基于深度残差网络的图像超分辨率算法  

An Image Super-Resolution Algorithm Based on Deep Residual Network

作　　者：王剑峰[1] 彭中 WANG Jianfeng;PENG Zhong(College of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China)

机构地区：[1]青岛科技大学信息科学技术学院,山东青岛266061

出　　处：《青岛科技大学学报（自然科学版）》2022年第4期113-119,共7页Journal of Qingdao University of Science and Technology:Natural Science Edition

基　　金：国家自然科学基金项目(61672305)。

摘　　要：单幅图像超分辨率(SR)是从单个低分辨率(LR)图像推断出高分辨率(HR)图像的任务。大多数基于卷积神经网络(CNN)的深层SR模型没有充分利用原始LR图像的分层特征,性能相对较低。同时,将像素级误差作为损失函数,易产生缺乏高频纹理、过度平滑的图像。针对以上问题,本工作提出了改进的基于深度残差网络的图像超分辨率算法,通过残差密集块,获得稠密的局部特征,使用全局特征融合以全局方式自适应地学习全局分层特征;使用更合理的损失函数组合,结合分割概率图和空间特征变换模块,利用对抗学习的原理训练CNN,优化感知相似度。实验结果表明,本工作方法在高缩放因子下实现了图像质量的显著提升,深度残差网络可以重建出更逼真、视觉上更合理的纹理。Single image super-resolution(SR)is a task to infer high-resolution(HR)images from a single low-resolution(LR)image.Most of the deep SR models based on convolutional neural networks do not make full use of the hierarchical features of the original LR images,and their performance is relatively low.At the same time,taking the pixel-level error as the loss function,it is easy to produce images that are lack of high-frequency texture and over-smooth.To solve these problems,this paper proposes an improved image super-resolution algorithm based on depth residual network,which obtains dense local features through dense residual blocks,and adaptively learns global hierarchical features by using global feature fusion.Using a more reasonable combination of loss functions,the semantic segmentation probability graph and spatial feature transformation module are merged;finally using the principle of adversarial learning to train CNN to optimize perceptual similarity.The experimental results show that this method can significantly improve the image quality under high scaling factor,and the depth residual network can reconstruct a more realistic and visually reasonable texture.

关 键 词：超分辨率图像重建 深度残差网络 感知度量 类别先验 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]