PDTNet: An Image-Based Model for PV Panel Defect Detection

PV panel modules are the core component of solar power technology and it is vital that defects are quickly detected and repaired to ensure their safe and stable operation. Traditional computer vision detection model have problems such as low detection efficiency, many missed detections and poor robustness. To address these problems, we propose a single-stage target detection model PDTNet, which can better extract defect features and can be better deployed with a small number of parameters. Firstly, we propose a multi-scale feature extraction module and then increase the network receptive field by increasing the convolutional kernel size. Secondly, an ECA attention mechanism is added between the backbone network and the FPN layer to increase the attention of the model for the channels. Finally, by switching the network to DW convolution, the model has a smaller number of parameters, and the inference speed of the detection device is accelerated. The experimental results show that, compared to the YOLOX-Tiny model, our model achieves a 2.46% performance improvement on the PV-Multi-Defect dataset, with a detection speed of 56.7 frames per second, enabling real-time object detection with fewer parameters. In addition, the model has better robustness and shows better results on the Pascal VOC dataset.