Abstract: A defect detection model named OD-YOLO is proposed and designed to improve the problem of complex background interfering with the defect detection effect in electroluminescence imaging of polycrystalline silicon solar cells. We use the secondary convolution module（TwiceConv-OD） to filter out the complex grain background interference and enhance the model’s focus on the defects themselves; we propose the anchor-plus1 allocation strategy to increase the number of defect positive samples obtained by the model in the face of the complex background, which improves the model’s accuracy and recall and reduces the omission of misdetections; and we use the K-means++ algorithm to initialize the anchor frames and cluster the anchor frames to be more representative of all the detected samples. The size of the anchor frame is initialized using the K-means++ algorithm, and the clustered anchor frame is more representative of the geometrical shape of all defects in the detected samples, which is better adapted to the differences in defect scales of polysilicon solar cells. Experimentally verified by the publicly available defect detection dataset PVEL-AD-2021: the mean average precision（mAP） of the OD-YOLO model reaches 89.4%, which is improved by 3% compared to the YOLOV5s defect detection model, the accuracy is improved by 4.8%, the recall rate is improved by 1.9%, the parameters are reduced by 4.5%, and the speed is 104 frames per second.