In industrial manufacturing, the detection of stitching defects in fabric has become a pivotal stage in ensuring product quality. Deep learning-based fabric defect detection models have demonstrated remarkable accuracy, but they often require a vast amount of training data. Unfortunately, practical production lines typically lack a sufficient quantity of apparel stitching defect images due to limited research-industry collaboration and privacy concerns. To address this challenge, this study introduces an innovative approach based on DCGAN (Deep Convolutional Generative Adversarial Network), enabling the automatic generation of stitching defects in fabric. The evaluation encompasses both quantitative and qualitative assessments, supported by extensive comparative experiments. For validation of results, ten industrial experts marked 80% accuracy of the generated images. Moreover, Fréchet Inception Distance also inferred promising results. The outcomes, marked by high accuracy rate, underscore the effectiveness of proposed defect generation model. It demonstrates the ability to produce realistic stitching defective data, bridging the gap caused by data scarcity in practical industrial settings.
Keywords: Deep convolutional generative adversarial network; Deep learning; Defect generation; Image synthesis.
© 2024 The Authors. Published by Elsevier Ltd.