摘要
In the domain of binary classification, Convolutional Neural Networks (CNNs) are a well-established technique, evidenced by numerous studies highlighting their superior accuracy and generalizability compared to other methods. We try to use CNNs on a semiconductor manufacturing company's pad defect dataset to evaluate their effectiveness in classification. However, due to the significant influence of noise, the results are unsatisfactory. Thus, we propose a preprocessing technique, which includes U-Net pad feature extraction and dilation operation in this work. This preprocessing technique allows the model to identify the features of pad defects more accurately, minimizing the impact of background interference. Consequently, the proposed approach is more adaptable across various production lines in that company.