A Thermal-Aware Infrared–Visible Image Fusion Method for Defect Detection in Power Transmission Equipment

To address the difficulty of single-modal images in simultaneously representing structural defects and thermal abnormal defects of equipment in complex transmission-line inspection scenarios, this paper proposes a transmission equipment defect detection method based on dual-stream thermal-aware infrared–visible image fusion. The proposed method takes spatially registered visible images and infrared thermal images as inputs. First, a dual-stream feature extraction structure is constructed to separately extract structural texture features from visible images and thermal radiation features from infrared images. Subsequently, a shared–specific feature transfer mechanism is designed to decompose and interactively model visible-specific structural information, infrared-specific thermal information, and cross-modal shared semantic information, thereby enhancing the complementary information representation capability between the two modalities. On this basis, a thermal-aware gated enhancement module is introduced to adaptively strengthen abnormal thermal response regions in infrared images, enabling the fused image to preserve equipment edge contours and texture details while highlighting potential thermal fault features. To further constrain the fusion results, a joint optimization function composed of thermal radiation priority loss, edge preservation loss, and contrast enhancement loss is constructed to improve the collaborative preservation capability of structural information and thermal abnormal information in the fused image. Experiments are conducted on an aligned infrared–visible image dataset collected from transmission-line inspection scenarios. The dataset includes normal insulator strings, normal strain clamps, damaged insulator strings, thermal faults in insulator strings, and thermal faults in strain clamps. The experimental results show that, compared with typical fusion methods such as EgeFusion, DANT-GAN, and ITFuse, the proposed method achieves superior performance in terms of average structural similarity, average mutual information, thermal radiation preservation rate, and edge preservation rate. Furthermore, images generated by different fusion methods are fed into a YOLO detection model for downstream defect detection validation. The proposed method achieves Precision, Recall, and mAP@0.5 values of 0.936, 0.921, and 0.943, respectively, outperforming the comparison methods. The results demonstrate that the proposed method can effectively fuse visible structural texture information and infrared thermal radiation information, enhance the discriminative representation of defect regions in transmission equipment, and provide an effective approach for intelligent perception and defect detection of transmission equipment in complex inspection scenarios.