Automatic Detection and Classification of Microscopic Defects in Optical Thin-Film Coatings Based on Deep Learning

This study presents an effective method for detecting and classifying microscopic defects in optical thin films, aiming to enhance quality control in thin-film manufacturing. The proposed system utilizes thin-film surface defect images captured by an imaging microscope. It combines image preprocessing techniques, such as translation, scaling, and mirroring, to expand the dataset, thereby generating a rich and representative set of defect images. All images are manually labeled by experts to ensure high-quality annotations and to optimize training efficiency. The YOLOv7 object detection framework is employed for model training and optimization. Model performance is rigorously evaluated using metrics such as the confusion matrix and mean average precision (mAP). The trained model achieved an accuracy of 87.3% on the test dataset, demonstrating both high detection accuracy and practical applicability. This method offers significant potential for automating microscopic defect detection, thus improving the efficiency of film quality inspection and contributing to better production yield in optical thin-film processes.