Exploring the Applications of Textural Features for Automatic Leather Characterization

Abstract

This paper examines the significance of tex- tural features in digital images of leather samples for automated visual quality inspection in industrial produc- tion. Leather defect detection involves tackling two im- portant challenges: first, accurately isolating the leather surface from the background in the acquired images, and second, conducting a detailed analysis of the extracted region to identify and classify potential defects. This study investigates the potential of textural descriptors for leather characterization, exploring their application as feature vectors in both supervised and unsupervised machine learning methods. We evaluate these meth- ods on two tasks: distinguishing between the leather surface and background in acquired images, and classi- fying leather defects. As anticipated, supervised methods demonstrate superior performance, achieving over 98% accuracy in leather-background separation and up to 90% in defect classification. In contrast, the unsupervised approach yields more modest results, with Rand Index and Fowlkes-Mallows Index values of 81% and 73%, respectively. Despite the limitations of textural descrip- tors in leather defect classification, the results highlight the potential of texture analysis and unsupervised learn- ing in automating image analysis and enhancing quality control in leather production.