Shape prior guided defect pattern classification and segmentation in wafer bin maps

In semiconductor manufacturing, patterns formed by defective dies in a wafer bin map (WBM) reveal possible problems during the wafer fabrication process. Therefore, the identification of these patterns is important for root cause diagnosis and yield enhancement. Recently, as the manufacturing process becomes increasingly complicated, mixed-type defect patterns have been frequently observed on the WBMs. The joint classification and segmentation of each pattern contained in the mixed-type pattern is challenging, especially when these patterns are connected or overlapped. This study proposes a shape prior guided method in which the shape templates are deformed to match the patterns with a spatial transformer network. The deformation based method is able to give plausible results while reducing computation cost of the network. Furthermore, the method is flexible and easily extended to deal with complex shapes by defining different shape templates. The experimental results demonstrate the effectiveness of the proposed method in the identification of the pattern types, as well as the separation of each isolated pattern.