Dynamic Domain Adaptation Object Detection for Photovoltaic Cell Defects

The task of defect detection of photovoltaic (PV) cell electroluminescence (EL) is an important part of its manufacturing process. There are differences in background, defect contrast and resolution (Domain Shift) during the quality inspection of photovoltaic cell due to intrinsic factors in machine. Classical object detection methods fail to perform as expected on domain-shifted test sets. Traditional domain adaptation methods also need access to all unlabeled test set, which cannot be implemented on the real-time inspection of the photovoltaic production line. According to the actual production line requirements, we proposes a dynamic unsupervised domain adaptation (DUDA) object detection method to improve the generalization of the detection model on the target domain. We believe that the convolutional layer and the fully connected layer in the deep learning model store knowledge related to image categories and textures, and the BN layer has domain-related knowledge, so update statistical information for each batch of data that continues to come from the production line. The DUDA method can dynamically adjust the model parameters to adapt to the changing data in the dynamic environment of the production line.