| 摘 要: 为了解决具有错位和遮挡处理条件的高分辨率虚拟试戴(HR-VITON)在处理复杂纹理表现和服装特征交互方面的局限性问题,在基于具有错位和遮挡处理条件的高分辨率虚拟试衣方法的基础上,提出了一种结合多任务判别器与注意力机制的虚拟试衣方法。首先,通过在条件构造器中加入高效通道注意力机制,有效地增强了特征融合;其次,在图像生成网络中采用多任务判别器,以增强对服装渲染的全局和局部尺度评估。通过不断调整网络的学习参数,最终将模型放在数据集VITON-HD Dataset上进行虚拟试衣实验。实验结果表明,与原方法相比,该方法的图像感知相似度(LPIPS)提升了6%、分布距离指标(FID)提升了4.8%,虚拟试衣效果更好。 |
| 关键词: 虚拟试衣;高效通道注意力机制;多任务判别器;特征融合 |
|
中图分类号: TP391.41
文献标识码: A
|
| 基金项目: 浙江省重点研发“领雁”计划项目(2022C01238) |
|
| Research on Virtual Try-On Based on Multi-task Discriminator and Attention Mechanism |
|
WEI Feng1, ZHENG Junhong1,2, HE Lili1,2
|
(1.School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.Zhejiang Provincial Innovat ion Center of Advanced Textile Technology, Shaoxing 312000, China)
sixcandy@126.com; zdzhengjh@sohu.com; llhe@zju.edu.cn
|
| Abstract: In order to address the limitations of High-Resolution Virtual Try-On with Misalignment and Occlusion-Handled Conditions(HR-VITON) in dealing with complex texture representation and garment feature interaction, this paper proposes a virtual try-on method combining multi-task discriminator and attention mechanism based on HRVITON method. Firstly, by incorporating an efficient channel attention mechanism in the conditional constructor, feature fusion is effectively enhanced. Secondly, a multi-task discriminator is adopted in the image generation network to enhance global and local scale evaluation of garment rendering. By continuously adjusting the learning parameters of the network, the model is finally tested on the VITON-HD Dataset for virtual try-on experiments. Experimental results show that compared to the original method, the proposed method improves the Learned Perceptual Image Patch Similarity (LPIPS) by 6% and the Fréchet Inception Distance (FID) by 4.8% , indicating better virtual try-on effect. |
| Keywords: virtual try-on; efficient channel attention mechanism; multi-task discriminator; feature fusion |
