Image classification of fabric defects using ResNet50 deep transfer learning in FastAI

Abstract

One of the most common issues in manufacturing is the inability to persistently maintain good quality, which can lead to product defects and customer complaints. In this research, the novel implementation of deep learning for fabric defect classification in FastAI was proposed. The residual network structure of ResNet50 was trained through transfer learning to classify the data set that contained five classes of fabric images: good, burned, frayed, ripped, and stained. A novel approach to constructing the data set was undertaken by compiling randomly downloaded fabric images within the aforementioned five classes with a broad variety from the internet. The effect of the two splitting methods in dividing the data into training and validation data was investigated. Random splitting divides the data into random class proportions, while stratified splitting maintains the original class proportions. Models were tested offline with unseen data and reached a mean accuracy of 92.5% for the 2-class model and 70.3% for the 5-class model. Based on the attained accuracy and precision, no splitting method was superior to the other. The feasibility of the system’s online implementation was evaluated by integrating a smartphone camera to capture and classify fabric samples, with a mean accuracy of 75.6% for the 5-class model.