基于多尺度注意力残差网络的图像超分辨率重建

Image Super-resolution Reconstruction Based on Multi-scale Attention Residual Network

数字图像在传递信息中起着重要的作用,图像超分辨率技术能丰富图像的细节信息.针对许多网络对低分辨率图像的有效特征复用不足和参数量过大的问题,本文结合不同大小的卷积核以及注意力残差机制构建图像超分辨率网络,用3个有差别尺度的卷积层来提取图像的特征,其中第2和第3层用小卷积核替代大的卷积核,对3层卷积融合之后引入注意力机制,最后用传统的Bicubic插值直接给网络提供低频信息.在减小参数量和减轻梯度消失的同时,让有效的高频信息得到更大的权重且能增强网络之间的非线性表达能力,这有利于网络训练的迭代收敛.实验结果表明,基于多尺度注意力残差网络能够在一定程度上增强图像的重建能力.

Digital images play an important role in information transmission,and image super-resolution technology can enrich image details.To address the problems of insufficient effective feature reuse of low-resolution images and excessive parameters in many networks,this study combines convolution kernels of different sizes and attention residual mechanism to construct the image super-resolution network.Three convolution layers of different scales are used to extract the image features,of which the second and third layers replace the large convolution kernels with small ones,and after the three-layer convolution fusion,the attention mechanism is introduced.Finally,the traditional Bicubic interpolation is used to directly provide low-frequency information for the network.By doing this,while reducing the number of parameters and mitigating the disappearance of gradients,the proposed network can make the effective highfrequency information gain greater weights and can enhance the nonlinear expression ability between the networks,which is conducive to the iterative convergence of network training.Experimental results show that the proposed network can enhance the image reconstruction ability to a certain extent.