Fabric Defect Image Generation Based on Improved CycleGAN

The accurate detection of fabric defects remains a significant challenge, with the construction of the dataset being a major obstacle. To address the issue of insufficient data in fabric defect datasets, we have introduced a novel defect image generation algorithm called SA-CycleGAN. Initially, a residual module, SABlock, is incorporated with a self-attention mechanism to enhance the generator's ability to capture details. Furthermore, the generator is redesigned with an AutoEncoder and SkipConnection structure to improve the quality and diversity of image generation. Concurrently, we employ LPIPS as a cyclic consistency loss function to optimize the visual quality and realism of the generated images. Experimental results demonstrate that the SA-CycleGAN algorithm has brought about a significant data augmentation effect for fabric defect images. The effective application of the data generation method not only enriches the training dataset for fabric defect detection models but also contributes to enhancing the performance and generalization capability of the detection models.