基于多特征门控反馈残差网络的超分辨率图像重建算法  

Super-Resolution Image Reconstruction Algorithm Based on Multi-Feature Gated Feedback Residual Network

作　　者：孙红[1] 张玉香 SUN Hong;ZHANG Yuxiang(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《电子科技》2023年第4期65-70,共6页Electronic Science and Technology

基　　金：国家自然科学基金(61472256,61170277,61703277)。

摘　　要：针对超分辨率重建领域中低分辨率图像特征利用不充分的问题,文中基于反馈机制与注意力机制,提出了一种多特征门控反馈残差网络。该网络模型结构简单,以循环的方式实现了网络参数复用,可以有效地节省计算资源。此外,对网络迭代中的输出特征进行保留也可实现多特征融合。采用进一步的特征精炼模块将重建后的高分辨率图像特征进行特征提取,得到了更好的重建效果。在5种测试数据集上的实验结果表明,当缩放因子为4时,该网络的峰值信噪比分别为32.50 dB、28.83 dB、27.75 dB、26.65 dB和31.12 dB。与对比网络相比,文中所提算法的测试结果显著提升。In view of the problem of insufficient feature utilization of low-resolution image in super-resolution reconstruction,a multi-feature gated feedback residual network is proposed based on feedback mechanism and attention mechanism.The network has a simple structure and realizes the reuse of network parameters with a circular way,which can save compute resource effectively.The output features of network iteration are retained to achieve multi-feature fusion.In addition,a further feature refine block is used to extract the reconstructed high-resolution image features to obtain better reconstruction result.Experimental results on five test data sets show that when the scale factor is 4,the peak signal-to-noise ratio of the proposed network is 32.50 dB,28.83 dB,27.75 dB,26.65 dB and 31.12 dB,respectively.Compared with the comparison networks,the test results of the proposed algorithm are significantly improved.

关 键 词：卷积神经网络 残差网络 反馈机制 注意力机制 门控单元 多特征 特征提取 超分辨率重建 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]