MSARN: A Deep Neural Network Based on an Adaptive Recalibration Mechanism for Multiscale and Arbitrary-Oriented SAR Ship Detection

Abstract

Ship detection plays an important role in synthetic aperture radar (SAR) image interpretation. However, there are still some difficulties in SAR ship detection. First, ships often have a large aspect ratio and arbitrary directionality in SAR images. Traditional detection algorithms can cause the detection area to be redundant, which makes it difficult to accurately locate the target in complex scenes. Second, ships in ports are often densely arranged, and the effective identification of densely arranged ships is complicated. Finally, ships in SAR images exist at a variety of scales due to the multiresolution imaging modes used and ship shape variations, which pose a considerable challenge for ship detection. To solve the above problems, we propose a multiscale adaptive recalibration network (MSARN) to detect multiscale and arbitrarily oriented ships in complex scenarios. The recalibration of the extracted multiscale features through global information increases the sensitivity of the network to the target angle, thereby increasing the accuracy of positioning. In particular, we designed a pyramid anchor and a loss function to match the rotated target. In addition, we modified the rotation non-maximum suppression (RNMS) method to solve the problem of the large overlap ratio of the detection box. The proposed model combines the positioning advantage of rotation detection with the speed advantage of a single-stage framework. Experiments show that based on the SAR rotation ship detection (SRSD) data set, the proposed algorithm has a faster detection speed and higher accuracy than some state-of-the-art methods. INDEX TERMS Ship detection, synthetic aperture radar (SAR), adaptive recalibration, neural network.

End‐to‐End Ship Detection in SAR Images for Complex Scenes Based on Deep CNNs

“…The SAR ship detection dataset which is called SSDD is established in 2017 to set the baseline of SAR ship detection algorithms [30]. It is publicly released and has been used by many other scholars [31,[33][34][35][36][38][39][40]. The SSDD data set contains various ships in multitudinous scenes, including different sensors, resolutions, polarizations, scenes, and work modes.…”

“…Compared with some other detection methods implemented on SSDD as shown in Figure 16, such as Reference [30], the detection speed is about 3 FPS, Reference [36] is about 35 FPS, and Reference [39] is about 48 FPS. Our methods have achieved a faster speed of 72 FPS than all the above methods.…”

“…Compared to YOLOv1, YOLOv2, and YOLOv3 [40], the mAP is also improved by 14%, 5%, and 0.2%, respectively. At the same time, some novel and superior methods are also used for comparisons, such as the network based on adaptive recalibration mechanism (MSARN) [36], dense attention pyramid networks (DAPN) [33], grid-based CNN (G-CNN) [39], and depthwise separable CNN (DS-CNN) [40]. The results show that our method has better performance than all the above methods and achieves a big improvement in the final mAP.…”

“…These novel network structures can achieve higher final detection accuracy but increase the computational complexity to some extent. At the same time, some methods are concerned about more precise location and use the rotatable bounding boxes to detect which could locate the ships more finely [35,36]. Besides, some novel training techniques such as training from scratch are also adopted in the SAR field, and the final results also outperform other pretrained ship detectors [37].…”

Learning Polar Encodings for Arbitrary-Oriented Ship Detection in SAR Images

“…Chen et al [52] built a feature alignment module to extract the features of the ship targets more accurately based on oriented anchors. Chen et al [53] designed multilayer anchors and rotation non-maximum suppression postprocessing to improve the detection performance for oriented ship targets. Pan et al [54] used rotating region proposal network (RRPN) to extract candidate target regions, and then multi-layer cascade network was employed to fine tune the OBB detection results.…”