YOLOBiEMA: SAR ship detection model based on spatial learning and feature fusion

The effectiveness of synthetic aperture radar (SAR) ship detection has been improved by deep learning. In spite of this, the task of locating vessels in intricate environments continues to be difficult, particularly because of the diverse dimensions and orientations. This paper presents YOLOBiEMA, an upgraded YOLOv8 model that incorporates a novel bidirectional attention mechanism for better spatial context. The model also incorporates dynamic upsampling and specialized convolutional layers to enhance resolution and feature representation. An updated loss function improves bounding box regression, ensuring better performance across different ship sizes. Tests on the HRSID and SSDD datasets show that YOLOBiEMA achieves 70.6% <i>mAP_50:95</i> and 79.4% <i>AP₇₅</i> on HRSID, and 73.2% <i>mAP_50:95</i> and 85.1% <i>AP₇₅</i> on SSDD. These results surpass the performance of current leading models in detection and localization making YOLOBiEMA a promising solution for real-time SAR-based maritime surveillance, disaster response, and vessel tracking.