A deep learning-based approach for semiconductor manufacturing images defect identification

Image-based defect detection techniques are widely used in semiconductor manufacturing, with deep learning approaches marking a significant breakthrough for this task. This thesis, resulting from an internship at Statwolf Data Science, aims to develop a deep learning technique for semiconductor manufacturing visual anomaly detection, specifically focusing on recognizing a particular defect: corrosion on aluminum chip images. During the research, it was discovered that the dataset exhibited a bias, with images containing corrosion defects consistently displaying specific aluminum layer patterns, while non-corrosion images had entirely different patterns. This bias, along with other issues such as a limited number of samples, class imbalance, and image complexity, rendered traditional supervised classification techniques impractical. This dissertation details the use of explainability techniques to identify the dataset bias and compares traditional classification methods with a segmentation approach. The segmentation approach effectively reduced the influence of the bias and significantly improved overall classification performance, demonstrating its suitability for this application.