Automatic fabric defect detection with a wide-and-compact network

Abstract

Automatic detection of fabric defects is an important process for the textile industry. Besides the detection accuracy, an automatic fabric defect detection solution for a resource-limited system also requires superior performance in terms of processing time and simplicity. This paper proposes a compact convolutional neural network architecture for the detection of a few common fabric defects. The proposed architecture uses several micro architectures with multilayer perceptron to optimize network. The main component of a micro architecture is constructed using techniques of multi-scale analysis, filter factorization, multiple locations pooling, and parameters reduction, to improve detection accuracy in a compact model. Experimental results show that, compared to mainstream convolutional neural network architectures, the proposed network achieved superior performance in terms of detection accuracy with a much smaller model size. It worked well not only for fabric defects detection, but also for object recognition on a few public datasets.