An Efficient YOLOX-Based Method for Photovoltaic Cell Defect Detection

Abstract

As the cumulative installed capacity of photovoltaic power generation continues to grow globally, defect detection plays an increasingly critical role in the healthy operation and maintenance of photovoltaic systems. However, accurate and efficient defect detection is a challenging task for small targets, various defect shapes, and complex background interference. Therefore, this paper proposes a high-efficiency photovoltaic cell defect detection method based on improved YOLOX. First, the transfer learning training strategy is adopted to accelerate model convergence, which can also avoid the problem of insufficient accuracy due to the small number of defect samples. Secondly, to suppress the interference of complex backgrounds, the SENet attention mechanism is added to the feature extraction process. Then, the ASFF strategy, which can adaptively learn the features of each scale, is introduced at the end of the PAFPN network to highlight the importance of defect features. Furthermore, the detection accuracy of the model is improved by improving the loss function of positioning, classification and confidence. Finally, our model is tested on the global public photovoltaic electroluminescence anomaly detection (PVELAD) dataset. The mAP of our model is 96.7%, and the detection speed is 71.47FPS. Experimental results show that our proposed method achieves excellent detection results, especially the recognition performance of small target defects is greatly improved.