基于多尺度与残差网络的图像超分辨率重建  

Super-Resolution Reconstruction of Images Based on Multi-scale and Residual Network

作　　者：魏一铭 徐岩[1] 王汇丰 韦春苗 WEI Yiming;XU Yan;WANG Huifeng;WEI Chunmiao(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730000,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730000

出　　处：《电光与控制》2022年第7期102-107,共6页Electronics Optics & Control

基　　金：国家自然科学基金(62063014)。

摘　　要：针对目前图像超分辨率重建算法在提取图像信息时过于单一地累加卷积层导致的深层网络梯度消失和图像信息丢失的问题,提出了一种基于多尺度与残差网络的图像超分辨率重建算法。该算法使用多尺度密集连接的卷积核代替单一累加的卷积核,充分提取输入低分辨率图像信息并实现通道特征维度的复用;利用残差网络,多层次地对丢失图像信息进行补充并对深层网络模型的梯度问题实现了抑制,在反向传播的过程中帮助全网络模型自适应地完成对权重更新;最后以非线性映射的方式输出最终重建图像。实验表明,所提算法在测试集上的峰值信噪比和结构相似性与对比算法相比均有所提升;在与目前主流算法对比中获得了细节信息更加丰富、边缘纹理更加清晰的重建图像。To solve the problems of deep network gradient disappearance and image information loss caused by superimposition of simplex convolution layer when extracting image information in current image super-resolution reconstruction algorithmsan image super-resolution reconstruction algorithm based on multi-scale and residual network is proposed.The proposed algorithm uses multi-scale dense connection convolution kernel instead of the simplex accumulated convolution kernel to fully extract the low-resolution image input information and realize reuse of channel feature dimension.The residual network is used to supplement the lost image information at multiple levels and suppress the gradient problem of the deep network modelwhich helps the whole network model to adaptively update the weights in the process of reverse propagation.In the endthe final reconstructed image is output through nonlinear mapping.Experiments show that:1)The peak signal-to-noise ratio and structural similarity of the proposed algorithm on the test set are improved compared with that of the contrast algorithms;and 2)In comparison with the current mainstream algorithmsthe proposed algorithm obtains a reconstructed image with richer detail information and clearer edge texture.

关 键 词：图像处理 神经网络 多尺度特征 密集连接 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]