PCENet: A Method for SAR Ship Detection Based on Pixel Confidence Estimation

Various target detection networks based on bounding box parameter regression have been proposed, but several challenges remain in maritime target detection tasks: 1) Limited detection box scale fails to accommodate ships' multi-pose nature; 2) In marine surface Synthetic Aperture Radar (SAR) images, ships occupy only a small portion of the scene, and the excessive number of negative samples results in a degradation of the network's detection capability; 3)Due to the limited quantity of SAR data, the network's generalization ability is compromised, especially affecting its robustness to noise. To address these issues, the Pixel Confidence Estimation Network (PCENet), based on a Piecewise Mean Square Error (PMSE) loss function, is proposed. Instead of directly regressing bounding box parameters, PCENet estimates pixel- wise probabilities of target presence, mitigating pose-related performance degradation. The method leverages the amplitude distribution characteristics of maritime SAR images to design the PMSE loss function, which enhances the network's ability to discern strong scatter points while effectively alleviating the performance degradation caused by sample imbalance. By incorporating the estimation error of pixels within the same amplitude range into the loss function, the network's noise robustness is significantly improved. Experiments conducted on the dataset demonstrate a significant improvement in the proposed algorithm's confidence estimation ability and noise robustness. Additionally, compared to the lightweight network You Only Look Once (YOLO) version 5 s, the parameter model achieves a reduction of over 20%. This makes the method promising for deployment on miniaturized platforms.