Abstract: Hot spot effect is one of the main causes of serious damage to photovoltaic panels. To quickly and timely detect hot spots and maintain them for resolution in this paper, a hot spot detection method based on improved YOLOv4-tiny is proposed. First, aiming at the problem that the infrared images of hot spots are scarce and the enhancement of mosaic in the original model is unstable, a method of gamma transform is proposed to effectively expand the hot plate data set. Secondly, in the model, focusing on important features in hot spot infrared images and suppressing unnecessary features, CBAM (convolutional block attention module) is added to the network structure. Finally, in response to the shortcomings of weak receptive field and insufficient information extraction in the original model, the feature pyramid structure in the model is integrated with the idea of PANet (path aggregation network), and a small number of convolutional layers with a convolutional kernel of 1×1 are added to the structure to reduce the number of parameters. The experimental results show that the AP50 of the improved YOLOv4-tiny model accounts for 98.42%, which is 3.63% higher than the original model, and the detection rate reaches 50.06fps on the equipment with the GPU of GTX (graphic processing unit) 1070Ti, with excellent detection accuracy and good real-time performance, which is basically close to practical application requirements.