Abstract:

Face super-resolution reconstruction is a low-cost technique for obtaining low-resolution face processing to obtain high-resolution face, also known as face illusion. In order to make the reconstructed face image have clearer detail texture, this paper proposes a densely connected face super-resolution network based on attention mechanism by studying the face-based super-division algorithm based on deep learning. The algorithm consists of two parts：feature extraction and image  reconstruction. By focusing on the information of the feature channel domain and the spatial domain, the Multi Attention Domain Module（MADM） is proposed. Among them, by changing the mutual relationship and weight of the channel and space, the features are weighted and reorganized adaptively, and the dense layered and long and short connections are used to fuse the features of different layers to improve network performance. The experimental results verify the correctness of the proposed algorithm. Compared with the existing algorithms, the superior performance of the proposed algorithm is demonstrated. The reconstructed face image has clearer texture details.

Key words: attention mechanism, dense connection, face super-resolution, feature fusion, neural network

摘要：

人脸超分辨率重建是一种对低分辨率人脸处理获取对应高分辨率人脸的低成本技术，又称人脸幻生。为了使重建的人脸图像有更清晰的细节纹理，通过对基于深度学习的人脸超分算法的研究，提出了基于注意力机制的稠密连接人脸超分算法。该算法主要由特征提取和图像重建两个部分组成，通过同时关注特征通道域和空间域的信息，创建了多注意力域模块MADM（Multi Attention Domain Module）。其中，通过改变信道和空间上的相互关系和权重，自适应地对特征进行加权重组，并且使用密集的稠密连接和长短连接将不同层的特征融合在一起，实现提升网络性能。实验结果验证了该算法的正确性；并与现有算法比较，表明了该算法的优越性能，重建的人脸图像具有更清晰的纹理细节特征。

关键词: 注意力机制, 稠密连接, 人脸超分, 特征融合, 神经网络