基于生成对抗网络的女上装图像属性编辑

Image Attribute Editing of Womens Tops Based on Generating Adversarial Networks

为解决当前服装图像属性编辑模型生成图像存在属性缺失或冗余的问题,提出一种基于Fashion-AttGAN的优化模型对女上装图像细节进行变换的设计方法;通过优化特征提取网络,将结构相似性损失项加入重构损失,提高生成器的属性编辑能力;使用CP-VTON数据集训练,对女上装图像中袖长和颜色的细节进行调整。结果表明,生成图像在袖型连贯性和颜色准确性方面得到提升,改进模型收敛趋势更平稳,重构图像的结构相似性指标提升了27.4%,峰值信噪比提高了2.8%。该优化模型有效减少了生成图像的属性冗余和残缺,为服装图像细节变换研究提供参考。

In order to solve the problem of attributes missing or redundant in the current clothing image attribute editing models of generate images,a design method based on Fashion-AttGAN model was conducted to transform the details of womens tops.This paper optimized feature network and added structure similarity index measure to the reconstructed loss function to improve the attribute editing ability of generator.The CP-VTON dataset was used for training to ultimately achieve fine-grained editing of womens tops sleeve length and color.The experimental results show that the generated image achieves the improvement in sleeve coherence and color accuracy,the improved model is shown to move more smoothly towards convergence trend,the reconstructed image structure similarity index measure realizes the growth of 27.4%and peak signal-to-noise ratio grows by 2.8%.The proposed model reduces atributes missing or redundant in generated images and provides a technical reference for its detail transformation.