基于改进CycleGAN的织物疵点数据集增强OA

In order to solve the problems of insufficient number of image samples and low balance of defect types in fabric defect detection,and the problems of contour blur and color distortion in the generated images during the process of dataset augmentation,an improved cyclic generative adversarial network(CycleGAN)was used for sample augmentation.The CycleGAN generation network was introduced into the DenseNet dense connection structure to enhance feature transfer,and a self-attention mechanism was added to strengthen the focus of the defect area.The discriminator structure design was used to improve the detail judgment ability and improve the detail restoration of the defect area.The experimental results showed that compared with original CycleGAN model,the defect image quality score FID generated by the improved CycleGAN model was reduced from 78.03 to 71.68,and the experimental effect was improved by 8.1％,which verified the validity of the improved CycleGAN model.It is considered that the improved CycleGAN model can expand the fabric defect dataset more clearly and accurately,providing sufficient data support for subsequent fabric defect recognition and detection applications.