Transfer Learning-Based Defect Detection System on Wafer Surfaces

This study addresses the significant decline in the accuracy of wafer inspection models when the imaging system changes post-training. We propose a domain adaptation method based on semantic segmentation models that maintains accuracy without the need for re-labeling data despite changes in the imaging system. This method was tested on wafers from an actual production line under two different detection environments: a custom-built simple optical system (source domain) and a precision measurement platform (target domain). To align the source and target domain datasets, we introduced an image preprocessing method that adjusts the contrast and brightness of the source domain data based on the histogram distribution of the target domain. We utilized adversarial learning to transfer features from the source to the target domain and modified the segmentation network architecture to prevent overfitting to the source domain data. Additionally, we extended the domain adaptation framework to handle multiple target domains using a multi-discriminator network strategy, enhancing the model’s adaptability to diverse production line environments. Our results demonstrate that compared to the original network, our model significantly improves accuracy with increases of 4.3%, 16.3%, and 11.4% for three different depths of the semantic segmentation model. Furthermore, our proposed network outperforms widely used style transfer methods with performance improvements of 13.2% and 17.3%. Post-processing the output segmentation maps yielded accuracy, precision, and recall scores of 96.8%, 93.5%, and 100%, respectively.