Automated Detection of Wet Blue Leather Defects Using Artificial Vision with YOLOv8s in a Tannery

León-León R.A.; Chiclayo-Fulgencio P.E.; Chiclayo-Fulgencio K.S.; Mendoza-Guerra R.M.; Arias-Rumay P.E.; Rodríguez-Velásquez J.L. · CISCI - Conferencia Iberoamericana en Sistemas, Cibernetica e Informatica (2025) · DOI: 10.54808/cisci2025.01.34

Resumen

The manual inspection of Wet Blue leather in the tanning industry is prone to inconsistencies, resulting in significant economic losses. This study addresses this issue by developing and validating a computer vision system for automated defect detection and classification. A YOLOv8s model was trained using a dataset of 4,158 images captured under real operational conditions in a local tannery, classifying defects into four categories: good, veins, fat nodules, and holes. The system incorporates post-processing logic to assign a final quality grade (First, Second, Third) according to an industrial rubric. Results demonstrate high effectiveness, with an overall mean Average Precision (mAP@0.5) of 98.2% for defect detection on a validation set, and 96.3% classification accuracy on a test set for quality grading. The model operates at 30 frames per second (FPS) on consumer-grade hardware, confirming its viability for real-time deployment. We conclude that the YOLOv8s-based system offers a reliable and efficient solution to standardize quality control, reduce human error, and optimize product value in the leather industry.

Instituciones: Universidad Privada del Norte