Development and Implementation of a YOLOv5-based Adhesive Application Defect Detection Algorithm

초록

This study investigated the use of YOLOv5 for defect detection in transparent adhesives, comparing two distinct training methods: one without preprocessing and another incorporating edge operator preprocessing. In the first approach, the original color images were labeled in various ways and trained without transformation. This method failed to distinguish between the original images with properly applied adhesive and those exhibiting adhesive application defects. An analysis of the factors con-tributing to the reduced learning performance was conducted using histogram comparison and template matching, with performance validated by maximum similarity measurements, quantified by the Intersection over Union values. Conversely, the preprocessing method involved transforming the original images using edge operators before training. The experiments confirmed that the Canny Edge Detection operator was particularly effective for detecting adhesive application defects and proved most suitable for real-time defect detection. © 2024, Korean Sensors Society. All rights reserved.

키워드