Target Detection in Visible and Infrared Image Matching Based on Improved YOLOv7

By integrating visible light and infrared images, the efficiency of maritime target detection, recognition, and tracking is enhanced. However, some low-cost infrared devices lack focusing capability, preventing the acquisition of individual target images. Consequently, locating ships in corresponding visible light and infrared images becomes particularly critical. When utilizing YOLOv7 for ship target detection, issues of missed detection and low recognition rates for small targets arise. To address these challenges, this paper constructs a dataset comprising maritime target images captured in both visible light and infrared. Furthermore, an improved YOLOv7 model is proposed, involving the replacement of the activation function in the RepConv module with LeakyReLU to enhance network robustness. Additionally, a CBAM module is introduced after the Concat operation in the Neck section to boost network performance and enhance ship target feature representation. Experimental results demonstrate that the modified YOLOv7 model achieves a mAP of 94.18%, representing a 1.91% improvement over the original model. This validates the efficacy of the approach, further enhancing ship target detection performance and reducing missed detection rates.