SAR Target Detection and Recognition Based on Pixel-Level Prior Knowledge and Cosine Similarity Discrimination

Synthetic Aperture Radar (SAR) automatic target recognition (ATR) technology is crucial for developing informatized and intelligent military and civilian monitoring systems. In response to the inexplicability of deep learning-based recognition methods under the same scene and target conditions, this paper proposes a target detection and recognition method based on prior knowledge and cosine similarity discrimination. During the detection phase, the differences in the pixel value distributions between the target and background are leveraged to extract potential target regions from panoramic images using prior knowledge. Subsequently, cosine similarity is employed to filter the predicted targets. Finally, the Convolutional Block Attention Module (CBAM)-ResNet50 classification network enables sufficient learning of target detail features. The experimental results show that the proposed method can effectively utilize prior target information and detail features to enhance detection accuracy, achieve physical interpretability of the detection process, and avoid positional dependency on the detection network.