Conference PaperPDF Available

Deep learning for automatic target recognition with real and synthetic infrared maritime imagery


Abstract and Figures

Supervised deep learning algorithms are re-defining the state-of-the-art for object detection and classification. However, training these algorithms requires extensive datasets that are typically expensive and time-consuming to collect. In the field of defence and security, this can become impractical when data is of a sensitive nature, such as infrared imagery of military vessels. Consequently, algorithm development and training are often conducted in synthetic environments, but this brings into question the generalisability of the solution to real world data. In this paper we investigate training deep learning algorithms for infrared automatic target recognition without using real-world infrared data. A large synthetic dataset of infrared images of maritime vessels in the long wave infrared waveband was generated using target-missile engagement simulation software and ten high-fidelity computer-aided design models. Multiple approaches to training a YOLOv3 architecture were explored and subsequently evaluated using a video sequence of real-world infrared data. Experiments demonstrated that supplementing the training data with a small sample of semi-labelled pseudo-IR imagery caused a marked improvement in performance. Despite the absence of real infrared training data, high average precision and recall scores of 99% and 93% respectively were achieved on our real-world test data. To further the development and benchmarking of automatic target recognition algorithms this paper also contributes our dataset of photo-realistic synthetic infrared images.
Content may be subject to copyright.
Deep Learning for Automatic Target Recognition with Real
and Synthetic Infrared Maritime Imagery
Samuel T. Westlakea, Timothy N. Volonakisb, James Jackmanc, David B. Jamesa, and Andy
aCentre for Electronic Warfare Information and Cyber, Cranfield University, Defence Academy
of the United Kingdom, Shrivenham, SN6 8LA
bMBDA UK Ltd, PO Box 5, Filton, Bristol, BS34 7QW
cUniversity of Oxford, Institute of Biomedical Engineering, Old Road Campus, Oxford, OX3
Supervised deep learning algorithms are re-defining the state-of-the-art for object detection and classification.
However, training these algorithms requires extensive datasets that are typically expensive and time-consuming
to collect. In the field of defence and security, this can become impractical when data is of a sensitive nature, such
as infrared imagery of military vessels. Consequently, algorithm development and training are often conducted
in synthetic environments, but this brings into question the generalisability of the solution to real world data.
In this paper we investigate training deep learning algorithms for infrared automatic target recognition
without using real-world infrared data. A large synthetic dataset of infrared images of maritime vessels in the
long wave infrared waveband was generated using target-missile engagement simulation software and ten high-
fidelity computer-aided design models. Multiple approaches to training a YOLOv3 architecture were explored
and subsequently evaluated using a video sequence of real-world infrared data. Experiments demonstrated
that supplementing the training data with a small sample of semi-labelled pseudo-IR imagery caused a marked
improvement in performance. Despite the absence of real infrared training data, high average precision and recall
scores of 99% and 93% respectively were achieved on our real-world test data. To further the development and
benchmarking of automatic target recognition algorithms this paper also contributes our dataset of photo-realistic
synthetic infrared images.
Keywords: Automatic target recognition, deep learning, infrared, anti-ship, synthetic, maritime, dataset
Most infrared (IR) anti-ship automatic target recognition (ATR) algorithms are designed around classical com-
puter vision concepts, such as adaptive thresholding and hand-crafted feature extraction.16However, in many
comparable domains, the state-of-the-art has been redefined by deep convolutional neural network (DCNN)-based
algorithms.711 The application of these algorithms to anti-ship ATR has the potential to improve robustness
in adverse conditions and yield significant gains in target recognition and identification performance. These
developments are crucial to emerging systems, facilitating the automatic prioritisation of high value targets and
contributing neutral and friendly shipping avoidance capabilities.12
DCNNs are heavily reliant on increasingly large collections of data and, in many applications, they require
large benchmark datasets.1315 Consisting of thousands—and sometimes even millions—of annotated examples,
these datasets have practically eliminated the time-consuming task of data collection and annotation in their
respective domains. In addition, such benchmarking promotes the direct comparison of algorithms, enabling
rapid identification of suitable approaches for further development. However, within the field of IR anti-ship
ATR, no large benchmark datasets are available, causing the development of new algorithms to be hampered
Further author information: (Send correspondence to Samuel T. Westlake.)
Samuel T. Westlake: E-mail:
by the cumbersome task of generating bespoke datasets. Furthermore, such datasets typically contain too few
examples to train deep neural networks or facilitate robust validation, and are rarely made openly available,
further hindering the direct comparison of algorithms.
For future IR anti-ship ATR algorithms to fully leverage state-of-the-art techniques, new approaches to both
dataset generation and algorithm training are required—and thus, in this paper, we explore both. In Section 3, we
present a new synthetic IR anti-ship-focussed dataset and detail its design, generation and online augmentation.
While in Section 4, this dataset is used to investigate several approaches to training the YOLOv3 object detection
algorithm in the absence of any real-world IR imagery. The resultant performance is presented in Section 5, its
associated implications and recommendations are discussed in Section 6, while Section 7 concludes the paper.
The principal deficiency of many ship-focussed datasets is their size, with many datasets consisting of fewer
than 200 IR images.1618 Such datasets are not suited to training DCNNs as, for complex tasks like target
recognition, training examples must be sufficiently numerous and diverse to comprehensively represent expected
test conditions. For comparison, the COCO object detection dataset, which includes 80 categories, contains ca.
120,000 training examples,13 and there are over 1 million labelled examples in the ImageNet classification and
localisation dataset.15 Consequently, deep learning algorithms trained with such small quantities of data cannot
be expected to operate reliably under the wide range of conditions that anti-ship missiles are expected to face.
To the best of our knowledge, the largest relevant dataset is the Singapore Maritime Dataset (SMD),19,20
which contains a mix of fully labelled off- and on-shore, visual-spectrum and NIR video sequences, totalling 31,653
frames. However, the absence of infrared imagery in the long wave infrared band—the waveband of choice for
modern IR anti-ship missiles—means this dataset is not wholly suitable for the naval domain. Furthermore,
despite its large size, the SMD still lacks the necessary diversity for training complex algorithms that contain
millions of learnable parameters.19 Nevertheless, the SMD constitutes the largest openly-accessible collection of
real-world maritime image data, and is thus used as a valuable benchmark in other applications, such as harbour
surveillance21,22 and collision avoidance.23
A second deficiency of many datasets relates to their lack of variation in environmental factors, such as at-
mospheric conditions, sea states, and the presence of background ‘clutter’. Broad ranges of possible conditions
are rarely represented, with often just a singular set of atmospheric and sea-surface conditions considered.18,24
Furthermore, background clutter is generally considered a challenging source of false positive detections, espe-
cially in littoral environments, yet such objects are rarely depicted in existing datasets. If ATR algorithms are
to become truly robust against the huge diversity of possible deployment conditions, we consider it crucial that
future datasets include increased combinations of these conditions.
Furthermore, missile seeker algorithms must be capable of detecting and recognising ships of any size or design
within a wide envelope of orientations. However, few datasets fully account for this, with some concerned solely
with broadside perspectives, while others ignore elevation and only consider a horizontal view.3,25,26 Moreover,
most existing datasets depict just a small sample of different ship classes—typically no more than six.1,18,27
This simplifies the detection task and hinders development of recognition and identification capabilities, and
therefore it is crucial that future datasets include more numerous and varied collections of both military and
civilian vessels.
The collection and open distribution of an IR dataset containing a comprehensive selection of military vessels
is unlikely. In response to this, there have been several attempts to generate datasets of synthetically generated
imagery. An early notable example used five wireframe CAD models to generate ca. 41,000 silhouettes at
incremental elevations and azimuths.1However, as binary images, these are geared towards the task of ship
classification only. Improving on this concept, later work used sophisticated missile target engagement software
to generate realistic long wave IR imagery of military vessels, which were then used to train a neural network
classifier for target discrimination.27 These approaches demonstrated the potential of synthetic data for enabling
the use of machine learning algorithms for IR ATR. It also brought several advantages into focus, such as the
reduced time cost, the collection of comprehensive metadata, and crucially, the opportunity to freely depict
military vessels.
Despite these advantages, however, there remain several fundamental challenges concerning the use of syn-
thetic data for anti-ship ATR. For example, existing synthetic datasets are yet to account for the wide range
of possible external factors, such as atmospheric conditions, sea states and background clutter. Also, despite
the potential to depict any number of ship designs, creation of the these CAD models is a time-consuming and
skilled task, and so current synthetic datasets remain just as limited as their real-world counterparts in this
respect.1,27,28 Furthermore, ship thermal signatures change dramatically with external and onboard conditions,
though this also is yet to be accounted for, as invariant ship surface temperatures are currently assumed.27
While synthetic data generation is an attractive solution to some of the many data-related challenges in IR
ATR, significant improvements are required if such data is to enable the effective training of robust detection
In our view, overcoming the restricted availability of real-world IR imagery and limited realism of synthetic
datasets will likely require a hybrid dataset which draws on the advantages of each. Consequently, this paper
details an improved approach to synthetic IR dataset generation and presents our own high-quality and openly
available dataset for the development of future machine learning IR anti-ship ATR algorithms. We also demon-
strate the use of this data for training complex high-performance deep learning object detection algorithm which
we evaluate using a sequence of real-world IR imagery.
This section describes the generation of our synthetic IR dataset: CAD model selection and design, thermal
property assignment and image generation. This section also describes the sea, sky and background augmentation
process, used to increase image complexity and diversity during algorithm training. This dataset has been made
openly available.29
3.1 Selection and design of ship models
Ten classes of ship were selected across four type designations, as summarised in Table 1. Three types of military
vessel were included: corvette, frigate and destroyer; each represented by three different ship classes, and a single
civilian vessel, the MV Armorique passenger ferry. These classes of vessels were selected to represent a variety
of different designs and all are currently in operation with a range of different nations. Visualised in Appendix
A, the ships were modelled using 3D CAD software and were designed to represent their real-world counterparts
as accurately as possible.
Table 1. Classes of ships selected for the dataset (displacement refers to standatd displacement).30
Class Type Commissioned Length (m) Displacement (tonnes)
Ada Corvette 2011 99.0 1,524
Independence Corvette 2010 128.5 3,188
Visby Corvette 2009 72.7 630
Alvaro de Bazan Frigate 2002 146.4 6,250
Jiangkai II (Type 054A) Frigate 2008 134.0 3,556
Oliver Hazard Perry Frigate 1977 135.6 2,794
Akizuki Destroyer 2012 151.0 5,050
Sejong Daewang (KDX-III) Destroyer 2008 165.9 7,600
Zumwalt Destroyer 2016 186.0 15,995
Armorique Ferry 2008 168.0 29,500
3.2 Assignment of thermal properties
Before the IR appearance of these ships could be simulated, it was necessary to apply thermal properties,
including temperature and emissivity, to the surfaces of each model. To account for the continuously varying
nature of ship thermal signatures, nine versions of each CAD model were created, with each to be prescribed a
unique set of surface temperatures.
A selection of real-world ship imagery was collected using a Tau2 long wave IR camera to inform the
design of an algorithm that could generate diverse and realistic thermal signatures. For a given ship model,
a mean temperature, µand standard deviation, σwere drawn from uniform distributions U(2,6) and U(5,20)
respectively. These were used to define a Normal distribution from which the temperature values for each surface
were drawn. If a given surface was an exhaust funnel, its temperature value was elevated by an amount drawn
from a N(40,102). The temperature of antenna and radome surfaces were also elevated in 50% of cases, by
an amount drawn from N(U(3,10), U(0,4)2). All surfaces were prescribed an emissivity value of 0.97, with
exception of glass surfaces, such as windows, which were given an emissivity value of 0.85.
Through this procedure, nine dictionaries of thermal signatures were generated for each ship and applied to
their corresponding CAD model. This resulted in nine unique versions of each vessel, as illustrated in Fig. 1
with the Akizuki class destroyer used in this study.
Figure 1. Illustration of the different thermal appearances generated for the Akizuki class destroyer.
3.3 IR image generation
Thermal imagery of the CAD models were generated using CounterSim, a target-missile engagement simulator
developed by Chemring Countermeasures Ltd. A virtual long wave thermal imager was defined with a temper-
ature range of 0–100 C and resolution of 1024 ×512. Using this virtual camera, thermal imagery for each ship
was generated at each ship azimuth in {0, 1, ..., 359}, each camera pitch in {0, -10, -20}, and each camera-ship
distance in {1000, 1111 1250, 1429, 1666, 2000, 2500, 3333, 5000, 10,000}m. These range values were selected
to give a linear reduction in the perceived height of a given target.
During image generation, the ocean was modelled as a flat surface with temperatures drawn from U(5,20),
background sky was modelled as a constant with temperatures drawn from U(5,25), and atmospheric transmis-
sion was modelled using the moderate resolution atmospheric transmission (MODTRAN4) atmospheric model.31
These arbitrary values were selected to cover a wide range of feasible temperature values for both the sea sur-
face and sky temperatures values, with constant flat surfaces assumed to facilitate augmentation at a later
stage. In total, 972,000 images were generated, along with a further 108,000 binary masks for use in online data
augmentation and for semantic segmentation.
3.4 Online image augmentation
To increase variation within our synthetic data, a three-stage stochastic online data augmentation pipeline was
designed to enable the addition of different sea and sky-states, and background clutter. Each of these steps relied
on collections of pre-processed images that were stochastically superimposed into a given image at runtime.
For sky-state augmentation, images that include the sky were collected from various online sources and pre-
processed by setting pixels that corresponded to blue sky to zero and converting to grayscale. At runtime, for a
given synthetic image, a cloudy image was selected, randomly resized and cropped to shape, with its pixel intensity
scaled according to Equation (1), where ˜
Iis the normalised real-world image and c is the average pixel value of the
sky region in the synthetic image. Values of 2 and 50 were chosen for the bounds a and b respectively, to provide
a feasible variety of pixel rages, and therefore temperature ranges. The cloudy image was then superimposed
above the sea-sky horizon line of the synthetic image, using the synthetic image’s corresponding binary mask to
preserve target pixels.
Ii, j =˜
Ii, j ×U(a, b) + c, (1)
The addition of background clutter followed a similar procedure. Images depicting plausible background
scenery and objects such as oil platforms, wind turbines, icebergs, small islands, and built-up coastlines were
collected from various online sources. Similar to before, these images were pre-processed by the removal of
background pixels and conversion to grayscale. At runtime, background images were selected at random and the
same process of resizing, cropping and pixel scaling was applied before the scene was then superimposed along
the sea-sky horizon line of the given synthetic image. Pixel scaling of the cluttered images was conducted in
accordance with Equation (1); however in this case a value of 0 was used for c and the bounds a and b were
selected depending on the nature of the cluttered scene. Values of 20 and 80 respectively were used in the case of
images that depicted human-made structures, values of 0 and 2 were used in the case of icebergs, and values of
15 and 60 were used in the remainder of cases. These value pairs were selected arbitrarily to correspond feasibly
with the nature of the background clutter.
Finally, for sea-state augmentation, two collections of images of ocean surfaces were collected. The first
of these include images taken from near sea-level and the second consisted of images taken from an elevated
perspective. At runtime, if the sea-sky horizon was visible in the given synthetic image, a real-world image
taken from the first collection was chosen, otherwise an image from the second collection was used. The same
pre-processing was applied, with Equation (1) used to rescale pixels; however in this case values of 5 and 30
were used for the bounds a and b respectively, and c corresponded to the average pixel intensity of sea region
in the synthetic image. There is also scope for the addition of sensor noise at this stage, and the effect of these
augmentation processes for both horizontal and elevated images can be seen in Fig. 2and Fig. 3respectively.
4.1 Detection Algorithm
The YOLOv3 algorithm32 was selected due its high accuracy score, as achieved on the COCO benchmark dataset.
Selection of YOLOv3 was also influenced by its potential to generalise well across domains, as indicated by the
ability of its predecessor33 to recognise people in art34,35 despite being trained on trained on the VOC 2007
dataset.36 Moreover, YOLOv3 is an efficient single-shot algorithm capable of real-time inference and several of
its variants have been further optimised for speed.33,37
The architecture of YOLOv3 is fully-convolutional, consisting of a backbone feature detector followed by a
Pyramid Feature Network-style architecture38 that enables inference across three scales. The backbone feature
detector is the Darknet-53 DCNN which, through its use of 1 ×1 convolutional layers and residual blocks, is
capable of achieving accuracies similar to the much larger ResNet-152 model,32,39 but at twice the inference
speed. The input to Darknet-53 is down-sampled by a factor of 32, whereby the output and two skip connections
at 1/16 and 1/8scale are concatenated with downstream feature maps before the final outputs are regressed.
The outputs of YOLOv3 relate to non-overlapping 8-, 16- and 32-pixel cells. For each of these cells, predictions
are inferred containing bounding box coordinates, classification predictions, and an “ob jectness” score, which
Figure 2. Examples of sea-level synthetic images before and after sky, sea and background state augmentation. (Contrast
enhanced for illustrative purposes.)
Figure 3. Examples of high-elevation synthetic images before and after sky, sea and background state augmentation
(Contrast enhanced for illustrative purposes.)
relates to the algorithm’s confidence that the prediction corresponds to a real object. This output structure
was modified to include capacity for the prediction of both object type and object class, in accordance with
recognition and identification paradigm that is commonplace within the field of ATR.
4.2 Algorithm training
Three approaches to model training were evaluated, in order to test whether synthetic imagery can in fact
be used as an effective substitute for real-world IR data. In the baseline experiment, training relied solely on
our synthetic dataset, while in subsequent experiments, the training data was augmented with the addition of
semi-labelled visual-spectrum and pseudo-IR imagery.
4.2.1 Baseline
Our synthetic dataset was the sole source of training data during the control experiment, with sea, sky and
clutter augmentation being applied as described in Section 3. Further augmentation included random shifts and
rotations and, to simulate noise that real systems typically incur, sensor noise was added in the form of random
Gaussian blur, motion blur and fixed pattern. After this all training images were resized to 576 ×288 ×1.
Model weights were optimised using the Adam method of gradient descent40 and YOLOv3 loss function.32,33
The model was trained for 50 epochs of 128,000 randomly selected training instances, with a batch size of 16
and a learn rate of 1e-4, which was gradually reduced by a factor of 0.01 using the cosine learning rate decay.41
4.2.2 Inclusion of semi-labelled visual-spectrum images
In order to improve generalisability to real-world imagery, the model was re-trained with the addition of 8,343 im-
ages in the visual spectrum, collected from various online sources. These images were collected programmatically,
using the ten class names of the ships considered in this study as search terms, and thus type and class labels
for each image were designated automatically. However, to avoid the labour-intensive task of manual bounding
box annotation, bounding box labels were omitted, and thus these images are referred to as semi-labelled.
An image classifier was added to the YOLOv3 architecture at the output of backbone feature detector,
consisting of global average pooling42 followed by a fully-connected layer. The output layer of this classifier
contained 14 output neurons; four of which related to classification of ship type, and the remaining ten being
for the classification of ship class. The resultant type and class predictions from this classifier were scored
independently using cross-entropy loss. During training, the quantity of semi-labelled data was artificially inflated
by duplication to provide a sampling ratio of ca. 5 synthetic images to every semi-labelled image.
4.2.3 Inclusion of semi-labelled pseudo-IR images
With the aim of maximising the impact of semi-labelled visual-spectrum data, the third approach used a set
of data transforms for the conversion of visual-spectrum data to pseudo-IR imagery. This was done by the
application of multiple stochastically linear transforms to the pixel intensity of each image.
4.3 Evaluation
Algorithm evaluation was conducted with a sequence of real-world imagery, collected by using a Tau2 IR
camera, sensitive in the long wave infrared waveband. This sequence depicts a navy frigate traveling under its
own power in a littoral environment with a wide range of dense background clutter, including: a navigation buoy,
and both rocky and build up shorelines. Metrics used for evaluation were precision, recall and intersection over
union (IoU), each of which ignored ship type and class predictions.
Regarding the training data, across each of the three experiment the YOLOv3 algorithm achieved average
precision, recall and IoU scores that exceeded 0.98, 0.92 and 0.86 respectively. After evaluation with the IR
validation sequence, the best performance was achieved by the model trained with a mix of synthetic and semi-
labelled pseudo-IR imagery. This version achieved peak values of 0.991, 0.932 and 0.872 respectively for precision,
recall and average IoU.
Fig. 4illustrates the F-scores achieved by each of the three sets of training conditions. When trained with
both synthetic and semi-labelled pseudo-IR data, the algorithm achieved a peak validation F-score of 0.96. Yet
training with synthetic data only, or with synthetic data supplemented with semi-labelled visual-spectrum data,
this score peaked at just 0.42 and 0.50 respectively. This shows that, despite being insufficient for training the
YOL0v3 algorithm on its own, the addition of semi-labelled data facilitated a marked increase in performance.
However, this was only possible after such data was tailored for the IR domain and its variance maximised.
Fig. 5illustrates algorithm inference on some examples of our synthetic training data, with ground truth boxes
drawn in green and predicted target detections drawn in blue with accompanying type and class predictions.
Fig. 6and Fig. 7show a selection of the semi-labelled visual-spectrum and pseudo-IR training images respectively,
with inferred target drawn in blue and ship type and class predictions written in green. Semi-labelled images
were not accompanied by bounding box labels, meaning the detection algorithm was not explicitly shown how
to correctly annotate these examples but was able to do so regardless. This demonstrated the models ability to
generalise and indicated that useful features from these real-world images were being learned and recognised by
the algorithm.
Figure 4. Algorithm validation F-scores.
Figure 5. Examples of algorithm outputs on synthetic data during training. (Contrast enhanced for illustrative purposes,
green boxes are ground truths, blue boxes and text are algorithm predictions).
Figure 6. Examples of algorithm outputs on semi-labelled data during training (blue boxes and text are algorithm
predictions). The depicted boxes were inferred by the algorithm, despite no ground truth bounding boxes being provided
for these images.
Figure 7. Examples of algorithm outputs on semi-labelled data during training (blue boxes and text are algorithm
predictions). The depicted boxes were inferred by the algorithm, despite no ground truth bounding boxes being provided
for these images.
In this paper, an end-to-end deep learning object detector was successfully trained for IR anti-ship ATR with
synthetic data as the sole source of training examples. Achieving this required the resolution of several issues
regarding the generation of synthetic data. Problems regarding the limited diversity of sky and sea states and
background clutter were resolved by the development of an augmentation pipeline that substantially increased
the complexity of our synthetic imagery. Secondly, more realistic and varied ship appearances were enabled
by increasing the standard of detail in the target CAD models and the development of an approach for the
automatic generation of target thermal signatures. These improvements resulted in a deep learning algorithm
that was capable of generalising in the domain of real-world IR imagery.
A marked improvement in validation performance was achieved by supplementing our synthetic training data
with a relatively small quantity of semi-labelled pseudo-IR imagery. This data appeared to enable the learning
of new features that were applicable to the IR validation sequence but not already present in our synthetic
dataset, which had the effect of improving ability to generalise in the real-world IR domain. Importantly, this
demonstrates that the task of image annotation is not always necessary, and that relevant image data can offer
benefits despite a lack of annotation.
That said, the use of semi-labelled visual spectrum data had little effect on algorithm validation performance.
The fact that the algorithm was able to successfully detect and recognise targets in these images indicates that
features were learned from this data. Though it is apparent that these new features offered little benefit when
applied in the IR domain, thus indicating that these visual spectrum images were not sufficiently representative
of IR data.
The discrepancy between training and test scores when the algorithm was trained with our synthetic data only
indicates that there remains room for improvement of this synthetic data. This synthetic data could be made
more representative of the test domain by the use of real IR imagery, as opposed to grayscale visual spectrum
data, in the image augmentation pipeline. Additionally, real-world conditions could be better represented by
the inclusion of foreground clutter, such as bow waves and other objects, which are known to cause the partial
occlusion of targets. Future iterations of this dataset would also benefit from the inclusion of more atmospheric
conditions, including rain, fog and solar glint, and the addition of even more classes of vessel.
Deep learning algorithms have the potential to redefine the state-of-the-art in the field of IR anti-ship ATR,
however they require large and diverse collections of annotated training data. This paper describes how such
data can be generated synthetically, using a series of high-fidelity CAD models, each with a range of unique
thermal appearances, and an online image augmentation pipeline. Moreover, we trained a modified version of
the YOLOv3 algorithm under various conditions and identified the use of pseudo-IR imagery as an effective
method to improve the algorithm’s ability to generalise.
Furthermore, this paper presents our new dataset of realistic longwave IR imagery for the training develop-
ment of anti-ship ATR algorithms. This dataset contains 972,000 annotated examples, ten different ship classes
and uses online augmentation of different sky- and sea-states and background clutter to maximise its complexity
and diversity.
[1] Alves, J., Herman, J., and Rowe, N. C., “Robust recognition of ship types from an infrared silhouette,”
tech. rep., Naval Postgraduate School, CA (2004).
[2] Bai, X., Liu, M., Wang, T., Chen, Z., Wang, P., and Zhang, Y., “Feature based fuzzy inference system for
segmentation of low-contrast infrared ship images,” Applied Soft Computing 46, 128–142 (2016).
[3] Gray, G. J., Aouf, N., Richardson, M. A., Butters, B., Walmsley, R., and Nicholls, E., “Feature-based
tracking algorithms for imaging infrared anti-ship missiles,” in [Technologies for Optical Countermeasures
VIII], 8187, 81870T, International Society for Optics and Photonics (2011).
[4] Hu, M.-K., “Visual pattern recognition by moment invariants,” IRE transactions on information theory 8(2),
179–187 (1962).
[5] Lan, J. and Wan, L., “Automatic ship target classification based on aerial images,” in [2008 International
Conference on Optical Instruments and Technology: Optical Systems and Optoelectronic Instruments], 7156,
715612, International Society for Optics and Photonics (2009).
[6] Tremblay, C. and Valin, P., “Experiments on individual classifiers and on fusion of a set of classifiers,” in
[Proceedings of the Fifth International Conference on Information Fusion. FUSION 2002. (IEEE Cat. No.
02EX5997)], 1, 272–277, IEEE (2002).
[7] Badrinarayanan, V., Kendall, A., and Cipolla, R., “Segnet: A deep convolutional encoder-decoder archi-
tecture for image segmentation,” IEEE transactions on pattern analysis and machine intelligence 39(12),
2481–2495 (2017).
[8] Girshick, R., “Fast r-cnn,” in [Proceedings of the IEEE international conference on computer vision], 1440–
1448 (2015).
[9] Krizhevsky, A., Sutskever, I., and Hinton, G. E., “Imagenet classification with deep convolutional neural
networks,” in [Advances in neural information processing systems], 1097–1105 (2012).
[10] Liu, S., Qi, L., Qin, H., Shi, J., and Jia, J., “Path aggregation network for instance segmentation,” in
[Proceedings of the IEEE conference on computer vision and pattern recognition ], 8759–8768 (2018).
[11] Zhang, H., Wu, C., Zhang, Z., Zhu, Y., Zhang, Z., Lin, H., Sun, Y., He, T., Mueller, J., Manmatha, R.,
et al., “Resnest: Split-attention networks,” arXiv preprint arXiv:2004.08955 (2020).
[12] “Long range anti-ship missile (lrasm).”
long-range-anti-ship-missile. (Accessed on 07/14/2020).
[13] Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll´ar, P., and Zitnick, C. L.,
“Microsoft coco: Common objects in context,” in [European conference on computer vision], 740–755,
Springer (2014).
[14] Milan, A., Leal-Taix´e, L., Reid, I., Roth, S., and Schindler, K., “Mot16: A benchmark for multi-object
tracking,” arXiv preprint arXiv:1603.00831 (2016).
[15] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,
Bernstein, M., et al., “Imagenet large scale visual recognition challenge,” International journal of computer
vision 115(3), 211–252 (2015).
[16] Zhaoying, L., Fugen, Z., and Xiangzhi, B., “Infrared ship target segmentation based on region and shape
features,” in [2013 14th International Workshop on Image Analysis for Multimedia Interactive Services
(WIAMIS)], 1–4, IEEE (2013).
[17] Tao, W., Jin, H., and Liu, J., “Unified mean shift segmentation and graph region merging algorithm for
infrared ship target segmentation,” Optical Engineering 46(12), 127002 (2007).
[18] Withagen, P. J., Schutte, K., Vossepoel, A. M., and Breuers, M. G., “Automatic classification of ships from
infrared (flir) images,” in [Signal Processing, Sensor Fusion, and Target Recognition VIII ], 3720, 180–187,
International Society for Optics and Photonics (1999).
[19] Moosbauer, S., Konig, D., Jakel, J., and Teutsch, M., “A benchmark for deep learning based object detec-
tion in maritime environments,” in [Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition Workshops], (2019).
[20] Prasad, D. K., Rajan, D., Rachmawati, L., Rajabally, E., and Quek, C., “Video processing from electro-
optical sensors for object detection and tracking in a maritime environment: a survey,” IEEE Transactions
on Intelligent Transportation Systems 18(8), 1993–2016 (2017).
[21] Andersson, M., Johansson, R., Stenborg, K.-G., Forsgren, R., Cane, T., Taberski, G., Patino, L., and Fer-
ryman, J., “The ipatch system for maritime surveillance and piracy threat classification,” in [2016 European
Intelligence and Security Informatics Conference (EISIC)], 200–200, IEEE (2016).
[22] Palmieri, F. A., Castaldo, F., and Marino, G., “Harbour surveillance with cameras calibrated with ais data,”
in [2013 IEEE Aerospace Conference], 1–8, IEEE (2013).
[23] Campbell, S., Naeem, W., and Irwin, G. W., “A review on improving the autonomy of unmanned surface
vehicles through intelligent collision avoidance manoeuvres,” Annual Reviews in Control 36(2), 267–283
[24] Greer, G., Advanced Algorithms and Countermeasures for Imaging Infrared Anti-Ship Missiles, PhD thesis,
Cranfield University (2013).
[25] Bizer, M. J., “A picture-descriptor extractions program using ship silhouettes,” tech. rep., Naval Postgrad-
uate School, CA (1989).
[26] Herman, J., “Target identification algorithm for the an/aas-44v forward looking infrared (flir),” tech. rep.,
Naval Postgraduate School, CA (2000).
[27] Gray, G., Aouf, N., Richardson, M., Butters, B., Walmsley, R., and Nicholls, E., “Feature-based recognition
approaches for infrared anti-ship missile seekers,” The Imaging Science Journal 60(6), 305–320 (2012).
[28] Fernandez, H., de Seixas, J., Neves, S., and Souza Filho, J., “Combining morphological mapping and
principal curves for ship classification,” in [International Symposium on Signals, Circuits and Systems,
2005. ISSCS 2005.], 2, 605–608, IEEE (2005).
[29] Westlake, S., “Irships dataset.” (2020).
[30] [Jane’s Fighting Ships ] (2019).
[31] Berk, A., Anderson, G. P., Bernstein, L. S., Acharya, P. K., Dothe, H., Matthew, M. W., Adler-Golden,
S. M., Chetwynd Jr, J. H., Richtsmeier, S. C., Pukall, B., et al., “Modtran4 radiative transfer modeling
for atmospheric correction,” in [Optical spectroscopic techniques and instrumentation for atmospheric and
space research III ], 3756, 348–353, International Society for Optics and Photonics (1999).
[32] Redmon, J. and Farhadi, A., “Yolov3: An incremental improvement,” arXiv preprint arXiv:1804.02767
[33] Redmon, J., Divvala, S., Girshick, R., and Farhadi, A., “You only look once: Unified, real-time object
detection,” in [Proceedings of the IEEE conference on computer vision and pattern recognition ], 779–788
[34] Cai, H., Wu, Q., Corradi, T., and Hall, P., “The cross-depiction problem: Computer vision algorithms for
recognising objects in artwork and in photographs,” arXiv preprint arXiv:1505.00110 (2015).
[35] Ginosar, S., Haas, D., Brown, T., and Malik, J., “Detecting people in cubist art,” in [European Conference
on Computer Vision], 101–116, Springer (2014).
[36] Everingham, M. and Winn, J., “The pascal visual object classes challenge 2007 (voc2007) development kit,”
University of Leeds, Tech. Rep (2007).
[37] Huang, R., Pedoeem, J., and Chen, C., “Yolo-lite: a real-time object detection algorithm optimized for
non-gpu computers,” in [2018 IEEE International Conference on Big Data (Big Data) ], 2503–2510, IEEE
[38] Lin, T.-Y., Doll´ar, P., Girshick, R., He, K., Hariharan, B., and Belongie, S., “Feature pyramid networks
for object detection,” in [Proceedings of the IEEE conference on computer vision and pattern recognition],
2117–2125 (2017).
[39] He, K., Zhang, X., Ren, S., and Sun, J., “Deep residual learning for image recognition,” in [Proceedings of
the IEEE conference on computer vision and pattern recognition], 770–778 (2016).
[40] Kingma, D. P. and Ba, J., “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980
[41] Loshchilov, I. and Hutter, F., “Sgdr: stochastic gradient descent with restarts. corr abs/1608.03983 (2016),”
arXiv preprint arXiv:1608.03983 (2016).
[42] Lin, M., Chen, Q., and Yan, S., “Network in network,” arXiv preprint arXiv:1312.4400 (2013).
... Chen [12], Cane [13], Marie [14], Chen [15], Chen [16], Liu [17], Shan [18], Gal [19], Lin [20], Chen [21], Lee [22], Feng [23], Shan [24], Fefilatyev [25] IR N/A Tang [26], Liu [27], Hu [28], Lin [29] MWIR Özertem [30], Wang [31] LWIR Sun [32], Lu [33], Bai [34], Leira [35], Bai [36], Mumtaz [37], Singh [38], Zhang [39], Xu [40], Zhou [41], Schöller [42], Li [43], Westlake [44] Visible + IR N/A Islam [45], Wei [46] Visible + MWIR Nita [47] Visible + LWIR Zhang [48], Ribeiro [49], Farahnakian [50] Visible + SWIR + LWIR Stets [51] Visible + SWIR + MWIR + LWIR Bouma [52] Visible: refers to visible-band images; IR: infrared images, including short-wave infrared (SWIR), medium-wave infrared (MWIR), and long-wave infrared (LWIR); N/A, means that the type of image was not specified in the paper. from the statistical data that infrared images are used the most, followed by visible images. ...
Full-text available
The ocean connects all continents and is an important space for human activities. Ship detection with electro-optical images has shown great potential due to the abundant imaging spectrum and, hence, strongly supports human activities in the ocean. A suitable imaging spectrum can obtain effective images in complex marine environments, which is the premise of ship detection. This paper provides an overview of ship detection methods with electro-optical images in marine environments. Ship detection methods with sea–sky backgrounds include traditional and deep learning methods. Traditional ship detection methods comprise the following steps: preprocessing, sea–sky line (SSL) detection, region of interest (ROI) extraction, and identification. The use of deep learning is promising in ship detection; however, it requires a large amount of labeled data to build a robust model, and its targeted optimization for ship detection in marine environments is not sufficient.
In this survey, we compile a list of publicly available infrared image and video sets for artificial intelligence and computer vision researchers. We mainly focus on IR image and video sets which are collected and labelled for computer vision applications such as object detection, object segmentation, classification, and motion detection. We categorize 92 different publicly available or private sets according to their sensor types, image resolution, and scale. We describe each and every set in detail regarding their collection purpose, operation environment, optical system properties, and area of application. We also cover a general overview of fundamental concepts that relate to IR imagery, such as IR radiation, IR detectors, IR optics and application fields. We analyse the statistical significance of the entire corpus from different perspectives. We believe that this survey will be a guideline for computer vision and artificial intelligence researchers that are interested in working with the spectra beyond the visible domain.
Full-text available
We present a survey on marine object detection based on deep neural network approaches, which are state-of-the-art approaches for the development of autonomous ship navigation, maritime surveillance, shipping management, and other intelligent transportation system applications in the future. The fundamental task of maritime transportation surveillance and autonomous ship navigation is to construct a reachable visual perception system that requires high efficiency and high accuracy of marine object detection. Therefore, high-performance deep learning-based algorithms and high-quality marine-related datasets need to be summarized. This survey focuses on summarizing the methods and application scenarios of maritime object detection, analyzes the characteristics of different marine-related datasets, highlights the marine detection application of the YOLO series model, and also discusses the current limitations of object detection based on deep learning and possible breakthrough directions. The large-scale, multiscenario industrialized neural network training is an indispensable link to solve the practical application of marine object detection. A widely accepted and standardized large-scale marine object verification dataset should be proposed.
Full-text available
We present a novel and practical deep fully convolutional neural network architecture for semantic pixel-wise segmentation termed SegNet. This core trainable segmentation engine consists of an encoder network, a corresponding decoder network followed by a pixel-wise classification layer. The architecture of the encoder network is topologically identical to the 13 convolutional layers in the VGG16 network [1]. The role of the decoder network is to map the low resolution encoder feature maps to full input resolution feature maps for pixel-wise classification. The novelty of SegNet lies is in the manner in which the decoder upsamples its lower resolution input feature map(s). Specifically, the decoder uses pooling indices computed in the max-pooling step of the corresponding encoder to perform non-linear upsampling. This eliminates the need for learning to upsample. The upsampled maps are sparse and are then convolved with trainable filters to produce dense feature maps. We compare our proposed architecture with the widely adopted FCN [2] and also with the well known DeepLab-LargeFOV [3] , DeconvNet [4] architectures. This comparison reveals the memory versus accuracy trade-off involved in achieving good segmentation performance. SegNet was primarily motivated by scene understanding applications. Hence, it is designed to be efficient both in terms of memory and computational time during inference. It is also significantly smaller in the number of trainable parameters than other competing architectures and can be trained end-to-end using stochastic gradient descent. We also performed a controlled benchmark of SegNet and other architectures on both road scenes and SUN RGB-D indoor scene segmentation tasks. These quantitative assessments show that SegNet provides good performance with competitive inference time and most efficient inference memory-wise as compared to other architectures. We also provide a Caffe implementation of SegNet and a web demo at
We present some updates to YOLO! We made a bunch of little design changes to make it better. We also trained this new network that's pretty swell. It's a little bigger than last time but more accurate. It's still fast though, don't worry. At 320x320 YOLOv3 runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster. When we look at the old .5 IOU mAP detection metric YOLOv3 is quite good. It achieves 57.9 mAP@50 in 51 ms on a Titan X, compared to 57.5 mAP@50 in 198 ms by RetinaNet, similar performance but 3.8x faster. As always, all the code is online at
The way that information propagates in neural networks is of great importance. In this paper, we propose Path Aggregation Network (PANet) aiming at boosting information flow in proposal-based instance segmentation framework. Specifically, we enhance the entire feature hierarchy with accurate localization signals in lower layers by bottom-up path augmentation, which shortens the information path between lower layers and topmost feature. We present adaptive feature pooling, which links feature grid and all feature levels to make useful information in each feature level propagate directly to following proposal subnetworks. A complementary branch capturing different views for each proposal is created to further improve mask prediction. These improvements are simple to implement, with subtle extra computational overhead. Our PANet reaches the 1st place in the COCO 2017 Challenge Instance Segmentation task and the 2nd place in Object Detection task without large-batch training. It is also state-of-the-art on MVD and Cityscapes.
Restart techniques are common in gradient-free optimization to deal with multimodal functions. Partial restarts are also gaining popularity in gradient-based optimization to improve the rate of convergence in accelerated gradient schemes to deal with ill-conditioned functions. In this paper, we propose a simple restart technique for stochastic gradient descent to improve its anytime performance when training deep neural networks. We empirically study its performance on CIFAR-10 and CIFAR-100 datasets where we demonstrate new state-of-the-art results below 4$backslash$% and 19$backslash$%, respectively. Our source code is available at
Conference Paper
We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 dif- ferent classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0% which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully-connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implemen- tation of the convolution operation. To reduce overfitting in the fully-connected layers we employed a recently-developed regularization method called dropout that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry