关键词: 合成孔径雷达, 舰船目标斜框检测, 深度学习, 多尺度, 特征增强

Abstract: Aiming at the problem that the parameter quantity in the current synthetic aperture radar (SAR) image ship target oblique frame detection algorithm is too large to meet the real-time detection requirements and the problem of boundary discontinuity, a lightweight SAR ship target oblique frame detection algorithm based on polar coordinate encoding and decoding is proposed. Firstly, based on the shuffle unit of ShuffleNetV2 and using lightweight and efficient PC and Ghost convolutions, a convolution collaborative unit is proposed to optimize convolution operations and reduce algorithm complexity. Then, the polar coordinate encoding and decoding method is introduced, and the cosine-harmonized IOU weighted loss function is proposed to dynamically adjust the polar coordinate encoding loss to solve the oblique frame detection boundary discontinuity problem; at the same time, the polar coordinate encoding and decoding method is improved by using double-peak max-pooling and elliptical two-dimensional Gaussian distribution to enhance the detection accuracy for the densely packed ships near shore. Finally, a multi-scale bar convolution attention module and a spatial selection feature enhancement module are proposed to obtain feature information at different scales to improve the network feature extraction capability. Experimental results on the SSDD+ and DSSDD datasets show that the algorithm achieves 85.7% and 90.6% detection accuracy, respectively, and a generalization test on the HRSID dataset with a detection accuracy of up to 82.3%.Compared with similar algorithms with similar detection accuracy, the parameter quantity and floating-point operations are less than one-tenth of similar algorithms, only 0.87MB and 1.21GB, and the detection speed is increased by 14%, up to 135FPS, meeting the real-time detection requirements.

Key words: synthetic aperture radar, ship target oblique frame detection, deep learning, multi- scale, feature enhancement