Classification of photovoltaic panel defects based on improved inception-ResNet-V2

To solve the problem that the photovoltaic panel defect classification method has too many parameters and too deep network depth, an algorithm based on the improved Inception-ResNet-V2 is proposed. Firstly, the channels in the network structure are reduced for optimizing the large parameters problem and reducing the amount of computation. Secondly, lager feature maps are retained. Finally, UniRepLKNetBlock is incorporated into the Inception-Resnet module for highlighting the advantages of a larger sensory field in capturing contextual information. Experiments are carried out on the public dataset of electroluminescent photovoltaic panel defect images from the Kaggle website, and the experimental results show that the proposed improved model can effectively alleviate the problems of low classification accuracy and excessive number of parameters. The classification accuracy of the improved model is 97.36%, which is a significant improvement compared to 88.61% of the original model, and the number of parameters of the improved model is only 28.9% of that of the original model. As far as we know, this is the first time that the Inception-ResNet-V2 model has been introduced to solve the classification of photovoltaic panel defects.