基于密集反馈网络的单幅图像超分辨率重建  

Super-Resolution Reconstruction of a Single Image Based on Dense Feedback Network

作　　者：刘锡泽 范红[1] 海涵 王鑫城 许武军[1] 倪林[1] LIU Xize;FAN Hong;HAI Han;WANG Xincheng;XU Wujun;NI Lin(College of Information Science and Technology,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620

出　　处：《计算机工程》2021年第11期254-261,共8页Computer Engineering

基　　金：国家自然科学基金(61801106)。

摘　　要：基于深度学习的单幅图像超分辨率网络模型体积庞大,导致参数利用率低且难以部署,对中间层特征利用不充分。提出一种密集反馈注意力网络(DFAN)模型。在同一特征图中通过多尺度残差注意力模块(MRAB)提取不同尺度的深层特征,以增加特征的多样性。同时将每个MRAB的输出均作为同组中其他残差模块的输入,使各层之间的信息流最大化,从而减小训练难度。实验结果表明,相比VDSR、DRRN、MemNet等模型,DFAN模型具有较优的重建效果,其在重建放大倍数为4的Set5数据集上计算复杂度仅为VDSR模型的0.14倍左右,而峰值信噪比提高了0.57 dB。The existing deep learning-based models for the super-resolution reconstruction of a single image are large-sized,which reduces the utilization of parameters and middle-layer features,and cause difficulty to model deployment.To address the problems,a model is proposed based on Dense Feedback Attention Network(DFAN).This model uses multi-scale residual modules with attention mechanism to extract deep features of different scales in the same feature map,so the diversity of features is increased.The output of each residual module is broadcast to the input of other residual modules in the same group,so that the information flows between layers are maximized,thereby reducing the difficulty of training.Experimental results show that the proposed network exhibits better reconstruction effects than VDSR,DRRN,MemNet and other models.When reconstructing the Set5 dataset at an amplification factor of 4,the Peak Signal to Noise Ratio(PSNR)of DFAN is 0.57 dB higher than that of VDSR,while the computational complexity of DFAN is only about 0.14 times that of VDSR algorithm.

关 键 词：单幅图像超分辨率重建 深度学习 密集反馈模型 注意力机制 残差模块 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]