Fabric image outpainting of colored spun yarn fabrics based on U-Net architecture with fast Fourier convolution

Abstract

Fabric image outpainting can expand small-sized images to generate visually consistent and texture-continuous images for virtual fitting systems. Colored spun yarn fabrics have periodic textures formed by the organizational structure in the weaving process, and random textures formed by the random mixing of fiber colors. Based on the characteristics of colored spun yarn fabrics, this study proposes a novel image outpainting method based on the U-Net architecture with fast Fourier convolution (FFC). The main novelties of the proposed method include: (1) the integration of fast Fourier convolution (FFC) with residual blocks to effectively capture the global periodic textures inherent in fabrics; (2) a novel temporal spatial predictor (TSP) module to model the spatiotemporal evolution of fabric patterns; and (3) a dedicated multitask loss function that ensures texture consistency and visual realism. A dataset of colored spun yarn fabrics is built as the benchmark to evaluate the proposed framework. Experimental results demonstrate that the proposed method is feasible and effective for image outpainting for colored spun yarn fabrics, being superior to the existing methods. The proposed method can be applied to virtual prototyping and e-commerce applications in the textile industry.