A Spatial Cross-Scale Attention Network and Global Average Accuracy Loss for SAR Ship Detection

A neural network-based object detection algorithm has the advantages of high accuracy and end-to-end processing, it been widely used in synthetic aperture radar (SAR) ship detection. However, multi-scale variation targets, complex background near-shore scenes, dense arrangement some ships make difficult to improve accuracy. To solve above problem, this paper, a spatial cross-scale attention network (SCSA-Net) for SAR image is proposed, which includes novel (SCSA) module eliminating interference land background. The SCSA uses features at each scale output from backbone calculate where needs space enhances feature pyramid (FPN) eliminate noise, backgrounds. In addition, paper analyzes reasons “score shift” problem caused by average precision loss (AP loss) proposes global (GAP problem. GAP enables distinguish positive samples negative faster than focal AP loss, achieve higher Finally, we validate illustrate effectiveness proposed method performing on Ship Detection Dataset (SSDD), SAR-ship-dataset, High-Resolution Images (HRSID). experimental results show that can significantly reduce noise results, accuracy, superior existing methods.