Improving Invisible Food Texture Detection by using Adaptive Extremal Region Detector in Food Recognition

Full text

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
68
ABSTRACT
The advancement of mobile technology with reasonable cost
has indulge the mobile phone users to photograph foods and
shared in social media. Since that, food recognition has
become emerging research area in image processing and
machine learning. Food recognition provides an automatic
identification of the types of foods from an image. Then,
further analysis in food recognition is performed to
approximate the calories and nutritional information that can
be used for health-care purposes. The interest region-based
detector by using Maximally Stable Extremal Region (MSER)
may provides distinctive interest points by representing the
arbitrary shape of foods through global segmentation
especially the food images with strong mixture of ingredients.
However, the classification performance on food categories
with less diverse texture food images by using MSER are
obviously low compared to the other food categories that have
more noticeable texture. The texture-less food objects were
suffered from small number of extremal regions (ER)
detection beside having low image brightness and small
resolutions. Therefore, this paper proposed an adaptive
interest regions detection by using MSER (aMSER) that
provide a mechanism to choose appropriate MSER parameter
configuration to increase the density of interest points on the
targeted food images. The features are described by using
Speeded-up Robust Feature Transform (SURF) and encoded
by using Bag of Features (BoF) model. The classification is
performed by using Linear Support Vector Machine and yield
84.20% classification rate on UEC100-Food dataset with
competitive number of ER and computation cost.
Key words : Food recognition, MSER, Local features, Bag of
Features
1. INTRODUCTION
There is strong correlation between obesity and overweight
with the occurrence of so-called diet-related chronic diseases
such as diabetes, heart disease, kidney diseases and even
cancers. Dietary assessment is a treatment undertaken by
medical practitioner and dietitian to combat obesity and
overweight problems. However, the traditional dietary
assessment is a tedious process that often lead to inaccuracy in
making evaluation to describe the information about the foods
consumed [1]. An adequate information is compulsory to be
deliberated such as the preparation methods, portion size,
brand, calories and nutritional contents that must be recorded
in daily basis. Furthermore, the traditional dietary assessment
tends to lead under-reporting problem [2].
An automatic dietary assessment via food recognition
algorithm has become active research area under the umbrella
of image processing and machine learning field [3]–[5]. By
using food recognition, the calories estimation of foods can be
calculated precisely. The mobile technology nowadays has
been equipped with good imaging quality and at reasonable
costs have provide the ubiquity way in acquiring images.
Capturing food images have also become a phenomenon with
the popularity of social media network. In fact, the explosive
amount of food images in social media has potential to
provide useful and real information about eating habits and
food preferences in our society that can be benefited by the
food and health-care industry.
Foods have complex appearance as food objects have
non-rigid deformation and high variations that widen the
intra-class variability and narrow the inter-class
inter-similarities [6], [7]. Thus, feature representation method
plays an important role in transforming the raw pixels of food
images into higher semantic of representation. The feature
representation by using local features are the common
practices in food recognition. This is because of the complex
appearance nature of foods can be effectively captured
through the properties of local feature that invariant to
illuminations, rotations, scale and orientation [6], [8] and a
compact and discriminative features can be produced [9].
There are numerous types of local features in the literature. A
research conducted by [10] employed the interest region
Improving Invisible Food Texture Detection by using Adaptive Extremal
Region Detector in Food Recognition
Mohd Norhisham bin Razali 1,2, Noridayu Manshor 1, Norwati Mustapha 1, Razali Yaakob 1, Mohammad
Noorazlan Shah Zainudin3
1 Faculty of Computer Science and Information Technology, Universiti Putra Malaysia, Malaysia
2 Faculty of Computing and Informatics, Universiti Malaysia Sabah, Malaysia
3 Faculty of Electronics and Computer Engineering, Universiti Teknikal Malaysia,
Melaka, Malaysia
ISSN 2278
-
3091
Volume 8, No.1.4, 2019
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse1181.42019.pdf
https://doi.org/10.30534/ijatcse/2019/1181.42019

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
69
detector by using Maximally Stable Extremal Region (MSER)
to detect food interest points, considering MSER as among the
best interest region detector in term of effectiveness and
efficiency [11]. MSER detects a set of connected regions from
an image to define the extremal regions (ER). In food
recognition, MSER has capability to deal with arbitrary shape
of foods and detects the grainy food objects via ER detection
by using global segmentation.
The common problem of any interest points detector such as
DoG and Hessian is its tendency to detect denser features only
on the textured surface [12], [13]. In contrast, low number of
interest points were detected for texture-less foods that affect
its classification performance. MSER encountered the similar
problem as the other interest points detector. In fact, according
to [11], [14], MSER detect even smaller number of interest
points among the interest points detector. Therefore, this
study proposed an adaptive approach for ER detection in
MSER (aMSER) that choose appropriate MSER parameter
configuration to increase the density of ER on the texture-less
food images.
The rest of the paper is organized as follows. In section 2, we
provide related works on food recognition with adaptive
approach. Section 3 describes the aMSER extraction
mechanisms and the MSER parameter configuration. Section
4 and 5 presents the feature representation, dataset and
performance measurement. Section 6 presents the
experimental results and section 7 conclude this paper with
the recommendation for future works.
2. RELATED WORKS
In general, a recognition process is composed by feature
extraction, feature encoding and classification. Each process
has its own components and configurations that impact the
classification performance. For instance, the feature
extraction required a rigorous evaluation to identify the types
of feature, the sampling techniques, descriptor size and so on.
The same case goes to feature encoding and classification
stage that required certain extend of evaluation on the
components used. The initial idea of an adaptation model in
object recognition was exposed by [15] where an adaptive
configuration of components in object recognition need to be
designed to cater the diversity appearance of the objects that
probably required different kind of component or
configuration in order to perform effectively. Due to different
nature of food objects with the other types of objects, food
recognition required different kind of methods adaptation [6].
The use of various type of features are inevitable to cater the
high variability of food objects. However, different foods
might require different features, for instance, the colour
feature provide a better description for ‘Potage’ and the shape
feature may provide better description for ‘hamburger’.
Concerning on this matter, an adaptive feature extraction by
using Multiple Kernel Learning (MKL) have been proposed
[16], [17] to measure the significance of the variety of features
for food objects. The overall classification accuracy reported
is 62.5% with poor recall rate on food category simmered
pork, ginger pork saute, toast, pilaf and egg roll due to less
diverse surface of these food objects.
The research conducted by [13] performed a technical
investigation on the components within BoF model to
determine the optimal sampling technique, descriptor size,
types of local features, clustering for generate visual
dictionary and the classifier. The classification accuracy was
reported as 78%. However, the evaluation of the local features
is only performed within the family of Scale Invariant Feature
Transform (SIFT). SIFT has problem at describing the image
with complex background [18]. As a result, the proposed BoF
model still unable to create a discriminative feature to
distinguish different class of foods.
In summary, the recognition performance on food categories
that consist many texture-less food objects need to be
improved by using suitable features and at reasonable
computation cost in feature detection, feature description and
feature encoding.
3. ADAPTIVE MAXIMALLY STABLE EXTREMAL
REGION (AMSER) EXTRACTION
An adaptive system can be described as the capability of a
system to react in a way according to the responses received
from its surrounding. An adaptive system is incorporated with
MSER (aMSER) in detecting the extremal regions in food
images. By using aMSER, the selection of MSER parameter
configuration can be performed based on the pre-defined
conditions. There are certain conditions of food images that
led to insufficient number of interest points and even worst,
resulted null interest points. The invisible texture such as the
liquid-based foods, small images and low images brightness
contribute to the low number of interest points detection.
Indeed, in any interest points detector including MSER, the
density level of interest points are governed by its parameter
configuration [19]. For this reason, tweaking on MSER
parameter has become necessary. However, we believed that
the food images with denser interest points will probably not
benefit much from the parameter configuration. In fact, the
overwhelming number of interest points will in turn drag the
timeframe for feature detection and feature encoding [20].
Thus, the detection of extremal region (ER) via aMSER will
be executed in more flexible and sensitive based on the food
images condition. Apparently, the aMSER will be expected to
increase the number of ER on the targeted food images by
configuring the intensity threshold (IT) value and maximum
area variation (MAV) value. The following section explain on
the MSER parameter configuration and flowchart to the
development of aMSER.
3.1 MSER Parameter Configurations
The low number detection of ER occurred due to inability of
ER to grow by using current intensity function as there are
less sensitive towards the existence of ridgelines in the
images. Samples of food with low ER detection are showed in
Figure 1.

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
70
Figure 1 : Samples of Food Image with Low Volume of ER
The sensitivity of MSER towards the invisible ridgelines as
shown in Figure 1 can be increased by manipulating of IT and
MAV value [21]. The MAV and IT are the parameters in
MSER that control the region density and uniformity. A
suitable threshold for MAV and IT should be determined to
produce a stable region [21]. The evaluation of the parameters
are necessary as the optimal value of IT and MAV are subject
to the test data [22]. With this concern, an evaluation on the
ITV and MAV have been conducted based on the parameter
configuration as shown in Table 1.
Table 1: MSER Parameter Configurations
The parameter evaluation is divided into two stages. The first
stage is to find the optimum value of IT and the second stage
is to find the optimum value of MAV. The MSER 1 is the
original parameter configuration of MSER. For each run, the
quantity of ER, the time taken for extraction and classification
performance were recorded. The range of IT and MAV value
are based on the recommendation in [23]. As the IT value
decreased and MAV value increased, it will detect more
ridgelines that produced a greater number of extremal regions.
3.2 Flowchart of aMSER algorithm
Figure 2 shows the flowchart of the execution of aMSER
algorithm.
Figure 2 : Flowchart of aMSER algorithm
The aMSER is executed in food category basis from category
1 to category n. Initially, the food images within a category
were accessed. The images were converted from RGB (Red,
Green, Blue) format into gray-scale format to reduce the
complexity [24]. Then, the ER detected by using MSER 1,
followed by counting the total of ER for each image and
stored in a cell array. After that, a re-evaluation on the
quantity of ER for each image is performed. During the
re-evaluation, a condition is set. The condition states if the
quantity of ER in an image is less than a 100, the image will
need to repeat the ER detection stage where an optimal
parameter of MSER will be applied. The observation indicates
that the food categories that consist lot of foods with ER
below than 100 has yield low classification rate. Then, all the
new set of the quantity of ER will be updated in cell array
before the features are extracted by using Speeded Up Robust
Feature (SURF) descriptor. SURF is chosen to be paired with
MSER due its balanced performance between accuracy and
efficiency, less sensitive to noise and more practical for real
time application [10], [11], [14]. Furthermore, SURF
generates shorter length of feature vector that was reasoned
for a speedy feature encoding process, produced a distinctive
feature and robust to the geometric and photometric
deformation.
4. FEATURE REPRESENTATION
The aMSER generate a huge and diverse amount of interest
points. Literally, local feature can be represented as
from n dimensional features from
an image. For instance, hundreds or even thousands of interest
points were generated per image and the amounts of interest
points for all images may reach up to hundred thousand of
interest points. With this condition, it was not feasible to feed
the feature descriptions into machine learning classifier as it
may incur lot of computational cost. Eventually, the
representation of local feature needs to be transformed into
another level of representation by using certain feature
encoding technique.
Hard assignment technique encodes local feature by assigning
each descriptor to the nearest visual word with indication of
response 1 and the rest of visual words with response 0. The
visual words are generated by using unsupervised learning
algorithm that provide a model of local feature interest points
distribution of X. Specifically, the clustering algorithm such
as k-means is chosen due to its simplicity and it was
commonly used in previous researches. The visual word is the
terminology used in BoF which is referring the cluster
centroid that was defined with cluster size of K or vocabulary
size. Let a set of interest points are described as
. Every interest point are assigned to visual
words . In hard assignment, for each
interest points is assigned to cluster k, then and
for and objective function can be defined as:
Stages
Configurations
IT
MAV
MSER 1
5
0.25
Stage 1
MSER 2
3
0.25
MSER 3
1
0.25
MSER 4
Optimum IT
0.5
Stage 2
MSER 5
Optimum IT
0.75
MSER 6
Optimum IT
1

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
71
(1)
Then, the coding representation v for the local feature x is
described as:
(2)
5. DATASET AND PERFORMANCE MEASUREMENT
The experiments are conducted by using UEC-Food100
dataset [25]. The UEC-Food100 dataset consists of 100 food
categories with total of 14,467 images. Each image is having a
different pixel dimensions and on average, there are around
150 images per category. These images are collected from the
World Wide Web from real world settings with multiple
classes of food types, great differences in image contrast,
lighting and appearance. Figure 3 shows the samples of image
from the dataset.
Figure 3 : Samples of UEC100-Food dataset
There are four performance measurement that were used to
measure the classification performance which are
classification rate, error rate, precision and recall. There are
calculated by using the following formula:
In addition to that, the performance of detector and descriptor
are measured based on the ideal properties recommended by
[26] which are the quantity of interest points and the execution
time. Both are mentioned as the most practical performance
measurement for the real-time applications. The descriptor is
also measured based in the compactness to describe the size of
representation.
6. EXPERIMENTAL RESULTS
The results of the experiments can be divided into three
sections. Section 6.1 presents the evaluation on MSER
parameter. Section 6.2 presents the performance comparisons
of aMSER with the other methods. Section 6.3 provide the
performance on the texture-less food categories.
6.1 Evaluation of MSER parameter configuration
This section provides the experiments results of the MSER
parameter configurations. The Intensity Threshold (IT) and
Maximum Area Variation (MAV) value have been configured
to increase the quantity of extremal region (ER) detection on
food images with the number of ER below than 100. Figure 4
shows the classification rate and the quantity of ER for each
MSER configuration. The effect of IT configuration can be
referred in MSER 2 and MSER 3. While the effect of MAV
configuration can be referred in MSER 4, MSER 5 and MSER
6. MSER 1 refers to the original parameter configuration.
Figure 4 : Classification rate and ER quantity of MSER
The graph in Figure 4 shows the significant improvement of
the classification rate through the IT configuration from
73.89% by using MSER 1 to 89.75% by using MSER 3. The
quantity of ER has also increased dramatically which is about
140%. However, the configuration of MAV has showed little
effect on both classification rate and ER quantity. Figure 5
shows the time taken in seconds for detection and encoding.
Figure 5 : Detection and encoding time of MSER parameter
configuration
The graph in Figure 5 showed regardless IT or MAV
configurations, both have consistently extended the detection
and encoding execution time. The detection time has been

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
72
increased by about 200%. The effect of the configurations is
even more obvious on the encoding time which has spike for
more than 100 times from the initial configurations in MSER
1. This is because this treatment (parameter configuration) has
been applied to all food images regardless their interest points
quantity. This problem has led to the idea of implementing
adaptive mechanism in the MSER extraction where only
certain food images are selected to undergo this treatment.
Figure 6 shows the effect of IT and MAV configuration on a
sample of food image. Figure 6 (c) and (d) show the effect of
IT configuration and Figure 6 (e), (f) and (g) show the effect
of MAV configuration.
Figure 6 : Sample of ER detection by using different MSER
parameter configurations
As aforementioned previously, the configuration of MAV has
little effect on ER density. In fact, the configuration of MAV
in Figure 6 (e), (f) and (g) have null effect on the ER quantity.
While, the IT configuration has increased the ER quantity
from 89 to 272. The ER from the background has increase as
well. Also, the ER detection have become grainier as it was
more sensitive towards region intensity. This finding shows
the capability of MSER to detect regions from the
fine-grained type of foods.
6.2 Evaluation of aMSER
This section presents the performance results of aMSER. By
using aMSER, the extremal region quantity on the targeted
food images have been increased by using the configuration
MSER 6. The food images that have the number of ER below
than 100 are usually food images with texture-less, small
images and low contrast. Table 2 shows the comparisons of
classification performance and extraction time between
aMSER, MSER 1 and MSER 6.
Undeniably, the MSER 6 yield the best classification
performance as the number of extremal regions or interest
points are much denser. Indeed, dense interest points
sampling tend to produce an informative feature
representation that lead to better classification accuracy [27].
Table 2: Classification Performance and Extraction Time of
aMSER
Performance
Measurement
MSER 1
aMSER
MSER 6
Detection (min.)
30.04
113.7
87.33
Encoding (min.)
19.73
41.01
231.51
Classification
Rate %
73.89
84.20
90.96
Error Rate %
0.40
0.20
0.10
Precision %
74.00
84.30
91.00
Recall %
73.90
84.20
91.00
Extremal
Regions
3,087,664
3,576,594
7,610,852
In the flipside, it has also increased the quantity of ER by
about 146% and 10 times for encoding time from the MSER 1.
The proposed aMSER has also improved significantly the
classification rate from 73.89% to 84.2% with only about 15%
rise in ER quantity. The encoding time also demonstrated just
a slight increase. Figure 7 and 8 showed a graph of
performance comparisons of aMSER with the features that
were used in previous food recognition including Histogram
of Gradient descriptor by using Different of Hessian detector
(DoH-HOG) [28], Speeded Up Robust Feature by using
Different of Hessian detector (DoH-HOG) [29] and Scale
Invariant Feature Transform by using Different of Gaussian
detector (DoG-SIFT)[6].
Figure 7 : Performance of features
Figure 8 : No of ER between the features

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
73
The results showed the classification rate of aMSER and even
MSER, has outperformed the DoH-HOG, DoH-SURF and
DoG-SIFT. The extraction time of DoG-SIFT is the lengthiest
and even more than MSER 6. Figure 8 shows the graph of the
quantity of ER generated by the features.
DoG-SIFT generates the highest amount of ER. The aMSER
produced lesser amount of ER than DoH-HOG, DoH-SURF
and DoG-SIFT but still managed to get better classification
rate as shown in Figure 7.
6.3 Evaluation of texture-less foods
This section presents the classification rate of food categories
that contained many texture-less foods images as shown in
Figure 9.
Figure 9 : Classification rate of texture-less food categories
Based on classification rate showed in Figure 9, the proposed
method aMSER has improved the classification rate of
texture-less food category by obtaining an average of 79.36%.
Meanwhile, the average of classification rate by using MSER,
HOG, SURF and SIFT are 61.07%, 63%, 54.9% and 58.97%
respectively. Figure 10 shows the improvement of ER
detection by using aMSER on few samples of texture-less
food images. The configuration of IT and MAV value and in
MSER has managed to increase the ER detection in the
respective food images. Thus, more features can be captured
and represented.
Figure 10 : Samples of extremal region detection on texture-foods
by using aMSER
7. CONCLUSIONS AND FUTURE WORKS
The proposed aMSER has provides a flexibility in detecting
interest regions to overcome the problem of interest points
scarcity on food images with smooth texture such as the
liquid-based foods, tiny images and low level of brightness.
Furthermore, the datasets are built from the real-world setting
or uncontrolled condition that make food images
characterized by the inconsistency and variability of image
quality. The aMSER has improved the classification rate of
the texture-less food categories to 79.36% from 61.07% by
using the traditional MSER as well as the other previous
methods. In overall, the aMSER has improved the
classification accuracy from 73.89% to 84.20%. The findings
also highlighted the efficiency aspect of the recognition where
a reasonable speed of interest points detection and feature
encoding as well as compact number of interest points have
been generated by using aMSER. In the future work, aMSER
can be extended to self-adaptive ER detector where a learning
algorithm can be incorporated to identify the most optimal
parameter for each individual food images. The problem even
can be modularized beyond using optimal parameter for the
lack of interest region density since there is cases where small
set of interest regions are already informative enough to
describe the characteristic of foods. There are might be the
other factors that can be considered other than tuning the
MSER parameter to improve the recognition performance. By
using self-adaptive detector, an optimal way in selecting
parameter tuning and an optimum of interest points for each
image can be performed.
ACKNOWLEDGMENT
The authors acknowledge the financial supported by the Putra
Grant (Cost Center: 9569000 ) funded by the Universiti Putra
Malaysia (UPM).
REFERENCES
[1] A. M. Coulston, C. j. Boushey, and M. Ferruzzi, G.,
Nutrition in the Prevention and Treatment of Disease,
in Nutrition in the Prevention and Treatment of Disease,
3rd ed., Academic Press, 2013, pp. 5–30.
[2] F. Ragusa, V. Tomaselli, A. Furnari, S. Battiato, and G.
M. Farinella, Food vs Non-Food Classification, in 2nd
International Workshop on Multimedia Assisted Dietary
Management, 2016, pp. 77–81.
https://doi.org/10.1145/2986035.2986041
[3] T. Ege and K. Yanai, Image-Based Food Calorie
Estimation Using Knowledge on Food Categories,
Ingredients and Cooking Directions, Proc. Themat.
Work. ACM Multimed. 2017, pp. 367--375, 2017.
https://doi.org/10.1145/3126686.3126742
[4] G. M. Farinella, D. Allegra, M. Moltisanti, F. Stanco, and
S. Battiato, Retrieval and classification of food images,
Comput. Biol. Med., vol. 77, pp. 23–39, 2016.
https://doi.org/10.1016/j.compbiomed.2016.07.006
[5] M. N. Razali and N. Manshor, Object Detection

Mohd Norhisham bin Razali et al., International Journal of Advanced Trends in Computer Science and Engineering, 8(1.4), 2019, 68- 74
74
Framework for Multiclass Food Object Localization
and Classification, Adv. Sci. Lett., vol. 24, no. 4, pp.
1357–1361, 2018.
https://doi.org/10.1166/asl.2018.10749
[6] F. Kong, H. He, H. A. Raynor, and J. Tan, DietCam:
Multi-view regular shape food recognition with a
camera phone, Pervasive Mob. Comput., vol. 19, no. C,
pp. 108–121, 2015.
[7] H. Kagaya and K. Aizawa, Highly Accurate
Food/Non-Food Image Classification Based on a Deep
Convolutional Neural Network, in International
Conference on Image Analysis and Processing, 2015,
vol. 9281, pp. 350–357.
[8] F. Zhu, M. Bosch, N. Khanna, C. J. Boushey, and E. J.
Delp, Multiple Hypotheses Image Segmentation and
ClassificationWith Application to Dietary
Assessment, IEEE J. Biomed. Heal. Informatics, vol. 19,
no. 1, pp. 377–388, 2015.
https://doi.org/10.1109/JBHI.2014.2304925
[9] Z. Zong, D. T. Nguyen, P. Ogunbona, and W. Li, On the
combination of local texture and global structure for
food classification, Proc. - 2010 IEEE Int. Symp.
Multimedia, ISM 2010, pp. 204–211, 2010.
[10] M. N. Razali, N. Manshor, A. A. Halin, R. Yaakob, and
N. Mustapha, Food Category Recognition using SURF
and MSER Local Feature Representation, in
International Visual Informatics Conference, 2017, pp.
212–223.
https://doi.org/10.1007/978-3-319-70010-6_20
[11] M. H. Lee and I. K. Park, Performance evaluation of
local descriptors for maximally stable extremal
regions, J. Vis. Commun. Image Represent., vol. 47, pp.
62–72, 2017.
[12] S. Krig, Local Feature Design Concepts,
Classification, and Learning, Comput. Vis. Metrics, pp.
131–189, 2014.
[13] M. M. Anthimopoulos, L. Gianola, L. Scarnato, P. Diem,
and S. G.Mougiakakou, A Food Recognition System for
Diabetic Patients Based on an Optimized
Bag-of-Features Model, IEEE J. Biomed. Heal.
Informatics, vol. 18, no. 4, pp. 1261–1271, 2014.
https://doi.org/10.1109/JBHI.2014.2308928
[14] P. Ma, M. Seeland, M. Rzanny, N. Alaqraa, and J. Wa,
Plant species classification using flower images — A
comparative study of local feature representations,
PLoS One, pp. 1–30, 2017.
[15] X. Zhang, Y.-H. Yang, Z. Han, H. Wang, and C. Gao,
Object Class Detection: A Survey, ACM Comput.
Surv., vol. 46, no. 1, pp. 1–46, 2013.
[16] T. Joutou and K. Yanai, A food image recognition
system with Multiple Kernel Learning, in 2009 16th
IEEE International Conference on Image Processing
(ICIP), 2009, pp. 285–288.
[17] H. Hoashi, T. Joutou, and K. Yanai, Image Recognition
of 85 Food Categories by Feature Fusion, in IEEE
International Symposium on Multimedia, 2010.
https://doi.org/10.1109/ISM.2010.51
[18] J. Yu, Z. Qin, T. Wan, and X. Zhang, Feature
integration analysis of bag-of-features model for
image retrieval, Neurocomputing, vol. 120, pp.
355–364, 2013.
[19] E. Nowak, F. Jurie, and B. Triggs, Sampling strategies
for bag-of-features image classification, in 9th
European Conference on Computer Vision, 2006, vol.
3954 LNCS, pp. 490–503.
[20] E. Salahat and M. Qasaimeh, Recent Advances in
Features Extraction and Description Algorithms : A
Comprehensive Survey, in IEEE International
Conference on Industrial Technology (ICIT), 2017.
[21] N. Takeishi, A. Tanimoto, T. Yairi, Y. Tsuda, F. Terui, N.
Ogawa, and Y. Mimasu, Evaluation of Interest-region
Detectors and Descriptors for Automatic Landmark
Tracking on Asteroids, Trans. Jpn. Soc. Aeronaut.
Space Sci., vol. 58, no. 1, pp. 45–53, 2015.
https://doi.org/10.2322/tjsass.58.45
[22] S. Krig, Interest Point Detector and Feature
Descriptor Survey, in Computer Vision Metrics, no. 1,
Apress, Berkeley, CA, 2014, pp. 217–282.
[23] J. Matas, O. Chum, M. Urban, and T. Pajdla, Robust
Wide Baseline Stereo from, Br. Mach. Vis. Conf., pp.
384–393, 2002.
[24] S. Jabeen, Z. Mehmood, T. Mahmood, T. Saba, A.
Rehman, and M. T. Mahmood, An effective
content-based image retrieval technique for image
visuals representation based on the
bag-of-visual-words model, PLoS One, pp. 1–24, 2018.
https://doi.org/10.1371/journal.pone.0194526
[25] K. Yanai and Y. Kawano, Food Image Recognition
Using Deep Convolutional Network with
Pre-Training and Fine-Tuning, in IEEE International
Conference on Multimedia & Expo Workshops
(ICMEW), 2015.
[26] A. Ziomek and M. Oszust, Evaluation of Interest Point
Detectors in Presence of Noise, Int. J. Intell. Syst. Appl.,
vol. 8, no. 3, pp. 26–33, 2016.
https://doi.org/10.5815/ijisa.2016.03.03
[27] B. Ionescu, J. Benois-Pineau, T. Piatrik, and G. Quenot,
Fusion in Computer Vision, Adv. Comput. Vis. Pattern
Recognit., p. 272, 2014.
https://doi.org/10.1007/978-3-319-05696-8
[28] Y. Kawano and K. Yanai, FoodCam: A real-time food
recognition system on a smartphone, Multimed. Tools
Appl., vol. 74, no. 14, pp. 5263–5287, 2015.
https://doi.org/10.1007/s11042-014-2000-8
[29] H. Pooja and P. S. A. Madival, Food Recognition and
Calorie Extraction using Bag-of- SURF and Spatial
Pyramid Matching Methods, Int. J. Comput. Sci. Mob.
Comput., vol. 5, no. 5, pp. 387–393, 2016.