Clothing retrieval and matching tasks require the use of model clothing images as input. Due to the limitation of shooting postures and angles, direct using of model images for clothing retrieval or matching often faces many challenges. In view of this, this paper aims to propose a novel tiled clothing image generation model based on improved conditional generative adversarial network (GAN) that can generate clear and accurate tiled clothing images from selected model images.
Aiming at the problems of local information loss and overall structure inaccuracy in tile clothing image generation, this paper optimizes pix2pixHD network model from three aspects: using spatial transformer network (STN) for spatial invariance optimization, using atrous spatial pyramid pooling (ASPP) for feature extraction optimization, using self-attention (SA) for global context information acquisition optimization. The improved network model is called fashion-tile, which can improve the quality and fidelity of tile clothing image generation.
The experimental results show that the proposed method is obviously superior to the existing methods not only in the evaluation metrics, but also in the generating clothing image quality and fidelity. The peak signal-to-noise ratio (PSNR) value is increased by at least 6.6%, the structural similarity (SSIM) value is increased by at least 2.1%, and the Fréchet inception distance (FID) value is reduced by at least 8.6% on the person2cloth dataset.
This work generates high-quality tiled clothing images that enhance the preservation of clothing details and structures, providing consumers with a clearer and more realistic visual experience, thereby increasing shopping satisfaction and purchase intention. With continuous technological advancements and deeper application, the proposed method is expected to play a greater role in the future of clothing e-commerce, offering consumers a richer and more authentic shopping experience.
The proposed method provides an effective solution for generating tiled clothing from model images, which will help to improve the accuracy of subsequent clothing retrieval and matching, and help to enhance the consumers shopping experience and effectively promote sales.
