A novel fuzzy pretrained neural network architecture for image recolorization on textile products

Abstract

An image recolorization task aims to generate synthetic color images on computers specifically for fabric color design proposals. To enhance processing efficiency and better preserve the original geometric and textile structures, this work proposes a novel fuzzy pretrained neural-network-based image recolorization architecture for fabric and clothing images, consisting of three phases: image decomposition, image segmentation, and image reconstruction. Intrinsic image components, which are initially estimated using a pretrained neural network VOS-Net, formulate an optimization process to achieve high-quality reflectance and shading images. The plug-and-play denoiser network DRU-Net is integrated to smooth these components. The image segmentation phase, which is grounded in the piecewise constant assumption and fuzzy membership functions, partitions image regions to be recolored through only one computation. Recolored images expressed through analytical solutions are produced using the given color palette. Numerical results demonstrate that, compared with classical model-driven methods, our model-data-driven method shows superior efficiency, achieving a remarkable 34.73% reduction in runtime. In terms of quantitative analysis, the proposed method excels at preserving original texture structures and yarn edges with better performance of Learned Perceptual Image Patch Similarity (LPIPS) and structural similarity (SSIM) scores. Furthermore, the color palette distance as an evaluation metric of recoloring indicates that the proposed method effectively reduces the color distortion and color leakage, and the color presentation are closer to the given color palette.