PDFormer: Efficient Vision Transformer for Photovoltaic Defect Detection

In industrial production, the quality of photovoltaic determines the power generation efficiency and service life. Therefore, only by improving the quality inspection automation capability of photovoltaic products can we ensure the quality of mass production. Recently, Vision Transformers (ViTs) have shown excellent performance in various visual tasks. However, the ViTs generally suffer severe performance degradation when small-scale datasets are used for training since ViTs overfit quickly. To alleviate this, we propose PDFormer, a simple but effective Transformer framework towards efficient learning for photovoltaic defects detection. Our proposed PDFormer boosts the performance of the ViTs by compound improvements, which generally consists in three levels: data level, structure level, and supervision level. Specifically, an image mixing augmentation method called QuadMix augmentation is first proposed to randomly mix the positive and negative samples for the binary classification task. Besides, we develop a novel attention-based module to reweight the deep features by intermediate classification scores. Finally, we adopt both the ViT and CNN networks as the compound teacher networks to perform compositional multi-teacher knowledge distillation for the transformer student. Benefitted from the overall efficient designs, PDFormer significantly improves the detection performance of the transformer baseline on the dataset. Experimental results demonstrate that PDFormer achieves a top-1 accuracy of 98.08%, surpassing other competitive methods on the photovoltaic dataset.