YOLO-IMS:An improved method for infrared mutilated ship target detection

The lack of visual information of infrared mutilated ship target usually leads to greater difficulty in identification. In order to alleviate the difficult problem of infrared mutilated ship target identification, the YOLOv5s-based improvement network YOLO-IMS is proposed. Firstly, the efficient spatial multi-scale attention (ESMA) module is proposed to improve the ability of the model to express the features of the mutilated ship. Based on feature grouping, the input tensors of two parallel branches are extracted and weighted with different features, and then the output features of the two parallel branches are aggregated through cross-spatial learning. And a branch is introduced to make the module have a long-range dependence relationship and enhance the spatial feature extraction capability of infrared mutilated ship targets. Secondly, we replace the original YOLOv5s' CIoU Loss with the normalized gaussian wasserstein distance (NWD) loss. This modification accelerates network convergence during training. Finally, the spatial pyramid pooling fast plus (SPPFP) module is introduced to enhance local feature fusion for incomplete targets.Experimental results demonstrate that YOLO-IMS not only maintains the detection efficiency of the original YOLOv5s but also achieves a 77.0% mAP@0.5 on our established infrared mutilated ship dataset, outperforming the original YOLOv5s by 10.5%. YOLO-IMS also performs well on the standard IRay infrared human-vehicle dataset.