Abstract: Automatic recognition of infrared thermal image is an important means of the defect and fault diagnosis of substation equipment. Aiming at the problems of substation equipment detection, such as being extremely vulnerable to background clutter interference, image visual effects and lack of intelligent methods, fast guided filtering was used to retain edge information when removing noise. A parameter self-adjusting Retinex algorithm was proposed to enhance image contrast and an improved YOLOv3 network was presented to increase the recognition accuracy of substation equipment. After testing, the mAP of five kinds of substation equipment can reach 94.85%, and the recognition speed of each picture is 7.88ms. The experimental results show the accuracy and rapidity of the proposed method, which provides conditions for realizing the status monitoring of substation equipment.