A robust real-time leather defect segmentation using YOLO

Abstract

Natural leather is a product made from animal skin which is treated through chemical procedure to preserve it. It is used in the manufacture of clothing, bags, furniture, automobile material, among others. Because of its capital value in industry, it is important to ensure its quality. Traditional inspection by human experts is expensive, time-consuming, and subjective to human errors. Consequently, automatic leather inspection has become an essential part of any production system as it rejects nonconformities, ensures product quality, reduces operating costs, and shortens production cycle times. This paper presents an artificial intelligent model using computer vision for the autonomous inspection of natural leather. Due to its good results in similar problems, the YOLO algorithm was chosen. More specifically, a comparison of the Small, Medium, Large, and Extra-Large models of YOLOv5 in leather defect detection was performed. We used images of leather with and without defects from the MVTec Anomaly Detection dataset. After training, the models were analyzed and compared based on some performance metrics. All models showed a great ability to detect defects in the dataset used.