A Low Complexity Gabor Based CNN Model for Defect Detection in Fabrics

In the textile industry, automatic fabric defect detection is the most important process to ensure the quality of the fabric. The current technology uses a learning-based methodology to identify defects in fabrics with simple patterns. In the current system, Convolutional Neural Network (CNN) models are used to identify the defects present in the fabric image with maximum accuracy. However, the complexity of the CNN becomes more and it takes much time to process the data. To cater to this problem, a novel Gabor based CNN (GCNN) architectures namely Gabor based VGG-16 (GVGG-16) and Gabor based MobileNet (GMN) models are developed for detecting defects in simple and complex patterned fabrics. In this case, the initial convolution layer uses the Gabor filter bank rather than traditional filters. Gabor filters are more beneficial for examining textures with varying sizes and orientations. The analysis of the texture and the information extracted from it are influenced differently by each of the Gabor filter's parameters. Thus, this paper examines how Gabor filter parameters are applied in GCNN architecture. The TILDA textile image database shows that the suggested model for fabric defect detection can detect defects with an accuracy of 99.17% using the fewest trainable parameters possible.