PV Cell Defect Detection Based on Improved YOLOv7 Modeling

Abstract

This paper proposed an improved model based on YOLOv7, with the aim of tackling the issues of low detection accuracy and heavy model weight of current defect detection algorithms for photovoltaic cells. Ghost convolution is introduced in the original backbone network and neck to reduce the weight of the model. Secondly, the DenseBlock module is improved in two parts to enhance feature extraction ability and feature fusion ability. The introduction of a Wise-IOU loss function to further enhance the detection accuracy of the model was finally successful. Experiments with four defect datassets featuring cracks, broken gates, thick lines and black cores revealed that the weight size of the improved model in this paper was reduced by 24.4%, and the defect detection accuracy and mean average detection accuracy were both increased by 9.7% and 3.7% respectively, in comparison to the original YOLOv7 network.