基于姿态的时装图像合成研究

Research on Posture-Based Fashion Image Synthesis

随着人工智能技术在服装时尚领域的深入,服装图像的合成技术成为了当今社会的一个热点研究方向。服装图像包含丰富的语义信息和细节信息,如何根据目标姿态合成服装图像是一个挑战性难题。提出一种新的服装图像合成框架,提出一种姿态与生成对抗网络相结合的图像合成方法。该方法首先通过一种形状编码从原始服装图像提取语义掩模图,从目标图像中提取目标姿态以及pose mask,将它们作为与语义编码器的输入,通过语义生成器合成新的语义掩模图;然后从原始图像提取纹理特征并与语义掩模图融合生成纹理特征图;最后将纹理特征图和语义掩模图融入到纹理生成器中合成新的服装图像。实验结果表明,与其他主流方法相比,此方法在图像合成质量以及定量评估指标上有明显提升。

With the deepening of artificial intelligence technology in the field of clothing fashion,the synthesis technology of clothing images has become a hot research direction in today's society.The clothing image contains rich semantic information and detailed information.How to synthesize the clothing image according to the target posture is a challenging problem.In this paper,a new clothing image synthesis framework is proposed,and an image synthesis method combining posture and generative adversarial network is proposed.The method first extracts the semantic mask image from the original clothing image through a shape coding,extracts the target pose and pose mask from the target image,and uses them as input to the semantic encoder to synthesize a new semantic mask image through the semantic generator;Then extract the texture features from the original image and merge with the semantic mask map to generate the texture feature map;Finally,the texture feature map and the semantic mask map are integrated into the texture generator to synthesize a new clothing image.Experimental results show that compared with other mainstream methods,this method has significantly improved the quality of image synthesis and quantitative evaluation indicators.