A novel defect generation model based on two-stage GAN

Abstract

The fabric defect models based on deep learning often demand numerous training samples to achieve high accuracy. However, obtaining a complete dataset containing all possible fabric textures and defects is a big challenge due to the sophisticated and various fabric textures and defect forms. This study created a two-stage deep pix2pixGAN network called Dual Deep pix2pixGAN Network (DPGAN) to address the above problem. The defect generation model was trained based on the DPGAN network to automatically “transfer” defects from defected fabric images to clean, defect-free fabric images, thus strengthening the training data. To evaluate the effectiveness of the defect generation model, extensive comparative experiments were conducted to assess the performance of the fabric defect detection before and after data enhancement. The results indicate that the detection accuracy was improved regarding the belt_yarn, hole, and stain defect.

Image synthesis of apparel stitching defects using deep convolutional generative adversarial networks

“…Traditional methods for feature extraction often involve manually designed extractors, required expertise and complex adjustments. These methods were customized to specific applications, resulting in inadequate generalization and robustness [ 6 ].…”

“…These techniques can provide some variation in the data, they only produce a limited amount of alternative data. However it's crucial to acknowledge that these techniques mainly focus on altering the existing data rather than enhancing the image information, and their generalization effect may be limited due to certain constraints [ 6 ].…”

Fabric defect image generation method based on the dual-stage W-net generative adversarial network

“…Enhancing defect datasets is a key method for improving the accuracy of defect detection. Utilizing generative adversarial networks (GANs) for data augmentation is a useful and efficient technique, and this approach has been applied in various domains, including architecture, 24 medicine, 25,26 facial recognition, 27 fabrics, 28,29 and more. Traditional GANs used in image generation face challenges such as model collapse, gradient vanishing, and gradient explosion.…”

“…Due to the diversity of textile textures, complex backgrounds, and a wide range of defect types and sizes, it is difficult to use simple GANs to generate fabric defect samples that meet the required standards. Zhang et al 28 introduced a two-stage GAN-based defect generation model, building on the Pix2Pix model. After two stages of training, fabric defects can be generated with relatively simple backgrounds.…”