Deep Learning-Based Surface Defect Detection on Leather Sheets for Automated Visual Inspection

Abstract

The leather sheet quality inspection process is a crucial step in the leather product manufacturing industry to ensure that the raw materials used meet established quality standards. The manual leather quality inspection process has various limitations, such as being relatively time-consuming, high operational costs, and prone to subjectivity and human error. These conditions encourage the development of an automated visual inspection system based on artificial intelligence (AI) capable of classifying defects on the leather surface more quickly, consistently, and accurately. This study proposes a Deep Learning-based approach to classify defects in leather images. The dataset used is a primary dataset collected independently through image capture using a camera at several Micro, Small, and Medium Enterprises (MSMEs) producing leather products in Manding, Bantul, Yogyakarta. The dataset consists of 142 images, including 89 images in the defect class and 53 images in the non-defect class. The proposed method includes several main stages: data preprocessing, splitting dataset, model training, and model performance evaluation. This study produced accuracy of CNN, SqueezeNet, and ResNet18 models of 80.95, 90.48, and 90.48 respectively. In addition, this study obtains f1-score of 81.48, 90.71, and 90.48.