• No results found

Application of Deep Learning Approaches in SAR ATR


Academic year: 2020

Share "Application of Deep Learning Approaches in SAR ATR"


Loading.... (view fulltext now)

Full text


2018 International Conference on Computer Science and Software Engineering (CSSE 2018) ISBN: 978-1-60595-555-1

Application of Deep Learning Approaches


Zhilong Lin, Changlong Wang and Yongjiang Hu



In the field of Synthetic Aperture Radar (SAR) image interpretation, the Automatic Target Recognition (ATR) has always been the focus and hotspot in this field, which is also the research difficulty in this field. The SAR ATR is generally composed of three steps: extraction of Region of Interest (RoI), target identification, and target classification. The complex flow not only limits the efficiency of SAR ATR, but also makes the overall optimization of the model difficult to be carried out, which restricts the accuracy. This paper discusses Faster R-CNN and SSD adopted from the computer vision and shows how those approaches enable significantly improved performance for SAR ATR. The validity of the deep learning method in the field of automatic target recognition for SAR images and the nature of the university are verified, which lays the foundation for further research.


Synthetic Aperture Radar (SAR) is widely used in military and civilian fields due to its all-weather observational features. It is urgent to propose high-performance SAR ATR technology. Compared with ordinary optical images, SAR images have great differences in imaging mechanism, geometric characteristics and radiation characteristics. SAR image has the following

Zhilong Lin, Changlong Wang and Yongjiang Hu


characteristics: First, the target edge blurred and image resolution is low caused by the uneven reflection of radar wave. Second, the noise in SAR images is larger, which has a great impact on the results of target detection. Thirdly, the SAR image is not sensitive to the intensity of the echoes of the ground objects so that the layering is worse. The target detection and recognition methods for SAR images are generally composed of several independent steps, such as filtering, segmentation, feature extraction and target recognition. Feature extraction as a key link, the ability to extract the target feature is the key to target recognition.

There are already many methods for feature extraction of SAR images. Zhao et al. [1] uses the original image pixel values directly as input to the Support Vector Machine (SVM) classifier. In [2], 2-D Discrete Fourier Transform (DFT) coefficients of the original image are extracted as recognition features. In [3], the global scattering center model is used to extract the scattering center features for SAR target recognition. Park et al. [4] proposed 12 kinds of discrimination features based on the target pixel and the projection length. Wang et al. [5] presents a novel texture feature extraction method based on a Gabor filter and Three Patch Local Binary Patterns (TPLBP) for Synthetic Aperture Rader(SAR) target recognition. However, all of the above methods use features designed based on domain expert knowledge rather than features learned automatically from the data. With the improvement and application of different resolution imaging radars in different wavebands, how to automatically extract the features that benefit the target recognition from the imaging data has become an important issue.

At present, the target detection algorithm based on convolutional neural network merges extraction of RoI, target identification, and target classification into deep learning framework and forms an end-to-end network model. The target detection algorithms based on deep learning are mainly divided into two categories, one is proposal based methods including R-CNN [6], Fast R-CNN [7], Faster R-CNN [8] and R-FCN [9], and the other is proposal-free methods like YOLO [10] and SSD [11]. It is of great significance to use the target detection framework based on deep learning for the SAR ATR in order to improve the accuracy.



Figure 1. Faster-R-CNN structure diagram.

Faster R-CNN as a representative of these methods for SAR ATR experimental studies.

The Faster-R-CNN is proposed, based on R-CNN and Fast R-CNN, and the four basic steps of proposal based methods (candidate region generation, feature extraction, classification and Bounding Box regression) are unified into one network. The accuracy of the Faster-R-CNN is slightly higher than that of the Fast R-CNN, but the detection speed is increased by a factor of 10. Faster-R-CNN target detection network structure shown in Figure 1:

The Faster-R-CNN candidate region was generated using the Region Proposal Network (RPN) instead of the Selective Search (SS) of the Fast R-CNN. Creatively adopting the RPN to generate proposal areas, and sharing the convolutional network with the target detection network, the proposed region is reduced from 2,000 to 300, and the quality of the proposed area has also been substantially improved. Algorithm flow is as follows:

Step 1 Get the feature maps by the shared feature extraction network

Step 2 Feature maps are sent to the RPN network and the CNN respectively. Step 3 RPN network generates the location of the region and the probability of belonging to the foreground and the background according to the feature maps.

Step 4 The higher dimensional features which sent to RoI pooling layers are obtained by the CNN.

Step 5 The RoI pooling layer transform the higher dimensional features to the same size.

Step 6 The same-size feature maps are sent into the fully-connected layer and finally output the target classification and bounding bound-box locations.



locations. SSD have some improvements compared with YOLO, including (1) using pyramid features for prediction at different scales; (2) using default boxes and aspect ratios for adjusting varying object shapes; (3) using small convolutional filters to predict categories and anchor offsets for bounding box locations using default boxes and aspect ratios for adjusting varying object shapes. This paper used SSD as a representative of proposal-free methods for SAR ATR experimental studies.

SSD consists of two parts, one part is the basic network, used to extract the characteristics of the input image. The other part is an additional network, which performs the classification of the target and the regression of the bounding box positions on the feature maps of different scales while convoluting the feature maps extracted from the basic network for extracting more advanced features. SSD structure shown in Figure 2.

Different scales of the characteristics of different receptive field, is conducive to the detection of different size targets. The characteristics of the target reflected by the different scale features are different, which is helpful for the accurate classification of the target. SSD target detection framework of the algorithm flow is as follows:

Step 1 First, we use the common convolutional neural network structure, such as VGG-16 as the base network to extract the feature maps from SAR image.

Step 2 Based on the feature maps extracted by the base network, the feature map of different scales is obtained through the additional network.

Step 3 Using different boxes with different aspect ratios at different pixel location on different scales feature maps to extract features.

Step 4 The position information and the target feature extracted from each pixel position of all the different scale feature maps are respectively used for the regression of the bounding box and the classification of the target through a convolutional neural network.



Figure 2. SSD structure diagram.

Figure 3. Ten-class military targets: optical image vs. SAR image.

The target evaluation index is mainly to evaluate the target detection framework of detection speed and accuracy. Detection speed Measured by the number of images per second (frames) to measure the real-time target detection framework. The detection accuracy is measured by Average Precision (AP). The average accuracy is calculated as follows:


0 ( )


P R dR (1)


SAR ATR Based on Faster R-CNN


On the basis of the VGG-16 network structure and pretrain model, Faster-R-CNN is used to train on the MSTAR database. The accuracy and results of SAR ATR are shown in Table I and Figure 4:



Figure 4. Faster R-CNN detection results.


Figure 5. SSD detection results.


SAR ATR Based on SSD

Based on the VGG-16 network structure and pretrain model, SSD is used to train on the MSTAR database. The accuracy and results of SAR ATR are shown in Table II and Figure 5.

The mean AP of SAR ATR based on SSD reach 95.36% and the detection speed is up to 31 pictures per second.


This paper used the target detection framework based on deep learning for the SAR ATR. The experiment results based on MSTAR show that the mean AP of SAR ATR based on Faster R-CNN and SSD reach 96.51% and 95.36% respectively. The speed of detection is 16 pictures per second and 31 pictures per second. Both Faster R-CNN and SSD can perform SAR ATR well. Faster R-CNN has higher precision, and SSD has faster speed, and it needs to be weighed in practical applications. Although the experiment has some limitations, this paper verified the effectiveness of applying the deep learning target detection framework to SAR ATR to some extent, which lays the foundation for further research in this direction.


This work was financially supported by National Nature Foundation of China (no. 51307183).


1. Zhao Q, Principe J C. Support vector machines for SAR automatic target recognition[J]. IEEE Transactions on Aerospace & Electronic Systems, 2001, 37(2):643-654.

2. Sun Y, Liu Z, Todorovic S, et al. Adaptive boosting for SAR automatic target recognition[J]. IEEE Transactions on Aerospace & Electronic Systems, 2007, 43(1):112-125.

3. Zhou J, Shi Z, Xiao C, et al. Automatic Target Recognition of SAR Images Based on Global Scattering Center Model[J]. IEEE Transactions on Geoscience & Remote Sensing, 2011, 49(10):3713-3729.

4. Park J I, Park S H, Kim K T. New Discrimination Features for SAR Automatic Target Recognition[J]. IEEE Geoscience & Remote Sensing Letters, 2012, 10(3):476-480.

5. Wang L, Zhang F, Li W, et al. A Method of SAR Target Recognition Based on Gabor Filter and Local Texture Feature Extraction[J]. Journal of Radars, 2015, 4(6).


7. Ren S, He K, Girshick R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE transactions on pattern analysis and machine intelligence, 2017, 39(6): 1137-1149.

8. Ren S, He K, Girshick R, et al. Faster r-cnn: Towards real-time object detection with region proposal networks[C]//Advances in neural information processing systems. 2015: 91-99.

9. Dai J, Li Y, He K, et al. R-fcn: Object detection via region-based fully convolutional networks[C]//Advances in neural information processing systems. 2016: 379-387.

10. Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 779-788.

11. Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//European conference on computer vision. Springer, Cham, 2016: 21-37.


Figure 1. Faster-R-CNN structure diagram.


Related documents

The measurements from the IMB 2012G, 2012H, 2012J, and 2012L, collocated with AMSR2 TBs, are used as the training data set for the different regressions to retrieve snow depth and

Methods: We identified the fetal sex on cffDNA extracted from 29 maternal plasma samples at early gestational ages, several of them not suitable for qPCR determination, using

Our results showed that oral signs and symptoms and blood examination abnormalities were more common and also more severe in patients with a history of gas- trectomy.. Resection of

On review of research literature in the area of reflective practice, particularly reflective journals, and consideration of existing learning activities and resources available in

This special report reviews the safety of the nuclear testing program and assesses the technical procedures used to test nuclear weapons and ensure that radioactive material produced

had carried several direct shear tests on clayey soil reinforced with 0.4% polypropylene fibers by weight of soil and reported that cohesion and internal friction angle of

Saranya et al , International Journal of Computer Science and Mobile Applications, Vol.7 Issue.. The second earth station conjointly communicates with the primary

Fakat, iyileşme sağlayacak kadar güçlü olan herhangi bir ilacın yanlış kullanılması durumunda, zehirlenmeye de neden olabilmesi gibi, reenkarnasyon öğretisi de