Textile Defect Classification Based on Convolutional Neural Network and SVM

Abstract

Textile defect detection and classification is an important part of the textile production process, however, to detect accurately and efficiently is still difficult. In this study, we present an effective formulation for textile defect detection. Unlike traditional textile detecting methods, a conventional neural network (CNN) support vector machine (SVM) is designed to extract the depth features of textile images and to classify defects. The effectiveness of textile feature extraction is improved by optimizing Alexnet to obtain a CNN with better feature extraction and less computation. Textile defect classification is more effective using SVM instead of Softmax. Experimental results demonstrated that the improved Alexnet-SVM gave 99% accuracy in defect classification using the TILDA database. This accuracy was 5% greater than Alexnet and GoogleNet.