| 摘 要: 针对虚拟试衣中特征提取不足、人物肢体被衣服遮挡的问题，在基于图像特征保留的虚拟试衣方法基础上，提出基于并行卷积核的Attention U-Net虚拟试衣方法。该方法采用并行卷积核代替原有的3×3卷积核来提取特征，并在U-Net网络中融入注意力机制形成新的Attention U-Net图像合成器，通过不断调整网络学习参数，将模型放在数据集VITON Dataset上进行虚拟试衣实验。实验结果表明，与原方法相比，该方法能提取出更多的细节纹理，在结构相似性上提升了15.6%，虚拟试衣效果更好。 |
| 关键词: 虚拟试衣；特征提取；并行卷积核；注意力机制；结构相似性 |
|
中图分类号: TP391.41
文献标识码: A
|
| 基金项目: 绍兴市技术创新计划(揭榜挂帅)项目(2020B41006). |
|
| Research on Attention U-Net Virtual Try-On Method based on Parallel Convolution Kernel |
|
SHU Xingzhe
|
(School of Information, Zhejiang Sci-Tech University, Hangzhou 310018, China)
1036413161@qq.com
|
| Abstract: Virtual try-on has problem of insufficient feature extraction in and people's limbs being covered by clothes. On the basis of the virtual try-on method with image feature retention, this paper proposes an Attention U-Net virtual try-on method based on parallel convolution kernel. In this method, parallel convolution kernel is used to replace the original 3×3 convolution kernel to extract features, and the attention mechanism is integrated into the u-net network to form a new Attention U-Net image synthesizer. By constantly adjusting the network learning parameters, the model is placed on the data set VITON (Virtual Try-On Network) Dataset for virtual fitting experiment. Experimental results show that compared with the original method, the proposed method can extract more detailed textures, improve the structural similarity by 15.6%, and the virtual fitting effect is better. |
| Keywords: virtual try-on; feature extraction; parallel convolution kernel; attention mechanism; structural similarity |
