25 June 2024 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Aiming to the problems of oriented bounding box based (OBB-based) SAR ship detection, which include the large model volume, the imbalance of positive and negative samples, and imprecise ship detection and positioning due to dense arrangement and directional diversity, we propose a lightweight OBB-based SAR ship detection model, GMDR-Net, which is based on gaussian mixture data augmentation and distance rotation IoU loss. Firstly, based on the Oriented RepPoints and ConvNeXt model, we propose a lightweight baseline model. Secondly, we propose a novel Gaussian Mixture Data Augmentation (GMDA) method, which effectively enhances the richness of samples and alleviates the problem of imbalanced positive and negative samples by combining the Gaussian Mixture Model (GMM) and SAR image dataset labels. Finally, we propose a distance rotation IoU (DRIoU) loss function, which introduces a distance penalty term based on GIoU, to achieve more accurate regression of the bounding box. Experimental results show that GMDR-Net achieved an Average Precision (AP) of 77.14% and 87.56% for the HRSID and SSDD datasets. Additionally, at 800×800 and 640×640 input resolutions, its floating point operations (FLOPs) are only 11.78 G and 7.54 G, and frame per second (FPS) reaches 29.62 and 44.76, striking a balance between accuracy and efficiency, and exhibiting wonderful insights in remote sensing field.