Computer vision-based automatic fabric defect detection models for the textile and apparel industries

Abstract

Faced with increasingly fierce market competition, textile and apparel enterprises need to control cost and improve the quality of their products in the manufacturing stage. Fabric quality inspection plays an important role in this stage because locating possible fabric defects before cutting is beneficial for reducing raw material waste and ensuring textile product quality. Traditionally, in most textile and apparel companies, fabric quality inspection is performed through a human visual system, which suffers from high labor cost and low effectiveness. Automatic inspection techniques, which can provide intelligent solutions for defect detection without human intervention, are beneficial for improving detection accuracy and efficiency. Over the past several decades, many computer vision-based defect detection models have been proposed to address the fabric inspection problem. Most existing approaches, such as statistical and structural approaches, rely on feature extractions of fabric defects and normal texture to distinguish possible defects. The difficulty in selecting the appropriate features that can be adapted to different types of fabric texture significantly influenced the effectiveness of these approaches. The primary purpose of this research is to develop effective computer vision-based defect detection models that aims at different types of fabric, such as plain fabric, twill fabric, and striped fabric. Woven fabric with uniform texture is one of the most basic types of fabric, which can be classified as plain fabric and twill fabric according to the fabric weaving structure. On the basis of the analysis of the characteristic of defects in plain and twill fabrics, two distinctive detection models were developed in this research. In addition, another hybrid detection model was proposed for striped fabric, which represents the most basic type of patterned fabric with a single element and is widely used in our daily lives, such as shirts and pants. Defects on plain fabric can be regarded as uneven edges on a very smooth surface. Thus, an optimal Gabor filter-based detection model was developed for plain fabric inspection. On the basis of a feature analysis of defects, only two Gabor filters optimized by differential evolutionary algorithm were used to significantly reduce the computational complexity. With defects regarded as random noise in the image, a defect detection model based on non-local sparse representation was proposed for twill fabric inspection. Through the restoration of non-defective fabric texture on the basis of dictionaries learnt from reference images, defects were successfully segmented from the original inspection fabric images. To detect the common fabric defects (errors in weaving structure) and pattern defects (repeated pattern variants), a hybrid defect detection model was developed to solve the inspection problem for fabric with a striped pattern. Gabor filters were initially used to detect variants of the repeated pattern and reduce the disturbance caused by the edge of repeated pattern. Sparse representation was adopted to segment the remnant defects in the resultant feature images in the second step. On the basis of a textile texture database (TILDA) and the amount of real fabric samples acquired from real apparel companies, extensive experiments were conducted to evaluate the performance of the proposed methodologies. Experimental results show that the proposed detection models are effective in detecting various types of fabric defects and are superior to several representative and popular defect detection models. The results of this research demonstrate that computer vision and artificial intelligence can offer satisfactory performance for the automatic defect detection problems of various types of fabric in the textile and apparel industries. They also reveal that considering the characteristics of different types of fabric texture is of paramount significance to the establishment of a practical fabric inspection model, which is highly sensitive, robust, and efficient.