditional methods in so many other applications also. CNNs have the capability to automati-
cally learn important features for a particular classification problem directly from the raw data
(i.e. images). CNNs consist of several layers where each layer learns some characteristic of the
data that can be used to perform the intended classification. Layers closer to the input learn
low-level generic features, whereas higher layers in the network learn more complex features
of the data, describing semantics and are considered higher level features. Babenko et al. [15]
recently investigated features from the various layers of a trained CNN model for image
retrieval. They showed that features extracted by a CNN (i.e. neural codes) are more discrimi-
native and robust than the traditional hand-crafted features. To accomplish SBIR, Qi et al. [9]
trained a Siamese CNN to map hand-drawn sketches to the edge maps of their corresponding
images. Their framework consisted of two identical CNNs whose loss function was linked
together. The sketch and edge map of the relevant image were forward propagated through the
corresponding models, which attempted to decrease the feature distance between relevant
pairs and increased the differences between irrelevant pairs. Their CNN consisted to three
convolutional layers, each followed by a max pooling layers, and one fully connected layer.
The output of the fully connected layer was input to the Softmax classifier. They showed supe-
rior retrieval performance than several state-of-the-art methods. However, they used a rela-
tively simpler model and ignored color and texture features of the images while performing
image matching. Wang et al. [18] presented a technique to train a CNN by mixing images as
well as their edge maps or sketches to construct the training dataset. This enlarged augmented
dataset consisting of both natural images as well as their sketches was used to train the CNN.
The network they used consisted of five convolutional layers and three fully connected layers.
During the training phase, they presented the network with 18 rotated versions of the sketch/
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 4 / 19
edge map to further enhance discriminative ability of the network. During test phase, they cre-
ated the 18 rotations of the query sketch and predicted the label by averaging output of the
Softmax layer.
Deep CNNs are powerful architectures capable of yielding state-of-the-art performance in a
variety of tasks. Their performance is limited by the availability of data which is usually solved
with data augmentation techniques. The majority of these techniques used for SBIR either
ignored color and texture features while representing images to be matched with simple
sketches, or used data augmentation on a relatively smaller scale. However, we believe, that the
touch screen devices make it far more convenient to draw partially colored sketches due to its
ease of use. In such a setting, it can be more beneficial to use features of the full color images
for searching relevant content to a partially colored sketch without requiring us to perform rel-
evance feedback from the users. The efforts that we demand from users in refining the search
results can be requested before entering the query. This can make the whole framework effi-
cient and convenient.
Visual search using partially colored sketches
This section presents the schematics of proposed framework including data augmentation,
architecture of the deep CNN and its training, features extraction for sketch representation,
and their retrieval. An abstract representation of the proposed framework is provided in Fig 1.
3.1 Data augmentation
Learning effectiveness of the deep CNNs are known to depend on the availability of sufficiently
large training data. Data augmentation is an effective method to expand the training data by
applying transformations and deformations to the labeled data, resulting in new samples as
additional training data. A key attribute of the data augmentation is that the labels remain
Fig 1. Schematic diagram of the proposed framework for SBIR.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 5 / 19
unchanged after applying those transformations. Generally, data augmentation involves crops,
rotations, translations, scaling, and mirroring, etc. of the labeled samples. It has been shown
that augmenting data during training phase improves the discriminative and generalization
ability of the model [17]. In the context of SBIR, data augmentation has been used by Wang
et al. [18] who expanded their training data by mixing sketches with real images. It allowed the
CNN to learn features of the sketches in addition to features of the full image. We propose to
use a more advanced method to augment training data by applying more transformations
aimed at allowing CNN to robustly recognize partially colored sketches. Training data is aug-
mented by mixing color images with salient edge maps, de-texturized, and de-colorized images
obtained through anisotropic diffusion as shown in Fig 2. De-colorized and de-texturized ver-
sions of the images will allow them to be matched with partially shaded sketches. Similarly, the
edge maps, hand-drawn sketches, and full color images will enable the CNN to effectively com-
pare partially shaded sketches with full color images. The addition of these varying versions of
images will enable CNN to model discriminative characteristics pertaining to these variety of
representations. Furthermore, it will enable users to query the database using both natural
images and sketches. We believe that training CNN with the augmented data will improve its
generalization on unseen samples.
The decolorized images are obtained by transforming full color images into grayscale.
Though this transformation can be obtained using a variety of methods, we opted to use the
weighted conversion from RGB to grayscale using Eq (1).
IGray ¼0:299 IRþ0:589 IGþ0:112 IBð1Þ
Fig 2. Data augmentation using semantic-preserving transformation for SBIR.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 6 / 19
where I
, I
, and I
are the red, green, and blue color channels respectively. The decolorized
images tend to serve the purpose of representing shaded versions of images to the CNN during
training. In a similar manner, the de-texturized images were formed by smoothing out fine
textural content using anisotropic diffusion approach [19]. Salient edges are most likely con-
tained in the hand-drawn sketch of any object. For allowing the CNN to model salient edges in
images, we presented to it, the edge enhanced versions of images as well. These images were
obtained by enhancing the salient edges using unsharp masking where a smoothed version of
the image is subtracted from the original image to obtain the unsharp mask. This mask is then
added to the original image to generate the edge enhanced image. We used Gaussian smooth-
ing to generate the unsharp mask as follows.
IEE ¼Iðx;yÞ þ Iðx;yÞ  1
  ð2Þ
where I is the input images, I
is the edge enhanced image, () is the convolution operation, σ
is the standard deviation of the filter and was set to 0.5, x and y are the spatial coordinates of
the image.
Four geometric transformations including two flips and two rotations were obtained and
added to the augmented dataset to allow for a certain degree of transformation invariance. In
the final dataset, each image had 8 other versions which sufficiently enlarged the dataset.
3.2 Deep convolutional neural network
Convolutional neural networks have emerged as powerful hierarchical architectures, capable
of learning features from data automatically. They have been applied to a wide variety of
applications in computer vision [20,21], natural language understanding [22], speech recogni-
tion [23], neuronal signal understanding [24], and drug discovery [25]. Their application to a
field is merely limited by availability of data and its representation to these architectures for
processing. A typical CNN is composed of a variety of data processing layers arranged in the
form of a hierarchy, where the output of a layer becomes the input of the succeeding layer. A
majority of these layers are convolutional layers which act as receptive fields for the visual data
being processed. In each convolutional layer, a set of learned kernels are applied on the entire
image to detect patterns at different spatial locations and generate feature maps. Pooling layers
are often used after convolutional layers, which attempt to extract the most meaningful infor-
mation from the set of feature maps. A common pooling strategy is to apply max pooling in
which maximum activations in small neighborhoods of the image are gathered. Consequently,
it reduces the dimensions of the feature maps based on the size of the local neighborhood con-
sidered for pooling. Stacks of convolutional and pooling layers are followed by fully connected
layers which model higher level abstractions. In such a hierarchical setting of layers, as we go
higher in the hierarchy, more abstract and semantically meaningful associations among the
data are modeled.
The CNN model we used for our experiments (shown in Fig 3) was trained by the visual
geometry group (VGG) of the University of Oxford [26]. The model receives input of size 224
x 224 x 3. It has increased depth (19 layers) and used smaller convolutional kernels throughout
the entire network (3 x 3 stride 1). It also used uniform pooling operations (2x2 stride 2) after
each stack of convolutional layers as shown in Fig 3. The first two stacks had two convolutional
layers each with 64, and 128 kernels, respectively. Two convolutional layers stacked together
effectively constitute a receptive field of 5x5. The remaining three stacks consisted of four con-
volutional layers having 256, 512, and 512 kernels, respectively. To allow the extended depth,
the input image was padded to preserve its size before each convolution operation. The two
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 7 / 19
fully connected layers had 4096 neurons each. The last fully connected layer was modified
according to our dataset and was set to 250 neurons, where each neuron correspond to one of
the 250 classes of sketches and associated images.
3.3 Training CNN with augmented data
Training of a CNN is accomplished by tuning the various parameters and biases in all the lay-
ers of the model according to the input data and classification problem. It involves two stages
namely forward and backward propagation phases. During the forward propagation phase,
input images are forward propagated through the network with existing parameters. The loss
cost is computed using the differences in predicted and ground truth labels. During the back-
ward propagation phase, gradients of each parameter are computed using chain rules in order
to adjust the parameters (weights and biases) and reduce error. These two phases are per-
formed many times and the parameters are adjusted until the loss cost has been sufficiently
reduced. We trained several models with our data and evaluated their performance for sketch
classification and SBIR. Individual models similar to the architectures of AlexNet [17] and
VGG-19 [26] were trained on our dataset. The models trained from scratch were able to obtain
classification accuracy of 64% and 68%, respectively, which was slightly below the state-of-the-
art. In order to improve the accuracy, we used the transfer learning approach where a pre-
trained model is fine-tuned on the new dataset [27]. This way the classification problem is
solved more effectively, thereby increasing accuracy. The final model (shown in Fig 3) had 166
million parameters. It was obtained by fine-tuning a pre-trained model (ImageNet dataset) on
the augmented dataset with 250 classes. The Softmax layer of this model outputs predictions
for all the classes.
Fig 3. Architecture of the deep CNN for features extraction.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 8 / 19
In recent studies, it has been shown that transfer learning approach can enhance classifica-
tion accuracy on new datasets [27]. During this approach, the classification function of pre-
trained CNN model is replaced with a new classification function, and optimized to reduce
classification error in the new domain. The learning rate is usually set very low (usually one-
tenth of the original learning rate) so that most of the parameters and weights of the pre-
trained model are only slightly modified. Consequently, the previous knowledge of the model
is used to solve the new problem more efficiently. We evaluated transfer learning using the
augmented dataset and consequently, a 12–15% improvement was noticed in the classification
accuracy. This improvement is due to the fact that the pre-trained model has been trained on a
very large dataset (ImageNet [28]) where it has learned very fine and highly discriminative fea-
tures. Reusing these features significantly improves retrieval accuracy.
3.4 Sketch representation with deep features
The hierarchical nature of the deep CNN allows it to learn multiple levels of features from the
training data. The lower layers learns relatively lower level features corresponding to edges,
curves, and color blobs. Subsequent layers learn higher level features and contain more
semantic features of the visual contents. Neuronal activations at various layers of the network
correspond to the intermediate representation of the image. Each of these intermediate repre-
sentations can be used to represent images for the task of classification or retrieval. However, it
has been noticed that the higher layers in the network learn more discriminative and domain
specific features [15]. Therefore they perform better than the lower layer features. We evalu-
ated features extracted from the last three fully connected layers (FC6, FC7, and FC8) and
found that the last fully connected layer (FC8) consisting of 250 neurons was the most suitable
for this task. Features from this layer are discriminative and yields lower dimensional features
which are favored in retrieval applications. Fig 4 shows sample sketches from the sketches
dataset. It can be seen that there exist a great degree of intra-class variations (Fig 4a) as well as
inter-class similarities (Fig 4b) in hand-drawn sketches which make their classification a very
challenging task. Fig 4a shows five different sketches of planes, bicycles, and laptops. Sketches
of computer mouse, guitar, giraffe, and chair, shown in Fig 4b exhibits inter-class similarities
among the hand-drawn sketches. Features extracted for some of the sketches, shown in Fig 5
reveal that similar features are extracted from sketches belonging to similar classes, despite the
differences in their visual appearances. It shows the discriminative capability of the proposed
model which is key to improved retrieval performance.
3.5 Sketch-based image retrieval with deep features
Features extracted from the last fully connected layer of the model are used to index sketches
as well as color images. When a query is submitted to the SBIR system, the same model is used
to extract features from the query sketch and compared with all the images in the dataset. The
comparison is performed by computing Euclidean distance between the query sketch and
database image features. Lower the score, greater will be the similarity and vice versa. The data-
base images are ranked according to this score in ascending order, where lower score images
are retrieved at higher ranks. In the current work, we developed a software for smart devices
including tablets and phones, which allow users to create sketches and partially color them
before submitting them as queries to the retrieval system. The application uses deep CNN
model trained using the Caffe framework [29] on the device. Though, it is relatively slower in
execution due to the computational limitations of the portable devices, the performance can be
significantly improved if cloud based service is used to perform the compute-intensive task of
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 9 / 19
features extraction and matching. Retrieval performance of the proposed framework is pre-
sented and discussed in the subsequent section.
Experiments and results
Sketch based image retrieval has been investigated for quite a long time. However, only limited
works has been seen using deep CNN for features extraction. Two of the most relevant works
to the proposed method are [18] and [9], who used convolutional neural networks to represent
sketches or match sketches with images. We provide a comparison of performance with these
methods and show that our method is better than both of them in terms of effectiveness and
Fig 4. Challenges in simple sketch representation (a) visual differences in same class objects (b)
inter-class similarities in sketches.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 10 / 19
4.1 Datasets
TU Berlin sketches dataset [30]. This dataset is composed of 20,000 hand-drawn sketches
made by non-experts. These sketches belong to 250 different categories, where each category
has 80 sketches. The size of each image is 1111 x 1111. Seventy five percent of the dataset was
used for training and fine-tuning the models, and the remaining data was used for testing. The
test set was used as query images for searching relevant images in full color image datasets.
Color images dataset. To assess the capability of deep CNN features in retrieving color
images in response to partially colored sketches, we collected more than 35,000 color images
from various datasets, corresponding to the 250 categories of TU Berlin sketches dataset.
These images were gathered from Corel-10k dataset [31], Multi-view objects dataset [32], and
Caltech256 [33].
4.2 Experiments design
We designed several experiments to evaluate performance of the proposed method on sketch
classification and SBIR for partially colored sketches. We are tested the representation capabil-
ity of CNNs for sketches with or without shading or colors. For training or fine-tuning CNNs,
the training set and test sets used had no overlap in order to allow fair comparison. Further-
more, we evaluated the effectiveness of inclusion of color into SBIR for improved performance
using deep features. CNN model training was accomplished on a PC running Ubuntu operat-
ing system, equipped with 64 GB RAM, Intel Core i5 CPU, and NVidia GeForce GTX TITAN
X (Pascal) with 12 GB onboard memory, with Caffe deep learning framework [29]. For evalu-
ating performance in sketch classification and retrieval based on deep features, MATLAB
2015a [34] was used. Further discussion on individual experiments and results is given in the
following sections.
A. Sketch recognition. The test dataset is taken from the largest publicly available hand-
drawn sketches dataset with 20,000 sketches organized into 250 categories. Twenty five percent
of this dataset (5000 sketches) were combined with sketches collected from the internet to test
the performance in sketch classification. Two separate experiments were performed using the
Fig 5. Sketch representation with deep features.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 11 / 19
selected model. During the first experiment, the model was trained using our augmented data-
set for 30 epochs. Classification results for the sketch dataset with this model are provided in
Fig 6. In more than 50% of the categories, the classification accuracy is above 70%. In only 20
categories, the accuracy is below 40%. In the second experiment, we used transfer learning
approach to fine-tune the same model on our dataset. Experimental results shown in Fig 7
exhibit the improvement in terms of classification accuracy. Only 8 sketch categories are
Fig 6. Sketch classification performance (a) without fine-tuning (b) with fine-tuned model.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 12 / 19
classified with less than 40% accuracy. Furthermore, recognition performance for most of the
categories is significantly improved raising the overall accuracy from 68% to 79%.
B. Sketch-based retrieval. In this experiment, we extracted features from both sketches
and full color images using the model trained on augmented dataset. Images were indexed
using these features. Retrieval performance of the proposed method is evaluated on a variety of
sketches and edge maps. Initially we assessed the representation performance of our model for
color-less sketches. In this experiment, we extracted features from the edge maps of natural
images using the model. Then, randomly chosen images were used as queries to retrieve simi-
lar images from the dataset. During the experiments, relevant images were retrieved from the
dataset in most cases, even if there was no color information involved during the features
extraction phase. Still, the features were discriminative enough to retrieve visually similar
images. It is interesting to note that the retrieved images had very similar edge maps, which
lead to their retrieval at top ranks. In some of the cases, the SBIR system failed to retrieve rele-
vant images, but when they were partially colored, their retrieval performance improved dra-
matically (shown in Fig 7). It showed that the introduction of colors significantly improves
performance even if they are only partially applied. It also corresponds to the ability of model-
ing colors by the deep CNN. In order to take advantage of the modeling capabilities of CNNs,
we propose to use partially colored sketches instead of simple strokes.
C. Effect of color on sketch-to-image retrieval. CNNs are powerful architectures capable
of modeling visual contents including colors, textures, and shapes, along with their spatial fea-
tures which lead of their semantic interpretation to a certain degree. However, sketches usually
lack colors or textures which limits the discriminative power of CNNs. It has been proved in
the past that color is a powerful descriptor [3538]. In this experiment, we study the effects of
colors on retrieval performance in the proposed framework. Several experiments were con-
ducted with colorless sketches as well as their partially colored or shaded versions. Though in
the past, coloring sketches was considered burdensome for the users, the convenience pro-
vided by the touch screen devices make it relatively convenient for them to sketch and apply
colors to it. During the experiments, we noticed that even a single stroke of shade or color on
Fig 7. Effect of color/shade on retrieval performance.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 13 / 19
the sketch improved retrieval performance significantly as can be seen in Fig 7. In the first
image, there is no shading or colors on the sketch. The ten images shown on the right are the
top-10 retrieved images. Only 3 relevant images have been retrieved out of 10 at ranks 2, 3,
and 7. When a single stroke was applied to the sketch the number of relevant images increased
to 5, retrieved at ranks 1, 3, 4, 8, and 10. Adding a few more strokes increased number of
relevant images to 6, and further addition of the red color stroke increased the number of rele-
vant images to 10. This experiment showed that addition of colors to sketches significantly
improves their representation by deep CNNs which eventually leads to improved retrieval per-
formance. Quantitative assessment of partial shading has also been carried out. Experimental
results presented in Fig 8 report the retrieval accuracy for top 25 retrievals. Results reveal that
adding only 5% shading in any sketch improves the retrieval accuracy by 12%. Similarly,
increasing the amount of shading to 20% increases the retrieval accuracy to 73.4% and a 30%
shade yields more than 78% retrieval accuracy. These results show that partial shading or col-
oring significantly improves image retrieval performance using the proposed approach.
D. Retrieval performance for deep features extracted from various layers. CNNs learn
multiple layers of features from the training data automatically. Neuronal activation from each
of these layers can be used to represent images for image retrieval. However, the retrieval per-
formance of the last fully connected layers are shown to outperform the early convolutional
Fig 8. Relationship between partial shading and retrieval performance.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 14 / 19
and pooling layers. Therefore, in this experiment, we evaluated the retrieval performance of
various layers. Features from the last three fully connected layers were extracted to represent
the database images and then retrieved with sketches having only 35–50% color. Table 1 shows
the retrieval performance of various layers for both sketch classification and retrieval perfor-
mance. Features extracted from FC8 showed improved performance than the other two layers.
E. Visual retrieval results for partially colored sketches. Access to visual information
can be made more convenient with the help of SBIR. Users can draw partially colored sketches
of objects they are interested in, and the retrieval system would attempt to retrieve the relevant
images. In this experiment, some hand-drawn sketches were partially colored and submitted
as queries to the proposed SBIR system to retrieve relevant images as shown in Fig 9. For thin-
ner shapes like bicycles, and glasses, there was no need to apply any shading or colors on the
sketch and relevant (visually similar) images were retrieved with high accuracy. However, the
retrieval performance for the rest of the sketches improved significantly when colors or shades
were applied to them. For instance, there are viewpoint changes in laptop, umbrella, and chair,
yet the proposed system was able to retrieve them. Although, some irrelevant images have
been retrieved for laptop, umbrella, chair, and watch, retrieval performance got improved as
more color was added to the sketch.
In addition, we also experimented with sketches and images other than the ones used in the
training or validation. In some cases, there exist less ambiguity between the sketch and corre-
sponding images without any shades or colors such as a tennis racquet, bicycle, and glasses,
etc. But for others, the ambiguity can be significantly reduced by adding some shades. Results
in Fig 10 suggest that the proposed method can perform well with a huge variety of images
other than the ones used during training. The first image in each row is the query sketch and
the remaining are top 10 retrieved images from a large dataset of images. Though some incor-
rect images have been retrieved within the top 10 results, the relevant images have been
retrieved at higher ranks. These results can be further improved if more colors or shades are
added to them. The results validate the effectiveness of proposed approach in real world
Conclusion and future work
In this paper, we present a method for sketch-based image retrieval system which uses partially
colored hand-drawn sketches to allow access to visual data in educational applications. We
slightly modified a deep CNN pre-trained on ImageNet dataset and fine-tuned it on aug-
mented dataset, composed of sketches, color images, edge maps, de-colorized, and de-textur-
ized images. The images belong to 250 categories and consisted of very challenging sketches.
Table 1. Comparison of sketch classification approaches.
Method Classification Accuracy
SIFT-Variant+BOW+SVM [30] 56.0%
Stargraph + KNN [39] 61.5%
MKL [40] 65.8%
SIFT+FV(FMM)+SVM [41] 68.9%
Humans Recognition [30] 73.2%
Sketch-a-Net [42] 74.9%
DeepSketch [18] 77.3%
Proposed Method (FC6-4096) 76.3%
Proposed Method (FC7-4096) 77.6%
Proposed Method (FC8-250) 79.1%
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 15 / 19
The model’s capabilities were extensively evaluated for representing hand-drawn sketches for
image retrieval applications. The main aim was to allow users to supply hand-drawn partially
color sketches as queries and access full color images from the dataset. We observed that the
model is able to retrieve thin shapes like eyeglasses and bicycles using rough colorless sketches
very effectively. The rest of the objects were retrieved with relatively less accuracy. However, it
Fig 9. Retrieval performance in response to partially colored sketches.
Fig 10. Retrieval performance on other categories.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 16 / 19
was interesting to note that the introduction of colors to sketches significantly improved the
retrieval performance based on the degree of color or shade applied to the sketch. Even a single
stroke of color or shade would improve retrieval performance for almost any sketch, and this
improvement was directly related to the amount of color or shade applied to it.
Traditionally, it was believed that drawing full color sketches in SBIR systems was very diffi-
cult for end users. However, we believe, that the touch screen devices have made it convenient
for users to quickly draw and color sketches using on screen controls and submit their sketches
as queries. Rather than devising relevant feedback strategies to refine search results, it is far
more convenient and efficient to attempt at retrieval with a little bit effort in preparing the
queries. The results show that there is promise in the proposed approach and further improve-
ments can be achieved if more work is done along these lines.
The authors thank Prof Marc Alexa and Prof Ugur Gudukbay for providing the sketches data-
set and multi-view objects dataset. The authors also thank the editor and anonymous review-
ers for their prolific and highly constructive comments which improved our manuscript
Author Contributions
Conceptualization: Jamil Ahmad, Sung Wook Baik.
Data curation: Khan Muhammad.
Funding acquisition: Sung Wook Baik.
Investigation: Jamil Ahmad, Sung Wook Baik.
Methodology: Jamil Ahmad.
Software: Jamil Ahmad.
Supervision: Sung Wook Baik.
Validation: Sung Wook Baik.
Visualization: Khan Muhammad.
Writing – original draft: Jamil Ahmad.
Writing – review & editing: Khan Muhammad, Sung Wook Baik.
1. Datta R, Li J, Wang JZ, editors. Content-based image retrieval: approaches and trends of the new age.
Proceedings of the 7th ACM SIGMM international workshop on Multimedia information retrieval; 2005:
2. Cao Y, Wang C, Zhang L, Zhang L, editors. Edgel index for large-scale sketch-based image search.
Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on; 2011: IEEE.
3. Wang S, Zhang J, Han TX, Miao Z. Sketch-based image retrieval through hypothesis-driven object
boundary selection with hlr descriptor. IEEE Transactions on Multimedia. 2015; 17(7):1045–57.
4. Cao B, Kang Y, Lin S, Luo X, Xu S, Lv Z, et al. A novel 3D model retrieval system based on three-view
sketches. Journal of Intelligent & Fuzzy Systems. 2016; 31(5):2675–83.
5. Hu R, Barnard M, Collomosse J, editors. Gradient field descriptor for sketch based retrieval and locali-
zation. 2010 IEEE International Conference on Image Processing; 2010: IEEE.
6. Xiao C, Wang C, Zhang L, Zhang L, editors. Sketch-based image retrieval via shape words. Proceed-
ings of the 5th ACM on International Conference on Multimedia Retrieval; 2015: ACM.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 17 / 19
7. Eitz M, Hildebrand K, Boubekeur T, Alexa M. Sketch-based image retrieval: Benchmark and bag-of-fea-
tures descriptors. IEEE Transactions on Visualization and Computer Graphics. 2011; 17(11):1624–36. PMID: 21173450
8. Qian X, Tan X, Zhang Y, Hong R, Wang M. Enhancing Sketch-Based Image Retrieval by Re-Ranking
and Relevance Feedback. IEEE Transactions on Image Processing. 2016; 25(1):195–208. https://doi.
org/10.1109/TIP.2015.2497145 PMID: 26552088
9. Qi Y, Song Y-Z, Zhang H, Liu J, editors. Sketch-based image retrieval via Siamese convolutional neural
network. Image Processing (ICIP), 2016 IEEE International Conference on; 2016: IEEE.
10. Chalechale A, Naghdy G, Mertins A. Sketch-based image matching using angular partitioning. IEEE
Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans. 2005; 35(1):28–41.
11. Shao T, Xu W, Yin K, Wang J, Zhou K, Guo B, editors. Discriminative Sketch-based 3D Model Retrieval
via Robust Shape Matching. Computer Graphics Forum; 2011: Wiley Online Library.
12. Cao X, Zhang H, Liu S, Guo X, Lin L, editors. Sym-fish: A symmetry-aware flip invariant sketch histo-
gram shape descriptor. Proceedings of the IEEE International Conference on Computer Vision; 2013.
13. Lowe DG. Distinctive image features from scale-invariant keypoints. International journal of computer
vision. 2004; 60(2):91–110.
14. Zheng Y, Yao H, Zhao S, Wang Y. Discovering discriminative patches for free-hand sketch analysis.
Multimedia Systems. 2016:1–11.
15. Babenko A, Slesarev A, Chigorin A, Lempitsky V. Neural codes for image retrieval. Computer Vision–
ECCV 2014: Springer; 2014. p. 584–99.
16. Cao B, Kang Y, Lin S, Luo X, Xu S, Lv Z. Style-sensitive 3D model retrieval through sketch-based que-
ries. Journal of Intelligent & Fuzzy Systems. 2016; 31(5):2637–44.
17. Krizhevsky A, Sutskever I, Hinton GE, editors. Imagenet classification with deep convolutional neural
networks. Advances in neural information processing systems; 2012: Curran Associates, Inc.
18. Wang X, Duan X, Bai X. Deep sketch feature for cross-domain image retrieval. Neurocomputing. 2016;
19. Perona P, Malik J. Scale-space and edge detection using anisotropic diffusion. IEEE Transactions on
pattern analysis and machine intelligence. 1990; 12(7):629–39.
20. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015; 521(7553):436–44.
21. Ahmad J, Mehmood I, Baik SW. Efficient object-based surveillance image search using spatial pooling
of convolutional features. Journal of Visual Communication and Image Representation. 2017; 45:62–
22. Collobert R, Weston J, editors. A unified architecture for natural language processing: Deep neural net-
works with multitask learning. Proceedings of the 25th international conference on Machine learning;
2008: ACM.
23. Deng L, Li J, Huang J-T, Yao K, Yu D, Seide F, et al., editors. Recent advances in deep learning for
speech research at Microsoft. 2013 IEEE International Conference on Acoustics, Speech and Signal
Processing; 2013: IEEE.
24. Jirayucharoensak S, Pan-Ngum S, Israsena P. EEG-based emotion recognition using deep learning
network with principal component based covariate shift adaptation. The Scientific World Journal.
25. Gawehn E, Hiss JA, Schneider G. Deep learning in drug discovery. Molecular Informatics. 2016; 35
(1):3–14. PMID: 27491648
26. Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv
preprint arXiv:14091556. 2014.
27. Bengio Y. Deep Learning of Representations for Unsupervised and Transfer Learning. ICML Unsuper-
vised and Transfer Learning. 2012;27:17–36.
28. Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L, editors. Imagenet: A large-scale hierarchical image
database. Computer Vision and Pattern Recognition, 2009 CVPR 2009 IEEE Conference on; 2009:
29. Jia Y, Shelhamer E, Donahue J, Karayev S, Long J, Girshick R, et al., editors. Caffe: Convolutional
architecture for fast feature embedding. Proceedings of the 22nd ACM international conference on Mul-
timedia; 2014: ACM.
30. Eitz M, Hays J, Alexa M. How do humans sketch objects? ACM Trans Graph. 2012; 31(4):44:1-:10.
31. Corel Dataset [cited 2017 12-July-2017].
32. C¸alışır F, Baştan M, Ulusoy O
¨, Gu¨du¨kbay U. Mobile multi-view object image search. Multimedia Tools
and Applications. 2015:1–24.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 18 / 19
33. Caltech-256 Object Category Dataset [cited 2016].
34. MathWorks. MATLAB 2015.
35. Ahmad J, Sajjad M, Rho S, Baik SW. Multi-scale local structure patterns histogram for describing visual
contents in social image retrieval systems. Multimedia Tools and Applications. 2016; 75(20):12669–92.
36. Ahmad J, Sajjad M, Mehmood I, Rho S, Baik SW. Saliency-weighted graphs for efficient visual content
description and their applications in real-time image retrieval systems. Journal of Real-Time Image Pro-
cessing. 2016:1–17.
37. Casanova D, Florindo J, Falvo M, Bruno O. Texture analysis using fractal descriptors estimated by the
mutual interference of color channels. Information Sciences. 2016; 346:58–72.
38. Ahmad J, Mehmood I, Rho S, Chilamkurti N, Baik SW. Embedded deep vision in smart cameras for
multi-view objects representation and retrieval. Computers & Electrical Engineering. 2017. https://doi.
39. Li Y, Song Y-Z, Gong S, editors. Sketch Recognition by Ensemble Matching of Structured Features.
BMVC; 2013: Citeseer.
40. Li Y, Hospedales TM, Song Y-Z, Gong S. Free-hand sketch recognition by multi-kernel feature learning.
Computer Vision and Image Understanding. 2015; 137:1–11.
41. Schneider RG, Tuytelaars T. Sketch classification and classification-driven analysis using fisher vec-
tors. ACM Transactions on Graphics (TOG). 2014; 33(6):174.
42. Yu Q, Yang Y, Song Y-Z, Xiang T, Hospedales T. Sketch-a-net that beats humans. arXiv preprint
arXiv:150107873. 2015.
Image retrieval using partially colored sketches
PLOS ONE | August 31, 2017 19 / 19
... Presently, an excessive utilization of advanced technological tools, the one needs to access query sketch more accurately and retrieve its relevant contents to be well-recognized through technologicalbased smart devices. However, to acknowledge the needs of the society and to balance with technological advancement, the researchers have been analyzed various novel tasks regarding sketch recognition [5,6], and sketch-based image retrieval [7,8], in a field of computer vision. The idea behind the sketch classification or recognition is to extract the information from the desired object class of labeled sketch-images among the pre-defined set of object-classes. ...
... Current approaches to object recognition make the necessary use of deep learning and machine learning techniques. However, the most existing work in sketch recognition is based on deep learning approaches using deep convolutional neural networks (DCNNs) and showed an impressive result than handcrafted approaches [6,8,15]. ...
... namely LeNet [22] and AlexNet [16] are used to extract features from sketch images and show improvement in the recognition results [15]. On the other hand, some recent attempts utilized different layers of various CNNs architectures for features extraction for the purpose of sketch classification and retrieval [8,18,23]. ...
Image-based object recognition is a well-studied topic in the field of computer vision. Features extraction for hand-drawn sketch recognition and retrieval become increasingly popular among the computer vision researchers. Increasing use of touchscreens and portable devices raised the challenge for computer vision community to access the sketches more efficiently and effectively. In this article, a novel deep convolutional neural network-based (DCNN) framework for hand-drawn sketch recognition, which is composed of three well-known pre-trained DCNN architectures in the context of transfer learning with global average pooling (GAP) strategy is proposed. First, an augmented-variants of natural images was generated and sum-up with TU-Berlin sketch images to all its corresponding 250 sketch object categories. Second, the features maps were extracted by three asymmetry DCNN architectures namely, Visual Geometric Group Network (VGGNet), Residual Networks (ResNet) and Inception-v3 from input images. Finally, the distinct features maps were concatenated and the features reductions were carried out under GAP layer. The resulting feature vector was fed into the softmax classifier for sketch classification results. The performance of proposed framework is comprehensively evaluated on augmented-variants TU-Berlin sketch dataset for sketch classification and retrieval task. Experimental outcomes reveal that the proposed framework brings substantial improvements over the state-of-the-art methods for sketch classification and retrieval.
... However, in some cases, it corresponds to a final feature representations and then feeds them into the classifier for a final decision. In this vein, utilisation of deep learning approaches to extract deep features from input images [10,11], and/or deep architecture [12], particularly designed for sketch recognition also achieved state-of-the-art performance. ...
... The final outcomes were presented with the label. Jamil et al. [11] demonstrated a fine-tune deep CNN architecture to recognise and retrieved partially colour sketch images. Moreover, in Reference [26], the authors develop SketchNet to learn the common structure between natural images and sketch images automatically. ...
Full-text available
An effective feature representation can boost recognition tasks in the sketch domain. Due to an abstract and diverse structure of the sketch relatively with a natural image, it is complex to generate a discriminative features representation for sketch recognition. Accordingly, this article presents a novel scheme for sketch recognition. It generates a discriminative features representation as a result of integrating asymmetry essential information from deep features. This information is kept as an original feature‐vector space for making a final decision. Specifically, five different well‐known pre‐trained deep convolutional neural networks (DCNNs), namely, AlexNet, VGGNet‐19, Inception V3, Xception, and InceptionResNetV2 are fine‐tuned and utilised for feature extraction. First, the high‐level deep layers of the networks were used to get multi‐features hierarchy from sketch images. Second, an entropy‐based neighbourhood component analysis was employed to optimise the fusion of features in order of rank from multiple different layers of various deep networks. Finally, the ranked features vector space was fed into the support vector machine (SVM) classifier for sketch classification outcomes. The performance of the proposed scheme is evaluated on two different sketch datasets such as TU‐Berlin and Sketchy for classification and retrieval tasks. Experimental outcomes demonstrate that the proposed scheme brings substantial improvement over human recognition accuracy and other state‐of‐the‐art algorithms.
... The challenge with developing machine vision-based classifiers for histopathological or clinical data lies in the limited number of images that are typically available to researchers. 65 While deep learning-based approaches eliminate the need to "hand draw" or manually design the features while improving classifier performance, 66 they often require more than a million images for training. 62 Obtaining and annotating clinical datasets on this scale is often not possible. ...
Tumor grade assessment is critical to the treatment of cancers. A pathologist typically evaluates grade by examining morphologic organization in tissue using hematoxylin and eosin (H&E) stained tissue sections. Fourier transform infrared spectroscopic (FT-IR) imaging provides an alternate view of tissue in which spatially specific molecular information from unstained tissue can be utilized. Here, we examine the potential of IR imaging for grading colon cancer in biopsy samples. We used a 148-patient cohort to develop a deep learning classifier to estimate the tumor grade using IR absorption. We demonstrate that FT-IR imaging can be a viable tool to determine colorectal cancer grades, which we validated on an independent cohort of surgical resections. This work demonstrates that harnessing molecular information from FT-IR imaging and coupling it with morphometry is a potential path to develop clinically relevant grade prediction models.
... Görüntü veri kümeleri üzerinde yapılan çalışmalarda yapay örnek üretimi örneklere çeşitli dönüşümler uygulayarak ya da bir parçasını alarak yapılabilmektedir [6,7]. Metin türündeki veri kümelerinde ise geri çevrim (back translation), kelimeleri eş anlamlılarıyla değiştirme gibi yöntemler kullanılmaktadır [8][9][10]. ...
Conference Paper
Full-text available
Synthetic data generation is one of the methods used in machine learning to increase the performance of algorithms on datasets. However, these methods do not ensure success on each dataset. In this study, it has been investigated that which type of synthetic data generation algorithms are useful in which datasets by examining the effects of SMOTE, Borderline-SMOTE and Random data generation algorithms on 33 datasets. For this, each dataset has been fully balanced as a result of synthetic data generation. In order to evaluate the results, datasets are divided into three groups as balanced, partially balanced-unbalanced and unbalanced in accordance with the unbalance ratio. The datasets formed as a result of the data generation of the algorithms and the original datasets have been trained with an ANN models and their performance has been evaluated on the test set. Experimental results have shown that adding synthetic data to the datasets with the above-mentioned algorithms generally increases the success in balanced and partially balanced-unbalanced datasets, but generally does not work in unbalanced datasets. Borderline-SMOTE, which produces border samples in balanced datasets, and SMOTE in partially balanced-unbalanced datasets have been more successful.
... e scholars applied manual augmentation techniques in conjunction with deep neural network that led to an enhanced achievement. Moreover, the experimenters implemented DA algorithm for hand-drawn dataset and a finetuned deep neural network to extract useful features from the introduced dataset [17]. Recently, the authors of [18] applied the DA Markov chain Monte Carlo (MCMC). ...
Full-text available
In healthcare applications, deep learning is a highly valuable tool. It extracts features from raw data to save time and effort for health practitioners. A deep learning model is capable of learning and extracting the features from raw data by itself without any external intervention. On the other hand, shallow learning feature extraction techniques depend on user experience in selecting a powerful feature extraction algorithm. In this article, we proposed a multistage model that is based on the spectrogram of biosignal. The proposed model provides an appropriate representation of the input raw biosignal that boosts the accuracy of training and testing dataset. In the next stage, smaller datasets are augmented as larger data sets to enhance the accuracy of the classification for biosignal datasets. After that, the augmented dataset is represented in the TensorFlow that provides more services and functionalities, which give more flexibility. The proposed model was compared with different approaches. The results show that the proposed approach is better in terms of testing and training accuracy.
... Augmentation helps avoid overfitting since the model will not see the same sample twice, and it achieves better classification since more samples generally lead to better learning (Ahmad et al., 2017). Fig 11 shows the original (Fig 11-left) and augmented (Fig 11-right) images. ...
Full-text available
Parkinson's Disease (PD) is a chronic, degenerative disorder which leads to a range of motor and cognitive symptoms. PD diagnosis is a challenging task since its symptoms are very similar to other diseases such as normal ageing and essential tremor. Much research has been applied to diagnosing this disease. This project aims to automate the PD diagnosis process using deep learning, Recursive Neural Networks (RNN) and Convolutional Neural Networks (CNN), to differentiate between healthy and PD patients. Besides that, since different datasets may capture different aspects of this disease, this project aims to explore which PD test is more effective in the discrimination process by analysing different imaging and movement datasets (notably cube and spiral pentagon datasets). In addition, this project evaluates which dataset type, imaging or time series, is more effective in diagnosing PD.
... Bhattacharyya (2011) provided a brief overview of additional color image preprocessing techniques. Ahmad et al. (2017) use detexturized, decolorized, edge enhanced, salient edge map based, and flip/rotate images to improve DCNN-based recognition in visual searches. More specialized versions of preprocessing are available if targeted tasks (e.g., in face recognition) can be narrowed down and underlying information can be modeled more precisely (see Zou et al., 2007;Han et al., 2013). ...
Full-text available
Most collaborative tasks require interaction with everyday objects (e.g., utensils while cooking). Thus, robots must perceive everyday objects in an effective and efficient way. This highlights the necessity of understanding environmental factors and their impact on visual perception, such as illumination changes throughout the day on robotic systems in the real world. In object recognition, two of these factors are changes due to illumination of the scene and differences in the sensors capturing it. In this paper, we will present data augmentations for object recognition that enhance a deep learning architecture. We will show how simple linear and non-linear illumination models and feature concatenation can be used to improve deep learning-based approaches. The aim of this work is to allow for more realistic Human-Robot Interaction scenarios with a small amount of training data in combination with incremental interactive object learning. This will benefit the interaction with the robot to maximize object learning for long-term and location-independent learning in unshaped environments. With our model-based analysis, we showed that changes in illumination affect recognition approaches that use Deep Convolutional Neural Network to encode features for object recognition. Using data augmentation, we were able to show that such a system can be modified toward a more robust recognition without retraining the network. Additionally, we have shown that using simple brightness change models can help to improve the recognition across all training set sizes.
... In recent years, with the emergence and development of touch screen devices and convolution neural networks [1] (CNNs), fine-grained sketch-based image retrieval (FG-SBIR) has been widely used. Sketch-based image retrieval [2][3][4][5][6][7][8][9] is an important direction of content-based image retrieval, but compared with some content-based image retrieval [10][11][12][13] that requires sample queries, sketch-based image retrieval can get rid of this shackle. You can get more intuitive and accurate information by just drawing a few strokes based on the impression of the object [14]. ...
Full-text available
This paper focuses on fine-grained image retrieval based on sketches. Sketches capture detailed information, but their highly abstract nature makes visual comparisons with images more difficult. In spite of the fact that the existing models take into account the fine-grained details, they can not accurately highlight the distinctive local features and ignore the correlation between features. To solve this problem, we design a gradually focused bilinear attention model to extract detailed information more effectively. Specifically, the attention model is to accurately focus on representative local positions, and then use the weighted bilinear coding to find more discriminative feature representations. Finally, the global triplet loss function is used to avoid oversampling or undersampling. The experimental results show that the proposed method outperforms the state-of-the-art sketch-based image retrieval methods.
Background: Although rheumatoid arthritis (RA) causes destruction of articular cartilage, early treatment significantly improves symptoms and delays progression. It is important to detect subtle damage for an early diagnosis. Recent software programs are comparable with the conventional human scoring method regarding detectability of the radiographic progression of RA. Thus, automatic and accurate selection of relevant images (e.g. hand images) among radiographic images of various body parts is necessary for serial analysis on a large scale. Objective: In this study we examined whether deep learning can select target images from a large number of stored images retrieved from a picture archiving and communication system (PACS) including miscellaneous body parts of patients. Methods: We selected 1,047 X-ray images including various body parts and divided them into two groups: 841 images for training and 206 images for testing. The training images were augmented and used to train a convolutional neural network (CNN) consisting of 4 convolution layers, 2 pooling layers and 2 fully connected layers. After training, we created software to classify the test images and examined the accuracy. Results: The image extraction accuracy was 0.952 and 0.979 for unilateral hand and both hands, respectively. In addition, all 206 test images were perfectly classified into unilateral hand, both hands, and the others. Conclusions: Deep learning showed promise to enable efficiently automatic selection of target X-ray images of RA patients.
Full-text available
Content based image retrieval systems rely heavily on the set of features extracted from images. Effective image representation emerges as a crucial step in such systems. A key challenge in visual content representation is to reduce the so called ‘semantic gap’. It is the inability of existing methods to describe contents in a human-oriented way. Content representation methods inspired by the human vision system have shown promising results in image retrieval. Considerable work has been carried out during the past two decades for developing methods to extract descriptors inspired by the human vision system and attempt to retrieve visual contents efficiently according to the user needs, thereby reducing the semantic gap. Despite the extensive research being conducted in this area, limitations in current image retrieval systems still exist. This paper presents a descriptor for personalized social image collections which utilizes the local structure patterns in salient edge maps of images at multiple scales. The human visual system at the basic level is sensitive to edges, corners, intersections, and other such intensity variations in images generating local structure patterns. Analyzing these patterns at multiple scales allow the most salient fine-grained and coarse-grained features to be captured. The features are accumulated in a local structure patterns histogram to index images allowing flexible querying of visual contents. The retrieval results show that the proposed descriptor ranks well among similar state-of-the-art methods for large social image collections.
Full-text available
Since the ancient times, free-hand sketch has been widely used as an effective and convenient intermediate means to express human thoughts and highly diverse objects in reality. In recent years, a great quantity of researchers realized the significance of sketch and gradually focused on sketch-related problems, such as sketch-based image retrieval and recognition. Despite so many achievements, very few works concentrate on exploring the intrinsic factors which potentially influence the vivid degree of sketch. In this paper, we propose a weak supervised approach to discover the most discriminative patches for different categories of sketches, which perhaps grasp the key to a good free-hand sketch. In the beginning, we randomly extract tens of thousands of patches at multiple scales. After that, pyramid histogram of oriented gradient is calculated to represent these patches as an effective and uniform feature representation. To find the most discriminative patches for each class of sketches, we design an iterative detection process which combines cluster merging and discriminative ranking. The experimental results on the TU-Berlin sketch benchmark dataset demonstrate the effectiveness of the proposed method, as compared to other available approaches. Moreover, a reasonable analysis and discussion about good and bad sketches is provided based on the visual results.
Active large scale surveillance of indoor and outdoor environments with multiple cameras is becoming an undeniable necessity in today's connected world. Enhanced computational and storage capabilities in smart cameras establish them as promising platforms for implementing intelligent and autonomous surveillance networks. However, poor resolution, limited number of samples per object, and pose variation in multi-view surveillance streams, make the task of efficient image representation highly challenging. To address these issues, we propose an efficient and powerful convolutional neural network (CNN) based framework for features extraction using embedded processing on smart cameras. Efficient, high performance, pre-trained CNNs are separately fine-tuned on persons and vehicles to obtain discriminative, low dimensional features from segmented surveillance objects. Furthermore, multi-view queries of surveillance objects are used to improve retrieval performance. Experiments reveal better efficiency and retrieval performance in different surveillance datasets.
Modern surveillance networks are large collections of computational sensor nodes, where each node can be programmed to capture, prioritize, segment salient objects, and transmit them to central repositories for indexing. Visual data from such networks grow exponentially and present many challenges concerning their transmission, storage, and retrieval. Searching for particular surveillance objects is a common but challenging task. In this paper, we present an efficient features extraction framework which utilizes an optimal subset of kernels from the first layer of a convolutional neural network pre-trained on ImageNet dataset for object-based surveillance image search. The input image is convolved with the set of kernels to generate feature maps, which are aggregated into a single feature map using a novel spatial maximal activator pooling approach. A low-dimensional feature vector is computed to represent surveillance objects. The proposed system provides improvements in both performance and efficiency over other similar approaches for surveillance datasets.
Traditional sketch-based 3D model retrieval methods are content-based, which return the search results by ranking the geometric similarities among a free-hand drawing and 3D model candidates. These conventional methods do not consider personal drawing characteristics and styles (abbreviated as styles), which are obvious and important in user's sketch queries. An ordinary user presumably is not a professional and skillful artist. Therefore, users are likely to introduce personal drawing style in sketching 3D model rather than faithfully render the model according to its geometric perspectives. For amateurs, such personal styles are unintentionally introduced due to their limited sketching capabilities. As determined by a person's sketching habit, personal drawing styles are largely personally consistent and stable. Ignoring such non-trivial personal styles while attempting to reconstruct intended models according to their sketch inputs does not usually produce satisfactory outcomes, in particular, for amateur sketchers. To overcome this problem, we propose a novel style-sensitive 3D model retrieval method based on three-view user sketch inputs. The new method models users' personal sketching styles and constructs joint tensor factorization to improve the retrieval performance.
3D models can be used in 3D printing and many other areas. At present, there are a lot of researches on 3D model retrieval and sketch is considered to be important for 3D model retrieval. In this paper, we develop a new 3D model retrieval prototype system based on style-sensitive 3D model retrieval method and three-view user sketches. We also implement user-friendly graphic interfaces for the 3D model retrieval system. We explore the performance of the system by conducting a series of 3D model retrieval experiments on the Princeton shape benchmark data set. Experimental results show that the new retrieval system can obtain satisfactory retrieval results. And some results of the new method are superior to some content-based 3D model retrieval methods, in terms of both quantitative search performance metrics and qualitatively measured user search experiences.
Deep learning has been proven be very effective for various image recognition tasks, e.g., image classification, semantic segmentation, image retrieval, shape classification etc. However, existing works on deep learning for image recognition mainly focus on either natural image data or binary shape data. In this paper, we show that deep convolutional neural networks (DCNN) is also suitable for cross-domain image recognition, i.e., using sketch as query to retrieve natural images in a large dataset. To solve this kind of cross-domain problem, we propose to train CNN jointly using image data and sketch data in a novel way. The learned deep feature is effective for cross-domain image retrieval - using simple Euclidean distance on the learned feature can significantly outperform the previous state-of-the-arts. In addition, we find that pre-training and a feasible data-argumentation for DCNN can largely surpass human-level performance in the standard sketch classification benchmark.
Conference Paper
The explosive growth of touch screens has provided a good platform for sketch-based image retrieval. However, most previous works focused on low level descriptors of shapes and sketches. In this paper, we try to step forward and propose to leverage shape words descriptor for sketch-based image retrieval. First, the shape words are defined and an efficient algorithm is designed for shape words extraction. Then we generalize the classic Chamfer Matching algorithm to address the shape words matching problem. Finally, a novel inverted index structure is proposed to make shape words representation scalable to large scale image databases. Experimental results show that our method achieves competitive accuracy but requires much less memory, e.g., less than 3% of memory storage of MindFinder. Due to its competitive accuracy and low memory cost, our method can scale up to much larger database.