A deep learning-based method was proposed， for defect classification and localization of photovoltaic panels to quickly and accurately determine the location and type of defects. To overcome the perspective limitations of traditional single-image defect detection methods， algorithms were adopted， such as image registration and stitching to generate high-resolution panoramic images of the photovoltaic panels. Deep learning techniques were then used to classify the infrared images of the photovoltaic panels and effectively identify the types of defects by comparing them with visible light images. The accuracy， precision， recall， and F1 score of the photovoltaic panel defect classification could reach 93.71%， 93.13%， 93.20%， and 93.11%， respectively. Compared with traditional methods， this approach had advantages， such as non-contact， high efficiency， and fast speed， making it suitable for detecting and locating defects in large-scale photovoltaic panels. It could provide accurate and comprehensive information about photovoltaic panel defects in a short time.