张淑芳,王沁宇
AuthorsHTML:张淑芳，王沁宇
AuthorsListE:Zhang Shufang,Wang Qinyu
AuthorsHTMLE:Zhang Shufang，Wang Qinyu
Unit:天津大学电气自动化与信息工程学院，天津，300072
Unit_EngLish:School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China
Abstract_Chinese:为了解决传统虚拟试穿方法存在的手臂遮挡与细节模糊问题，提升重建图像的视觉质量，提出一种基于生 成对抗网络的虚拟试穿方法．通过纹理提取模块和残差样式编码模块提取服装细节信息，并结合人体表征输入与人 物姿势来重建试穿图像，解决了手臂遮挡问题，实现了对扭曲失误服装的修复还原，且重建图像服装边缘清晰．定 性分析表明，改进虚拟试穿方法得到的重建图像能清楚地展示试穿人物的手臂部分与服装纹理细节，具有很好的视 觉逼真度和视觉质量．定量分析表明，该方法结构相似性指标提升了 8.56%，与原始参考的像素结构更相似；感知 相似性指标减少了 5.24%，与原始参考的卷积特征更相似；Inception 分数提升了 0.95%，具有更高的清晰度和更好 的多样性．
Abstract_English:To solve the problems of arm occlusion and detail blurring in traditional virtual try-on networks，a new virtual try-on method based on generative adversarial networks is proposed. The information on clothing details was extracted and encoded using the texture extraction and residual style encoding modules，respectively，and the try-on image was reconstructed using the extracted clothing information，the target pose，and the human representation as inputs. Our method could solve the arm occlusion problem，repair distorted garments，and generate images with clear details. Qualitative analysis showed that the try-on images reconstructed by our method could clearly show the model’s arm and clothing texture details with good visual fidelity and quality. Meanwhile，a quantitative analysis showed that using our method，the SSIM improved by 8.56%，which is similar to the original clothing’s pixel structure；LPIP reduced by 5.24%，which is similar to the ground truth’s convolution features；and inception score is improved by 0.95%，which has better definition and diversity.
Keyword_Chinese:图像重建技术；虚拟试衣；图像分析；生成对抗网络
Keywords_English:image reconstruction techniques；virtual try-on；image analysis；generative adversarial network

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6678