Abstract:During the task of targets identifying infrared ship images, due to the presence of targets of various scales in the infrared images and noise interference that leads to loss of feature information, the resulting algorithm in high missed and false detection, propose a YOLOv8-based ship target recognition algorithm named CL_YOLOv8.Firstly, for the input ship data, a large amount of information is lost during feature extraction and spatial transformation layer by layer, a programmable gradient information structure CbPGI (Combined programmable Gradient Information) is designed to enable the model to obtain reliable gradient information to update network weights and improve detection accuracy. Secondly, to address the problem of increased model computational complexity and parameters caused by the CbPGI module, we introduced the high-speed inference convolution PartialConv. We constructed the LWPC (Lightweight PartialConv) structure. This structure applies conventional convolution on only a portion of the input channels for spatial feature extraction and fusion, reducing the model storage requirements. The algorithm computational complexity of constructing the LWPC structure decreased by 31.3%, and the model parameter decreased by 22.8%. Under the same experimental conditions, the CL_YOLOv8 algorithm achieved a detection accuracy of 91.0% in mAP50 on the self-built ship dataset, which is 3.05% higher than the baseline model. |