An Improved YOLOv5 Model with FRB Method for Product Surface Defect Detection

In factory production, defect inspection is a crucial section in quality assurance. However, manually is inefficient, expensive, and unreliable. As Industry 4.0 becomes mainstream, the enhancement of production processes through Industrial Artificial Intelligence is an important topic. The deployment of defect inspection systems in automated production lines can reduce employee training, save labor costs, and enhance product quality. This study proposes a new approach for defect detection tasks based on the YOLOv5 object detection method. In this paper, this proposed method combines the YOLOv5 model with a Feature Retaining Block (FRB) to preserve fine features and is referred to as FRB-YOLO. It aims to tackle the issue of declining product quality caused by minor defects during the manufacturing process for most products. Minor defects which may resemble the natural patterns of the product can result in defective products being incorrectly classified as normal. The study conducts experiments using the publicly available NEU-DET datasets. The FRB-YOLO method proposed outperforms YOLOv5 in terms of mean average precision (mAP) across the NEU-DET datasets.