A Novel Salient Feature Fusion Method for Ship Detection in Synthetic Aperture Radar Images

Ship detection of synthetic aperture radar (SAR) images is one of the research hotspots in the field of marine surveillance. Fusing salient features to detection network can effectively improve the precision of ship detection. However, how to effectively fuse the salient features of SAR images is still a difficult task. In this paper, to improve the ship detection precision, we design a novel one-stage ship detection network to fuse salient features and deep convolutional neural network (CNN) features. Firstly, a saliency map extraction algorithm is proposed. The algorithm is applied to generate saliency map by using multi-scale pyramid features and frequency domain features. Secondly, the backbone of the ship detection network contains a two-stream network. The upper-stream network uses the original SAR image as input to extract multi-scale deep CNN features. The lower-stream network uses the corresponding saliency map as input to acquire multi-scale salient features. Thirdly, for integrating the salient features to deep CNN features, a novel salient feature fusion method is designed. Finally, an improved bi-directional feature pyramid network is applied to the ship detection network for reducing the computational complexity and network parameters. The proposed methods are evaluated on the public ship detection dataset and the experimental results shows that it can make a significant improvement in the precision of SAR image ship detection.


I. INTRODUCTION
With the expansion of maritime rights and interests, it has become particularly important to have a comprehensive, timely and accurate grasp of ocean dynamics. Compared with optical imaging systems, synthetic aperture radar (SAR) has all-day and all-weather surveillance capabilities, making it possible to continuously monitor targets at sea [1], especially ship surveillance. However, with the application of high-resolution SAR, ship features appearing in images are more complex, making it more difficult to detect ships quickly. Therefore, it is very important to study how to quickly realize the ship detection in SAR images.
After years of development, a large number of ship detection algorithms have been proposed. These algorithms can be classified into two categories. One is the traditional ship The associate editor coordinating the review of this manuscript and approving it for publication was Weimin Huang . detection algorithms, the other is the ship detection networks based on deep convolutional neural network (CNN).
In the traditional ship detection algorithms, the constant false alarm rate (CFAR) algorithm is the most in-depth researched and it is widely used in actual ship detection systems [2], [3]. There are variants of CFAR algorithms, such as bilateral CFAR [4], intensity-space CFAR [5], superpixel-based CFAR [6], proposal CFAR [7], variable window CFAR [8] and pixel CFAR (SUMO) [2]. However, CFAR algorithm is a sub-optimal detection method and complicated and cluttered backgrounds severely affect CFAR detection performance.
The ship detection networks based-CNN include two-stage ship detection networks [10]- [15] and one-stage ship detection networks [16]- [21]. The two-stage ship detection network is a coarse-to-fine architecture. The first stage is the coarse detection (region proposal) stage, which extracts regions where ships may exist. The second stage is the fine detection stage, which classifies and regresses the regions reserved from first stage to obtain the category probability and target position. The two-stage detection networks have high precision and slow speed. The main reason is that the region proposal stage is very time-consuming. To improve the detection speed, the one-stage network abandons the region proposal stage and directly obtains the class probability and target position. Although the one-stage detection network has very high detection speed, it is very difficult to train [22] and its detection precision is lower than two-stage networks.
In order to improve the detection precision of the one-stage network, a lot of tricks are applied [18], [22], [23]. Adding salient features to the network has proven to be the most effective trick. The salient features can be obtained by using attention model and saliency map. To use salient features more effectively, most of them use concat and add fusion methods. Du et al. [18] and Li et al. [24] also studied feature fusion, which called Du fusion and Li fusion. All of the above fusion methods can improve the performance of network. However, the concat method will increase the spatial complexity and network parameters. The add method is easy to change the distribution of features and even cause some features to disappear. The Du fusion will sacrifice the diversity of features. The Li fusion is mainly aimed at the fusion of aesthetic features and it is not suitable for ship detection in SAR images.
In this paper, we propose a novel one-stage detection network for fusing salient features and deep CNN features. The network is composed of a saliency map extraction (SME) algorithm and a ship detection network (SAR-ShipDetNet). SME algorithm joints multi-scale pyramid features and frequency domain features to generate saliency maps. The backbone of SAR-ShipDetNet is a two-stream network. The upper-stream is a deep feature network (DFNet), which is used to extract multi-scale deep CNN features. The input of DFNet is the original SAR image. The lower-stream is saliency feature network (SFNet) for extracting multi-scale salient features. The input of SFNet is a corresponding saliency map. Unlike the S-SSD [18], SFNet has a very simple structure with 5 mobile inverted bottleneck convolution (MBConv) layers. For integrating salient features to the deep CNN features without adding feature dimensions and network parameters, we design a novel salient feature fusion method. In addition, to reduce the computational complexity and parameters of the network, the complex bi-directional feature pyramid network (BiFPN) is improved. It can ensure that the network obtains better ship detection performance with fewer input features. Some examples of the proposed methods for SAR ship detection are displayed in Fig. 1, where red rectangles, green ellipses and yellow ellipses indicate the detected ships, missing ships and false alarms, respectively. Our contributions can be summarized as follows: • A novel one-stage ship detection network is proposed to fuse salient features and deep CNN features. The network mainly includes a saliency map extraction algorithm and a SAR-ShipDetNet. • Aiming at the characteristics of SAR images, a novel saliency map extraction algorithm is designed. It combines multi-scale pyramid features and frequency domain features to generate saliency maps.
• For integrating salient features to deep CNN features without adding feature dimensions and network parameters, a novel salient fusion method is proposed.
• To reduce the computational complexity, the bi-directional feature pyramid network is improved.
• Extensive validation experiments are designed on SARship-detection dataset [35]. The results show that the proposed methods can make a big leap forward in improving the performance of SAR image ship detection.
The rest of this paper is organized as follows. Section II briefly reviews previous works that are closely related to our method. In Section III, the proposed methods are illustrated in detail. A detailed datasets and experiments results are presented in Section IV. Conclusions are given in Section V.

II. REALTED WORK
Ship detection of SAR images has always been a research hotspot in the field of marine surveillance. With the application of high-resolution SAR satellites, traditional CFAR ship detection algorithms have been unable to meet actual detection requirements. The CNN-based ship detection network has been attracted by many researchers. The existing CNN ship detection networks are mainly divided into two-stage networks [10]- [15] and one-stage networks [16]- [21]. Li et al. [10] first designed multi-scale feature fusion on Faster-RCNN [9]. To solve multi-scale and multi-scene ship detection, Jiao et al. [11] proposed a densely connected multi-scale neural network based on Faster R-CNN. Kang et al. [12] combined Faster-RCNN with CFAR algorithm to improve the detection performance. Gui et al. [13] designed a multi-layer fusion light-head network, which used the fusion features of shallow high-resolution and deep semantic feature to generate region proposals. Zhao et al. [14] proposed an exhaustive ship proposal network and an accurate ship discrimination network for solving dense small ship detection in SAR images. Wang et al. [15] designed a hierarchical CNN to detect ships VOLUME 8, 2020 in spaceborne SAR images. Chang et al. [16] proposed a ship detection network based on you only look once version 2 (YOLOv2), which was called YOLOv2-reduce. YOLOv2reduce had fewer layers without affecting the precision of ship detection. Wang et al. [17] used data augmentation and transfer learning to realize the ship detection of SAR images in the single shot multi-box detector (SSD). Du et al. [18] proposed a saliency-guided SSD (S-SSD), which used salient features to guide the deep CNN features. Due to the difference of ship scales and the interference of inshore scenes, Fu et al. [19] designed a feature balancing and refinement network. To solve the problem of detecting small ships in SAR images, Cui et al. [20] designed a dense attention pyramid network, which made the features extracted by the network contain rich resolution and semantic information. In order to reduce the size of the ship detection model, Zhang and Zhang [21] proposed a lightweight ship detector that used feature fusion module, feature enhance module and a scale share feature pyramid module to improve detection performance. The two-stage detection networks have higher precision and slower speed than the one-stage networks.
To improve the detection precision of the one-stage networks, adding salient features to the network has proven to be the most effective trick [18], [23]. The S-SSD [18] used two identical sub-networks to extract features from the original SAR image and the corresponding saliency map at the same time. Then, the salient features were integrated to the deep CNN features. Zhang et al. [23] introduced a channel attention model and a spatial attention model in the high-speed and high-precision SAR ship detection network and obtained very excellent detection performance with a very light detection model. It is well known that the quality of the saliency information of the image will directly affect the detection precision. To integrate salient features to the deep CNN features, concat and add fusion methods are widely used [33], [35]. The concat method is to splice features in dimensions and will increase the spatial complexity and network parameters. The add fusion method is to add two tensors point by point. For example, two features, ''−0.766'' and ''+0.766'', are fused by the add method. The fused feature is ''0''.This fusion leads to the disappearance of features. Du et al. [18] and Li et al. [24] also studied feature fusion, which called Du fusion and Li fusion. When using Du fusion, the salient features are normalized. Then, the normalized salient feature is element-wise multiplied to the deep CNN features. The Li fusion is that the edge features of the deep CNN features are replaced by aesthetic features.
In this article, we design a novel salient feature fusion method, the dimensionality of the fused feature remains unchanged, and the salient features will not make the deep CNN features value disappear. The details of proposed fusion method are described in Section III-B.

III. PROPOSED METHOD
The proposed one-stage ship detection network fuses the salient features and deep CNN features for improving the  ship detection precision in SAR images. The one-stage detection network is mainly composed of a SAR-ShipDetNet and a saliency map extraction (SME) algorithm, which will be introduced in Section III-A and Section III-B, respectively. The proposed salient feature (SF) fusion method and improved BiFPN in SAR-ShipDetNet will be explained in Section III-C and Section III-D, respectively. Fig. 2 shows the detailed architecture of one-stage ship detection network, where SME is the proposed saliency map extraction algorithm and SAR-ShipDetNet is the ship detection network for SAR images. D0-D7 are the output features of deep features net (DFNet). S1-S5 are the output features of saliency features net (SFNet). SF1-SF3 are the proposed salient feature fusion methods. P4-P6 are the fused features and C3-C6 are the detection headers. IBiFPN represents the improved BiFPN. Conv and FC means convolution layer and fully connected layer, respectively.

A. THE ARCHITECTURE OF SAR-ShipDetNet
The SAR-ShipDetNet has a two-stream network. The upper-stream network is called DFNet for extracting multi-scale deep CNN features. Its input is the original SAR image with 512 × 512. The output features of DFNet are D0-D7. The feature sizes of D0-D7 are 256 × 256, 256 × 256, 128 × 128, 64 × 64, 32 × 32, 32 × 32, 16 × 16 and 16 × 16, respectively. The lower-stream is the SFNet for extracting multi-scale salient features. The input of SFNet is the corresponding saliency map with 512 × 512, which is obtained by using SME algorithm. The output features of SFNet are S1-S5. The feature sizes of S1-S5 are 256 × 256, 128×128, 64×64, 32×32 and 16×16, respectively. Table 1 and Table 2 respectively show the detailed configuration of DFNet and SFNet, where Stage represents the feature name. The Operator, In, Out and L represent the operation, input features, output features and the number of operator, respectively. The h, w and c indicate the height, width and channels.
The features of D3 and S3 are feed into SF1 to obtain fused feature P4. The fused feature P5 is generated by fusing The detection headers are feed into class prediction net and box prediction net to obtain the class probability and ship position. Each class prediction net and box prediction net is composed of two layers of 3 × 3 convolution (Conv) and a fully connected (FC) layer. It is worth noting that each layer of MBConv and Conv layers follows a batch-normalization layer and a swish activation function.

B. SALIENCY MAP EXTRACTION ALGORITHM
In this paper, the proposed saliency map extraction algorithm for SAR images divides the process into three sub-tasks, which are speckle noise suppression, image segmentation and saliency map extraction. Firstly, to reduce the influence of speckle noise in SAR images, refined Lee filter [26] is applied to suppress speckle noise. Secondly, to segment the despckled SAR image, the K-means clustering algorithm is used. Finally, the final saliency map of the SAR image is obtained by using segmented image.

C. SALIENT FEATURE FUSION METHOD
Concat and add fusion methods are widely used in feature fusion. They have been integrated to many deep learning frameworks, such as PaddlePaddle, TensorFlow and Pytorch. The concat method is to splice two or more feature maps in the channel or number dimensions. The spliced features increase channels or the number dimensions. The add method completes fusion by performing an element-wise add operation on the two feature maps. It is not difficult to find that the add operation will not increase the dimensionality of the features. Even then, the add method will destroy the distribution of deep CNN features. Du fusion method [18] is only used in S-SSD and it destroys the diversity of deep CNN features. For integrating salient features to the deep CNN features without adding feature dimensions and destroying features diversity, a novel salient feature (SF) fusion method is designed.
where Am(·) and Ph(·) represent taking amplitude and phase operator, respectively; 8: where log is a logarithm operator; 9: The f o (i, j), f c (i, j) and f s (i, j) represent the fused feature, deep CNN feature and the salient feature at the position (i,j), respectively. λ is a weight with the range (0, ∞). In Eq. 1, for non-target regions, the deep CNN features are remained the same. For the target regions, the salient features with weights are element-wise multiplied to the deep CNN features for enhancing the deep CNN features. The purpose is not only to enhance the attention of network to the target regions, but also to ensure the diversity of the deep CNN features.

D. IMPROVED BiFPN
The feature pyramid network (FPN) is used to extract multi-scale features of images and improve the detection performance of the network for different object sizes. It has experienced no fusion, top-down fusion, simple bi-directional fusion and complex bi-directional fusion. Bi-directional feature pyramid network (BiFPN) [49] has the best performance in object detection. BiFPN is a complex bi-directional fusion pyramid feature network and its inputs are 5-level features. The output of BiFPN is 5-level features. The smallest feature is 4 × 4, equivalent to the image with 512 × 512 being down-sampled 128 times. However, most ship target sizes are smaller than 64 × 64 in SAR images. Therefore, we have improved the structure of BiFPN.
The detailed framework of improved BiFPN is shown in Fig. 4, where P1-P6 represent the feature pyramid extracted by the backbone. C3-C6 are the detection headers. The meaning of color blocks and arrows in Fig. 4 are marked on the lower right. The input of improved BiFPN is 3-level features and the output is 4-level features. There are two differences between improved BiFPN and BiFPN. The one is that the input of BiFPN is a 5level features, while the input of improved BiFPN is a 3-level features. The second is that the output of BiFPN is 5-level features, while the output of improved BiFPN is 4-level features. The purpose of such improvement is to reduce network parameters and calculation time while maintaining BiFPN performance.

IV. DATASETS AND EXPERIMENTAL RESULTS
In this section, the datasets and the experimental results will be explained in detail. The experiments mainly include two parts. The first is the verification of SME algorithm and the second is the evaluation of one-stage ship detection network. In the verification of SME algorithm, we have displayed comparison results both subjective metrics and objective vision. In the evaluation of one-stage ship detection network, it mainly includes the selection of backbone, the improved BiFPN and our fusion method. The standard COCO metrics [34] are used to evaluate the performance of detection networks. For fair comparison, all the experiments are trained and evaluated on an Hewlett-Packard (HP)workstation Z4 with an NVIDIA RTX2080Ti GPU.

A. DATASETS
The public SAR-Ship dataset (SDD) [35] is used to train and test the performance of the proposed methods. The SDD is generated from 102 Gaofen-3 images and 108 Sentinel-1 images. The spatial resolutions of Gaofen-3 in range and azimuth are 3m×3m, 5m×5m, 8m×8m and 10m×10m, respectively. Its imaging modes are ultrafine strip map, fine strip map-1, full polarization-1, full polarization-2 and fine strip map-2, respectively. For Sentinel-1, the spatial resolutions are 20m×22m and 1.7m×4.3m to 3.6m×4.9m. The imaging modes of Sentinel-1 are S3 strip map, S6 strip map and interferometric wide swath.
The SDD has 43,819 ship chips and 59,535 ship targets in total. The pixels of each image is 256 × 256. The ship targets of each image are marked and the marking style is PASCAL VOC. The statistical distribution of the ship sizes is listed in Table 3, where ''Size'', ''Min'' and ''Max'' represent ship pixels, minimum ship size and maximum ship size, respectively. ''Count'' represents the total number of ships. ''Percentage'' means the percentage of the ship in whole ship targets. It is worth noting that the definitions of ''small ship'', ''medium ship'' and ''large ship'' are the same as the large-scale SAR ship detection data et-v1.0 (LS-SSDD-v1.0) [36]. In the experiment, we randomly divided SDD into two small-scale datasets, named SDD-1 and SDD-2 respectively. The reasons for this are as follows. First, using small datasets can prevent overfitting of the model. Second, it can verify the performance of the proposed complex network on small-scale datasets. SDD-1 contains 21,945 images and SDD-2 has 21,874 images. To obtain training, validation and testing sets, SDD-1 and SDD-2 are split with the 7:2:1. The training set and the validation set are used for training models and the testing set is used for testing models.

B. PERFORMANCE OF SME ALGORITHM
To verify the performance of the proposed SME algorithm, we compared with 8 saliency map extraction algorithms in objective metrics and subjective vision. They are Itti et al. [31], frequency-tuned (FT) [30], luminance contrast (LC) [28], spectral residual (SR) [29], global contrast (GC) [32], F 3 Net [37], attentive feedback network (AFNet) [38] and Hou et al [40]. Itti, FT, LC, GC and SR are traditional saliency map extraction algorithms. F 3 Net, AFNet and Hou et al. are deep CNN saliency map extraction algorithms. In addition, since the SDD does not contain the ground truth of the saliency map, we made a testing set called SED. SED contains 300 pairs of images. Each pair of images is the SAR original image and the ground truth map. The production of SED is as follows. First, 300 images are randomly  selected from the SDD. Second, the corresponding ground truth maps are generated by using Labelme software [39].
The mean absolute error (MAE) [40], mean Fmeasure(mF) [41], structure similarity measure(S α , α = 0.5) [42] and E-measure(E ξ ) [43] are used to evaluate the SME algorithm. The comparison results of SME algorithm and other 8 algorithms are shown in Table 4. The bold indicates the better result. ↑ represents that larger is better and ↓ means that smaller is better. It can be seen from Table 4 that the proposed method has a significant improvement compared with the traditional algorithms and CNN methods, especially the mF reached 0.482. Fig. 5 shows the visual comparison between SME and other 8 algorithms. The first column is the original images. The second column is the ground truth maps. The third column is the results of the proposed SME algorithm. The fourth to eighth columns are the results of other 8 saliency map extraction algorithms.

C. PERFORMANCE OF SAR-ShipDetNet
In verifying the performance of SAR-ShipDetNet, all experiments are trained on PyTorch 1.5.0. During training, the image size is resized to 512 × 512. There are two reasons. Firstly, in the object detection task, it has been proved that the greater the resolution of the input image is, the better the detection performance is [33], [35]. Secondly, the backbone of the proposed network is EfficientNet-b0 [49]. The input image resolution of the EfficientNet-b0 is 512×512. Notably, the resized method of image is bilinear interpolation, which is widely used in the computer vision task. This interpolation method can effectively reduce the artifacts caused by image distortion during image scaling. VOLUME 8, 2020 In our experiments, the batch size is set to 16. The initial learning rate is set to 0.01 and the learning rate is reduced using the plateau decay rule. Each model is trained using Adam optimizer with β1 = 0.9 and β2 = 0.999. Same as [33], we use swish activation and commonly-used focal loss with α = 0.25 and γ = 1.5. Each model is trained 60 epochs. Notably, we do not use auto-augmentation for any of our models.

1) CHOOSING A BACKNONE FOR SHIP DETECTION
In objects detection tasks, the backbone is used as feature extractor and it will directly affect the detection performance. In recent years, various backbones have been proposed, such as ResNet, SENet, DetNet and EfficientNet.
From the experimental results, we found that the detection performance of EfficientNet-b0 is better than other 6 backbones. Therefore, EfficientNet-b0 was used as the backbone in the rest experiments.

2) PERFORMANCE OF IMPROVED BiFPN
From  Table 6, where ours is the results of using improved BiFPN. From the experimental results, it can be found that the improved BiFPN will not affect the detection performance of BiFPN. It can also reduce the network parameters and calculation time.
In addition, to demonstrate the influence of the input features, an ablation experiment is designed. We selected P3, P4, P5 and P6 to analyse the influence. The results are shown in Table 7, where Set 1 , Set 2 , Set 3 and Set 4 are the input features set of improved BiFPN. Set 1 is {P3, P4, P5}. Set 2 is {P3, P4, P6}. Set 3 is {P3, P5, P6}. Set 4 is {P4, P5, P6}. From the experimental results, it can be found that the performance of Set 4 is the best. The main reason is that when the other three sets are used, two consecutive up-sampling or down-sampling operations are used, which will destroy the consistency of features. Table 8 shows the comparison results of our, concat, add and Du fusion methods. The No means the experimental results without using salient features. The Add, Cat, Du and Ours refer that the fusion method is add, concat, Du fusion [18] and salient feature fusion method, respectively. The t denotes the detection time cost of different methods. From the experimental results in Table 8, it can be found that the    proposed fusion method has greatly improved the detection performance. Compared with No, the detection time cost of the proposed method only increases 0.015 s. In the experiment of our fusion method, the λ is set to 5.

3) PERFORMANCE OF SALIENT FEATURE FUSION METHOD
In Section III-C, we have introduced the salient feature fusion method in detail. The fused features are calculated according to Eq. 1. The λ is a hyper-parameter and needs to be set. In order to obtain the best performance of ship detection, λ is set to different values. In the experiments, we set λ to 1, 2, 5, 10, 20, 50 and 100,respectively. The results of different λ are shown in Table 9. It can be found that when the λ is set to 5, the performance of ship detection is the best.

D. COMPARISON WITH STATE-OF-THE-ART METHODS
In this section, we compared two-parameter CFAR (2p-CFAR) [8], Faster-RCNN [35], SSD-300 [35], SSD-512 [35], Modified SSD-300 [35], Modified SSD-512 [35], Mask-RCNN [50], YOLO-v3 [51], YOLO-v4 [52], YOLO-v4-tiny, YOLO-v5 [53] and EfficientDet-d0 [33] with ours on SDD. The experimental results are displayed in Table 10, where AP s 50 represents the average precision of small ship targets when IoU is 0.5, and AP 50 denotes the average precision when IoU is 0.5. The T (h) means the training time and FPS is defined as 1/t. The calculation formula of F1 is the same as the calculation formula in [36]. During training and testing, the SDD is split into training set, validation set and testing set at a ratio of 7:2:1. The hyper-parameters of proposed one-stage detection network is the same as in Section IV-C. The λ is set to 5 and improved BiFPN is used in SAR-ShipDetNet. The input features of improved BiFPN is {P4,P5, P6}. Fig. 6 shows precision-recall (PR) curves on different detection models. Notably, the experimental results of Faster-RCNN, SSD-300, SSD-512, Modified SSD-300 and Modified SSD-512 come from [35]. Therefore, their PR curves are not displayed in Fig.6. From the results we can obtain the proposed method has the best detection performance. Although the proposed method is slower than other methods (except Mask-RCNN) in detection speed, the proposed method has a significant improvement in accuracy. In addition, in the detection of small ships, the proposed methods can reach 75.35% on AP s 50 . The proposed method is 4.61% higher than YOLO-v5 on AP s 50 . The loss versus (vs.) epoch plots are given in Fig. 7.

V. CONCLUSION
In this paper, we proposed a novel one-stage ship detection network for fusing salient features and deep CNN features in SAR images. The network includes a SME algorithm and a SAR-ShipDetNet. SME algorithm combines multi-scale pyramid features and frequency domain features to generate saliency map. The backbone of SAR-ShipDetNet is composed of upper-stream and lower-stream networks. The upper-stream network is a deep features net, which used to extract multi-scale deep CNN features. Its input is the original SAR image. The lower-stream network is a saliency features net (SFNet) and its input is a corresponding saliency map for extracting multi-scale salient features. Unlike the S-SSD detector, the SFNet only has 5 MBConv layers. For fusing the salient features and the deep CNN features, the salient feature fusion method is designed. The fusion method can maintain distribution and diversity of deep CNN features without increasing the network parameters and feature dimensions. In addition, in order to reduce the computational complexity, the complex BiFPN is improved. The proposed methods are evaluated on SDD, and the results show that the proposed methods have a significant improvement in the precision of SAR image ship detection.
GANG ZHANG received the bachelor's degree from Xi'an Polytechnic University, in 2011, and the master's degree from Xidian University, in 2014. He is currently pursuing the Ph.D. degree with Space Engineering University, Beijing. His research interests include synthetic aperture radar imaging, image interpretation, targets detection, and deep learning.
ZHI LI received the B.E. and master's degrees from the National University of Defense Technology, in 1994 and 1997, respectively, and the Ph.D. degree from the Institute of Geology China Earthquake Administration, in 2003. He is currently a Professor with Space Engineering University. He has authored or coauthored more than 60 articles and 12 books. He holds ten invention patents and ten software copyrights. His research interests include space system applications and artificial intelligence.
XUEWEI LI received the bachelor's and master's degrees from the Xi'an University of Technology. She is currently pursuing the Ph.D. degree with the Beijing University of Posts and Telecommunications. Her current research interests include image aesthetic assessment, image processing, and machine learning.
CANBIN YIN is currently an Associate Professor with Space Engineering University. He has participated in more than ten research projects, including the National Natural Science Foundation of China and the National Defense 973. He has published more than 30 articles and four monographs. He also holds more than ten national invention patents. His research interests include teaching and scientific research on space-based situational awareness, reconnaissance, and detection.