Abstract:In order to meet the requirements of recognition speed and accuracy for complex photovoltaic (PV) cell defect detection tasks, this paper proposed an improvement strategy based on the Yolov5 network model. First, the model's ability to localize defects was enhanced by adding shallow feature inputs to strengthen the model's fusion of shallow feature information; second, the Squeeze-and-excitation (SE) attention mechanism was introduced to expand the receptive field and strengthen the dependency between feature channels; finally, a smoothing downsampling module was constructed to replace the convolutional sampling, which reduced the loss of information while retaining the important features. The experimental results show that the Yolov5 model based on this improved strategy achieved an average accuracy of 92.06% in the multiclassification defect detection task, and the average recognition speed reached 25.31 frames per second (FPS), which took into account the requirements of performance and real-time performance, and can satisfy the needs of PV cell defect detection in industrial scenarios.