Scratch and Defect Detection System for Painting Images: A Deep Learning Approach

This paper presents a deep learning-based methodology for detecting scratches and defects in paint polishing images. Paint polishing is a critical stage in steel manufacturing, and ensuring the quality of the polished surface is essential to meet industry standards and customer satisfaction. Traditional methods for defect identification usually depend on labor-intensive, human error-prone manual examination. The objective of this system is to prevent welding dents from appearing on the sheet metal surface during spot welding finishing treatment. We present a novel deep learning mechanism capable of automatically identifying and localizing scratches and defects in paint polishing images. We leverage convolutional neural networks for their potential to understand difficult features directly from the image data. Extensive experiments demonstrate the effectiveness of our approach in accurately detecting and classifying various types of defects, improving the efficiency and reliability of the quality control process in steel manufacturing. Our examination with various YOLOv8 model weights demonstrated that YOLOv8m produces relatively good results with a smaller dataset and less processing time. Meanwhile, using the same dataset, YOLOv8x obtains a higher mAP50 value of around 59%, but with the drawback of a longer inference time.