面向边缘设备的EL图像光伏电池缺陷检测模型OA

To address the issues of high complexity and difficulty in real-time application of existing models,a lightweight and efficient photovoltaic cell defect detector(PVDet)model is proposed for electroluminescence(EL)image.First,for the backbone,we adopted a lightweight and efficient downsampling module and designed a new feature extraction module,with the aim of reducing network parameters while enhancing sensitivity to information at different scales.Secondly,in order to enhance the information interaction at the connection between the network backbone and neck,we propose a new feature pyramid structure.This structure increases the receptive field and further improves classifi-cation performance.Finally,we introduce a lightweight dynamic sampling,to improve the model's upsampling effect.Experimental results on the public EL image dataset PVEL-AD demonstrate that the proposed PVDet model achieves 89.4%average precision with only 2.0×105 parameters(Params)and 5.0×108 floating-point operations(FLOPs).Compared to the latest YOLO11 model,the number of Params and FLOPs is reduced by nearly 92%.It outperforms other mainstream YOLO models in terms of speed,making it suitable for deployment on industrial edge devices.