MaskFused‐DynamicPerception Generative Adversarial Network (MF‐DPGAN): High‐Quality Fabric Defect Samples Synthesis Algorithm for Improving Visual‐Based Detection Performance

Deep learning–based fabric defect detection has emerged as a pivotal approach to improving textile quality consistency. However, the scarcity of high‐quality defect samples remains a widespread bottleneck that limits the performance of defect detection networks. To address this issue, an innovative hierarchical‐synthesis structure‐based fabric defect samples generation model: the MaskFused‐DynamicPerception Generative Adversarial Network (MF‐DPGAN) was proposed. The dual Defect‐Mask Generator (DM‐G) with hierarchical attention, which decouples defect foregrounds from complex textile backgrounds, focuses on generating defect detail features and enhances the generation quality while suppressing interference from intricate background textures, endowing MF‐DPGAN with flexible and controllable defect migration capability. A multilayer perception module incorporating multibranch mixed dynamic selective kernel mechanism, DSK‐MixMLP block, strengthens the adaptability of MF‐DPGAN to multiscale defect features and morphological variations with complex backgrounds. A composite loss function, L _MF‐DPG , additionally integrating identity loss and mask loss, holistically regulates the balance between generated defects and background textures. Combined with DM‐G and DSK‐MixMLP, it achieves defect‐background decoupling and high‐fidelity defect synthesis. Extensive experiments demonstrate that MF‐DPGAN can flexibly generate various multiscale defect samples with physical plausibility and distribution diversity, achieving a 20.87% improvement in Fréchet Inception Distance (FID). Furthermore, the dataset augmented by MF‐DPGAN is energetically able to train more accuracy models and then significantly enhances the performance of defect detection networks, yielding a state‐of‐the‐art mAP of 98.83%. The designed MF‐DPGAN provides an exceedingly effective solution for roundly facilitating the promotion of defect detection performance with the conditions of scarce defect samples. Furthermore, human‐centered visual perception experiments have confirmed the exceptional subjective authenticity and visual naturalness of the defect images generated by MF‐DPGAN. Additional validation using a dataset constructed from real‐world defect samples demonstrates that MF‐DPGAN maintains competent generalization capability for fabric defect generation even in complex industrial environments.