Abstract: Unmanned aerial vehicle-borne thermal infrared equipment is used to carry out aerial photography and defect detection of solar photovoltaic modules. Aiming at the limitation of UAV’s memory and computing power,as well as the problems of the existing defect detection model based on deep learning in the complex environment of large photovoltaic power plants,such as large model parameters and large computational overhead,YOLOv5 LiteX is designed as an ultra-lightweight defect detection model. Firstly,the weighted bidirectional feature pyramid BiFPN is selected to replace the original feature pyramid PANet,so that the features can be effectively connected and fused across scales. Secondly,a larger detection scale is added on the basis of feature fusion to improve the performance of the model in detecting smaller defect targets. Focal-EIoU Loss is introduced to improve the prediction loss of the original frame coordinates so that the network could focus on the computation of difficult samples. In addition,the data enhancement method is used to overcome the problem of too little data. The mean average precision（mAP）of the improved network is 7.32 percentage points higher than that of the baseline network（Lite-YOLOv5）,and the mAP of difficult samples（foreign body occlusion）is greatly improved.