Abstract: The tempered glass on the surface of photovoltaic modules will cause reflection noise in the collected thermal infrared images,which is similar to the characteristics of hot spots,which will often leads false detections in hot spot detection task. This paper proposes a lightweight DeepLabv3+ semantic segmentation model called LD-MA（Lightweight DeepLabv3+ with Multi-scale integrated Attention Mechanism）for hot spot detection. LD-MA is based on the DeepLabv3+ network architecture,First,MobileNetV2 is used as the backbone feature extraction network to reduce the amount of network parameters to improve training efficiency. Then,a multi-scale feature fusion module is designed and a CBAM attention mechanism is introduced to retain the multi-stage target features and strengthen the learning of hot spot target feature information and location information. The hot spot detection experiment was carried out on the selfbuilt photovoltaic hot spot data set,and the results showed that the parameters of the LD-MA model were greatly reduced,and at the same time,false detection and missed detection were effectively avoided. In the test set,mIoU and mPA reached 90.82% and 94.39%.