Leather Defect Classification and Segmentation using Self-supervised Learning

Animal skins are tanned in order to produce leather, which is a natural material and the price of leather is expensive since it is very susceptible to its quality and condition of surface flaws. Controlling the quality of the final goods in a leather production business requires the manual examination of defects; however, it is time-consuming and frequently prone to human mistakes, and there are very few papers in the literature that investigate leather flaw identification using computer vision approaches. Access to annotated datasets is now one of the most significant obstacles to the application of deep learning techniques in industry, where Self-supervised learning (SSL) techniques attempt to deliver effective, deep feature learning without requiring a massive amount of labeled data sets, which result in easing the labeled data sets obstacle to the actual implementation of deep learning. In recent years, these approaches have improved fast, with their effectiveness reaching and occasionally better than fully supervised pretraining options across a number of data categories where we will focus on images. This article focuses on the emerging field of defect detection in the industrial sector on a real world dataset, utilizing the Patchification approach and YOLOv5, which were both pretrained on Self-supervised learning output models. This experiment’s initial challenge was execution time, which was drastically lowered by applying efficient strategies in the classification process. We utilized YOLOv5, which has flexible bounding boxes, to solve the second difficulty, which was a defect that was larger than the patch size we had specified. The positive findings indicate that using these sorts of solutions can lead to an effective telchnological tool for assisting the community in addressing the issue of low quantity of labeled data to have more accurate and faster results. Furthermore, we have adopted the same approaches to PCB and Fabric datasets for the task of defect detection.