Wafer Pattern Counting, Detection and Classification Based on Encoder-Decoder CNN Structure

This paper designs an automatic wafer pattern counting pipeline based on a convolutional neural network (CNN) based structure, also called the WPCCNN network pipeline. The study will utilize deep learning algorithms to detect, binary classify and count wafer patterns. In the sample dataset, over two hundred wafers have been scanned by industrial computed tomography, containing 11 different patterns for the dataset images. Each image includes three processed steps. Moreover, it utilizes the lightweight CNN structure to demonstrate detection, classification, and estimated counting [1, 3]. Besides, the study also uses encoder and decoder structure on the CNN algorithm to obtain the closest expected output. Compared to traditional object counting methods, such as localization and density estimation, using this new method to count objects will be more accurate, faster, and more accessible [1-3]. The experiment results indicate that our algorithm is highly accurate with the paring between the original patterns and the labeled markers. The average counting accuracy is 99.6% in a single wafer.