基于蓝图可分离残差蒸馏网络的图像超分辨率重建  

Super-resolution Reconstruction of Images Based on Blueprint Separable Residual Distillation Network

作　　者：熊荣盛 王帮海[1] 杨夏宁 Xiong Rong-sheng;Wang Bang-hai;Yang Xia-ning(School of Computer Science and Technology,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《广东工业大学学报》2024年第2期65-72,共8页Journal of Guangdong University of Technology

基　　金：国家自然科学基金资助项目(62072119)。

摘　　要：标准卷积的单图像超分辨率重建性能受限于堆叠网络层的冗余性,算法难以实施,特征提取层单一的残差结构也无法高效地利用卷积得到的特征信息。为改善上述问题,本文改进残差蒸馏结构,提出残差蒸馏复用模块,以减少残差蒸馏过程中图像高频信息的损失;此外,将基础残差块替换为蓝图可分离卷积,解耦特征图的空间相关性,以降低高相关性特征的权重,提高卷积的效率,降低参数量。为验证算法的性能,在Set5等标准数据集中对算法进行验证。实验结果表明,该算法模型的峰值信噪比(Peak Signal-To-Noise Ratio, PSNR)和结构相似度(Structural Similarity, SSIM)相比于基于残差蒸馏网络的轻量级图像超分辨率重建网络分别有0.06~0.25 d B与0.004~0.012的提升。The performance of single image super-resolution reconstruction based on standard convolution is limited by the redundancy of the stacked network layers,making it difficult to implement the algorithm on the ground.Moreover,the single residual structure of the feature extraction layer cannot efficiently utilize the feature information obtained from convolution.To address these,this paper proposes a residual distillation reuse module based on the existing residual distillation-based structure to reduce the high-frequency information of the image lost in the residual distillation process.In addition,the base residual block is replaced by a blueprint separable convolution to decouple the spatial correlation of the feature map,such that the weight of highly correlated features can be reduced.As a result,the efficiency of convolution can be improved and the number of parameters can be reduced.We conduct comparative experiments on standard datasets such as Set5 to evaluate the proposed algorithm.The experimental results show that the peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)of the proposed algorithm can be improved by approximately 0.06~0.25 dB and 0.004~0.012,respectively,over the lightweightresidual distillation image super-resolution networks.

关 键 词：图像超分辨率重建 残差蒸馏 蓝图可分离卷积 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]