Fabric Defect Detection Based on Lightweight Neural Network

Abstract

Owing to the variety and complexity of defects in the fabric texture image, automatic fabric defect detection is a challenging task in the fields of machine vision. Deep convolutional neural network (CNN) has achieved remarkable progress in the field of target detection, and the application of deep CNN model to fabric defect detection has achieved better results. However, with the detection accuracy increasing of deep CNNs, comes the drawbacks of significant increase in computational costs and storage services, which seriously hinders the usages of CNN on resource-limited environments such as mobile or embedded devices. In this paper, a fabric defect detection method using lightweight CNN is proposed. We introduce an extremely computation-efficient CNN architecture named YOLO-LFD, which adopts continuous 3 \(\times \) 3 and 1 \(\times \) 1 convolution layers for dimensionality reduction and feature fusion. We use multi-scale feature extraction method to improve the detection ability of the model for different size targets, and adopt K-means clustering method on the fabric defect image dataset to find the optimal size of anchor boxes. Experimental results demonstrate that the proposed scheme improves the detection accuracy of fabric defects, greatly reduces the model parameters, and improves the detection speed, which can provide real-time fabric defects detection on embedded devices.