Infrared-Visible Image Feature Interaction-Guided Substation Equipment Target Detection Method

To address the issues of missed detection, false detection, and poor accuracy in single-modal target detection of Substation equipment in complex scenarios, this paper proposes a target detection method guided by infrared-visible image feature interaction. Based on the YOLOv8 network framework, a dual-branch backbone network capable of simultaneously inputting infrared and visible images is first constructed. Secondly, to tackle the problems of blurred target details, long-distance small targets, and occluded target points in infrared images, as well as the insufficient distinction between targets and backgrounds in visible images under low-light conditions, which lead to poor feature capture capability of the network, a Channel-prior Convolutional Attention Module (CPCAM) is introduced into the dual-branch feature extraction network to enhance the feature extraction capability for key targets. Finally, to solve the problem that existing image fusion methods are prone to noise inundation during the fusion process, a Multi-modal Cross-guided Feature Module (MCFM) is utilized the image's light and contrast as predictive weights to guide bidirectional crossmodal adaptive feature fusion through a guided feature sharing mechanism. Experiments show that the image fusion detection method proposed in this paper improves the mAP@0.5 by 19.98% and 35.25% respectively compared with YOLOv8 single-modal network on visible and infrared image detection, and improves the detection accuracy by 11.08% on average compared with mainstream image fusion methods, with no missed detection or false detection.