Texture Image Generation for Carbon Fiber-Reinforced Composites Based on Periodic Fine-grained Texture Generation Network

To address the issue of insufficient texture image data during the quality inspection of Carbon Fiber-Reinforced Composites (CFRCs) layup process， we propose a periodic fine-grained texture generation network (PFTN) and a defect transfer PFTN (DT-PFTN) with defect transfer functionality. The PFTN effectively balances the randomness and periodicity of texture images by introducing a fine-grained texture compensation module and a learnable periodic sine wave， ensuring the generation of high-quality， fine-grained CFRCs texture images. The defect transfer network DT-PFTN generates fiber texture images with real surface defects by effectively encoding and embedding defect features. We contruct a CFRCs texture image dataset using the proposed models， including 12 480 normal texture image samples and 3 000 defective texture image samples. Experimental results demonstrate that both the PFTN and DT-PFTN models achieve competitive texture image generation effects on both defect-free and defective datasets， with SIFID (Single Image Fréchet Inception Distance)(10^-6)， LPIPS (Learned Perceptual Image Patch Similarity)， and FID (Fréchet Inception Distance) scores reaching 9， 0.35 and 60.81， respectively.