Abstract:

Single image super-resolution plays an important role in the field of computer vision. This technology aims to reconstruct high-resolution images from low-resolution images. In recent years, deep neural networks make performance in SISR task significantly improved. However, recently works based on convolutional neural network equally treat high-frequency and low-frequency features, which makes the reconstruction of high-frequency details poor, the output too smooth and the texture information lack. On the other hand, very deep convolutional network is not easy to converge, and as the depth of the neural network grows, the long-term information from the former layer can easily be weakened or lost in the latter layer, which makes the benefit not proportional to the depth of the network and the computational complexity. To solve these above problems, it proposes a spatial attention module and a channel attention module as the basic block of convolutional neural network for SISR. Firstly, in the same channel, the information of different locations is given different weights by the spatial attention module. Secondly, the weights between different channels are determined by the channel attention module, which makes the high-frequency information gain a higher position in the reconstruction task. The reconstruction performance is improved. It further proposes a short-term and long-term feature modulation module to transform the layer depth of the network into the block depth, which greatly reduces the depth of the network, in order to solve the problem of long-term information loss in the front layer. Compared with other methods based on deep convolution neural network, the Peak Signal-to-Noise Ratio（PSNR） on several benchmark datasets are better, which proves the effectiveness of the proposed method.

Key words: channel attention, spatial attention, convolutional neural network;image super-resolution reconstruction

摘要：

单幅图像超分辨率（Single Image Super Resolution，SISR）在计算机视觉领域占有重要地位，该技术旨在从低分辨率图像中重建出高分辨率图像。近年来，深度神经网络在SISR领域起到了至关重要的作用，然而，目前利用卷积神经网络平等地对待高频与低频特征，使得高频细节的重建表现不佳，输出过于平滑，缺少纹理信息。另一方面，过于深的网络不容易收敛，并且随着神经网络的深度增长，来自前一层的长期信息很容易在后期层中减弱或丢失，使得重建收益不能正比于网络的深度与计算复杂度。针对以上问题，对用于SISR的卷积神经网络的基本块提出了空间注意力模块与通道注意力模块，在同一通道中，不同位置的信息被空间注意力模块赋予不同的权重，不同通道间的权重由通道注意力模块决定，这使得高频信息在重建任务中获得更高的地位，提高了重建指标。进一步地提出了长期特征调制模块将网络的层深度转化为块深度，大大缩小了网络深度，以解决前层长期信息的丢失问题。在Set5等多个基准数据集上的峰值信噪比（PSNR）均比目前其他基于深度卷积神经网络的方法有所提升，这证明了提出的方法的有效性与先进性。

关键词: 通道注意力, 空间注意力, 卷积神经网络, 图像超分辨率重建