In screw inspection processes, quality control is predominantly performed through manual visual inspection, where inspectors examine several thousand screws per day. Consequently, inspection accuracy is highly influenced by individual skill levels and fatigue, leading to variability in results and an increased risk of missed defects. These challenges highlight the need for a robust and efficient automated inspection system that operates independently of human subjectivity. To address this issue, we propose an anomaly detection framework for screw inspection based on machine learning and image processing techniques. The proposed method integrates defect regions extraction from single frames with a decision algorithm that leverages the tracking results of abnormal regions across consecutive frames. Furthermore, to enable efficient implementation of the proposed framework, an inspection apparatus capable of evaluating multiple samples simultaneously is developed. Experimental results obtained from 105 samples demonstrate a recall of 100% and a precision of 46%, indicating effective localization of defect regions and a reduction in false detections through the incorporation of temporal tracking. These results confirm that the proposed approach is effective as a visual inspection method that explicitly accounts for the rotational inspection process encountered in practical manufacturing environments.

This paper proposes a method for augmenting training data targeting automatic defect classification (ADC) of semiconductor wafers using scanning electron microscope (SEM) images. In semiconductor devices, new manufacturing processes and materials are being introduced, increasing defect variations. To achieve high classification performance in ADC, it is necessary to obtain and train a large number of defect images at each manufacturing process. Therefore, we propose a new image-based data augmentation method that takes into account SEM and semiconductor characteristics. This method generates augmented images by the following three techniques: (1) image rotation/flipping considering the orientation of shadows generated in a image that includes defect topographic information, (2) shape deformation that imitates local pattern fluctuations (line edge roughness (LER)), which is one type of structural variation, and (3) generation of defect images of various sizes with physically consistent shadows considering the cast shadow caused by defects. Using a total of 773 actual device images at three processes, we confirmed that the proposed method improved ADC success rates from 51.2∼67.5% to 95.9∼97.9%. This method effectively generates pseudo images that contribute to improving ADC performance, thereby achieving sufficient classification performance required to identify the causes of defects.