SPSNet: A Selected Pyramidal Shape-Constrained Network for SAR Target Detection

The complex background and coherent speckle noise in synthetic aperture radar (SAR) images presents a significant challenge for the detection and recognition of SAR small targets. For deep neural networks, the robust feature learning method and effective loss function could enhance the accuracy of SAR target detection and reduce false alarm rates. However, many of feature enhancement networks based on feature pyramid network (FPN) have limited ability to capture feature interaction between different branches. In addition, the design of loss function cannot generate samples that better match the shape of SAR targets for network training. In this article, we propose a selected pyramidal shape-constrained network (SPSNet) to alleviate these problems. A feature fusion paradigm, including a spatial selection block and a dynamic channel attention module, are inserted into FPN for adaptive multiscale feature selection and feature enhancement in spatial-channel feature dimension. Both of these modules could capture the distinguishable features of SAR targets. Furthermore, the shape information of SAR target is utilized into detection loss to enhance the quality of SAR detection box sampling points in a soft threshold style, thereby enhancing the model's detection for SAR targets. The experimental results of three challenge SAR detection datasets illustrate that SPSNet gains superior performances.