DCGAN-based synthetic image generation of denim jeans defects

Background Automated defect detection in denim jeans manufacturing is crucial for maintaining quality control efficiency. However, for automated defect detection of denim jeans, machine learning algorithms suffer due to limited data availability because manufacturing industry remains reluctant to share data due to privacy concerns. Moreover, remote manufacturing units make it more difficult to gather necessary defected images. Furthermore, trained personnel are required to capture standard images for training effective models. Traditional image augmentation approaches extend the datasets from seed images; however, there is a lack in image diversification and they do not expand data distribution and thus may lead to overfitting. Deep learning models, especially generative adversarial networks, have the potential to provide effective solutions for industrial problems, such as synthetic image generation for denim jeans defect detection. Methods This article proposes the use of a deep convolutional generative adversarial network (DCGAN) for generating diversified and realistic synthetic images of common denim jeans defects, including broken loops, broken stitches, skipped stitches and twisted legs. The DCGAN model was trained on an initial dataset of 3,930 defective images and subsequently augmented using techniques such as flipping, random zooming, and color space augmentation. Results The generated synthetic images were subjectively validated by domain experts, achieving an average accuracy of 81.5% Objective evaluation using the Fréchet inception distance metric also demonstrated the effectiveness of the proposed approach, with scores of 12.26, 6.75, 7.68 and 27.59 for broken loop, broken stitch, skipped stitch and twisted leg defects, respectively. This work not only contributes to addressing the challenge of data scarcity in defect detection but also paves the way for more accurate automated defect detection systems in denim jeans manufacturing.