This study aims to solve the problem of reduced detection accuracy caused by a complex target-position scene and an uneven size in the detection of the abnormal heating point in an infrared image of a high-voltage switchgear. According to the YOLO v3 algorithm, the basic network architecture was optimized by including a convolution module and adjusting some hyper-parameters to realize rapid detection and identification of abnormal heating points in high-voltage switchgears. Simultaneously, a dataset for abnormal heating points of infrared images in high-voltage switchgears was established, and appropriate weights were obtained through training. The experimental results indicated that the detection method had a fast recognition speed, high accuracy, and strong generalization ability. The test accuracy reached 91.83%, indicating that the method can be initially applied to the detection of abnormal heating-point targets in high-voltage switchgears.