Akademik Veri Yönetim Sistemi
TEKSTIL VE KONFEKSIYON, cilt.32, sa.3, ss.208-219, 2022 (SCI-Expanded)
Fabric defect detection is generally performed based on human visual inspection. This method is not effective and it has various difficulties such as eye delusion and labor cost. To deal with these problems, machine learning and computer vision-based intelligent systems have been developed. In this paper, a novel real-time fabric defect detection system is proposed. The proposed industrial vision system has been operated in real-time on a loom. Firstly, two fabric databases are constructed using real fabric images and new defective patch capture (DPC) algorithm. One of the main objectives in this study is to develop a CNN architecture that focuses only on fabric defect detection. One of the most unique aspects of the study is to detect defective pixel regions of fabric images with Fourier analysis on a patch-based and integrate it with deep learning Thanks to the novel developed fast Fourier transform-based DPC algorithm, defective texture areas become visible and defect-free areas are suppressed, even on complex denim fabric textures. Secondly, an appropriate convolution neural networks (CNN) model is developed. Thus the new dataset dataset is refined using negative mining method and CNN model. However, traditional feature extraction and classification approaches are also used to compare classification performances of deep models and traditional models. Experimental results show that our proposed CNN model integrated with negative mining can classify the defected images with high accuracy. Also, the proposed CNN model has been tested in real-time on a loom, and it achieves 96.5% detection accuracy. The proposed model obtains better accuracy and speed performance in terms of detection accuracy with a much smaller model size.