FESAR：Detection Model Based on Local Spatial Relationship Capture and Fused Convolution Enhancement

Abstract

Abstract Synthetic Aperture Radar (SAR) plays a crucial role in ship monitoring due to its all-weather and high-resolution capabilities. In SAR images, ship targets often exhibit blurred or mixed boundaries with the background, and there may be occlusion or partial occlusion. Furthermore, the multi-scale transformation and the presence of small targets pose challenges to ship detection. To address these challenges, a novel SAR ship detection model, FESar, is proposed. First, to address the problem of large-scale transformations in ship detection, a network FCEM with fused convolutional enhancement is proposed, in which different convolutional branches are designed to capture local and global features, respectively, and are fused and enhanced. Secondly, an FPE module containing a spatial-mixing layer is designed to capture and analyze local spatial relationships in the image, and effectively combine local information to discriminate the feature information between ship targets and the background. Finally, a new backbone network, SPD-YOLO, is designed to perform deep downsampling for the comprehensive extraction of semantic information related to ships. To validate the performance of the model, experiments are conducted on the publicly available dataset LS-SSSDD-v1.0, and the experimental results show that the performance of the proposed FESar model outperforms many SOTA models, and based on the base model, FESar improves the AP by 5.5% on the dataset LS-SSDD-v1.0. Compared with the SAR ship detection model on the SSDD dataset, the comprehensive performance of FESAR is better than other SAR ship detection models. To verify the generalization of the model, we experiment with FESAR with numerous SOTA models on the dataset HRSID, and the experimental results show that, based on the base model, the FESAR model improves AP by 2.6% on the dataset HRSID.