Low-Light Recognition-Guided Segmentation of Infrared Thermal Images for Power Equipment

Infrared thermal imaging plays a vital role in power equipment monitoring, but its effectiveness is often limited by low contrast and blurred boundaries. Accurate temperature measurement typically requires low-light visible-light data. To address this issue, we propose a low-light recognition-guided segmentation method for infrared thermal images. Low- light images are first enhanced using URetinex-Net to improve visual quality. YOLOv8 detection is then applied to detect power equipment, and the guided position boxes are spatially mapped from visible-light image to infrared thermal images, to generate region-of-interest (ROI) guidance maps. Finally, superpixel clustering with adaptive thresholding performs fine-grained segmentation. Experimental results show that the proposed method outperforms segmentation approaches without visible-light guidance, attaining IoU, Dice, and BF1 scores of 0.8019, 0.8613, and 0.8221. These results confirm that visible-light guidance significantly enhances segmentation accuracy and robustness, providing reliable technical support for intelligent power equipment monitoring.