基于PATN和自注意力机制的多分辨率人体姿态迁移方法

打开文本图片集

Multi-resolution human posture transfer algorithm based on PATN and self-attention mechanism

HUANG Qingdong 1，2* ， SU Yuhui1，LIU Yihua¹， CHEN Zihuang1，YAO Yongqil

（1. School ofCommunications and Information Engineering，Xi'an University ofPosts and Telecommunications，Xi'an 710l21，China; 2. Shaanxi Key Laboratory ofInformation Communication Network and Security， Xi'an 71Ol21，China）

Abstract：Aiming at the problem that the existing human pose transfer methods suffer from image deformation and distortion due to improper feature processing in the encoding stage，a multi-resolution human pose transfer method based on pose-attentional transfer network （PATN） and self-attention mechanism is proposed.Firstly，a pose-guided self-atention module is designed to enhance the weight of the feature channel of the key body region through the multi-head atention mechanism，reduce the influence of backgroundirrelevant features，and adaptively explore the correlation between the two branch features.Secondly，a multi-scale attention module is added in the decoding stage to enhance the expression of pose information at different scales，efectively improve the fidelity of local details and overalltexture.Finally，the ternary pixel loss is introduced to constrain the generated image，which improves the feature consistency and structural consistency of the image. The experimental results are verified on the DeepFashion and Market1501 datasets.They show that the proposed method is superior to the existing PATN method in terms of structural similarity（SSIM），initial score （IS）and perceptual similarity（LPIPS），and has improved visual perception and edge texture，showing important potential inthe downstream task of person re-identification.

Key words： image processing； pose transfer; CGAN； self-attention mechanism；multi-resolution

1引言

随着深度生成模型的不断演进，人体姿态迁移（HumanPostureTransfer，HPT）技术近年来在智能照片编辑、行人重识别等领域得到广泛应用，成为当前计算机视觉领域的研究热点。（剩余11128字）