Fabric Defects Detection for Multicolor Yarn Shoe Upper Using Morphological Operations

Abstract

This study proposes a method for detecting defects in shoe upper fabrics with multicolored yarns, where the pattern is similar to the defects, which leads to false positives. Image preprocessing was used to simplify the complex background pattern, highlighting the defect. A dataset of shoe-upper defects was created using images captured using a vision system. Faster region convolutional neural network-which can detect defects with high accuracy under complex backgrounds-was used to detect shoe-upper defects. Precision increased by 4.3-95.3% after preprocessing; thus, preprocessing reduced the false detection of the background as defects. The detection precision of the defect-detection models was compared according to fabric type. YOLOv3 had higher detection precision for linen fabrics with simple background patterns and regular patterns, whereas faster region convolutional neural network exhibited higher detection precision for fabrics with complex background patterns, such as multicolor yarn fabrics.