Using deep learning to detect the presence/absence of defects on leather: On the way to build an industry-driven approach

Abstract

In textile/leather manufacturing environments, as in many other industrial contexts, quality inspection is an essential activity that is commonly performed by human operators. Error, fatigue, ergonomic issues, and related costs associated to this fashion of carrying out fabric validation are aspects concerning companies' strategists, whose mission includes to watch over the physical integrity of their employees, while aiming at enhanced quality control methods implementation towards profit maximization. Considering these challenges from a technical/scientific perspective, machine/deep learning approaches have been showing great skills in adapting a wide range of contexts and, in particular, industrial environments, complementing traditional computer vision methods with characteristics such as increased accuracy while dealing with image classification and segmentation problems, capacity for continuous learning from experts input and feedback, flexibility to easily scale training for new contextualization classes – unknown types of occurrences relevant to characterize a given problem –, among other advantages. The goal of crossing deep learning strategies with fabric inspection processes is pursued in this paper. After providing a brief but representative characterization of the targeted industrial context, in which, typically, fabric rolls of rawmaterial mats must be processed at a relatively low latency, an Automatic Optical Inspection (AOI) system architecture designed for such environments is revisited [1], for contextualization purposes. Afterwards, a set of deep learning-oriented training methods/processes is proposed in combination with neural networks built based on Xception architecture, towards the implementation of one of the components that integrate the aforementioned system, from which is expected the identification of presence/absence of defective textile/leather raw material at a low-latency. Several models powered by Xception were trained with different tunning parameters, resorting to datasets variations that were set up from raw images of leather, following different annotation strategies (meticulous and rough). The model that performed better reached 96% of accuracy.