级联残差优化Transformer网络的图像超分辨率重建  

Cascade residual-optimized image super-resolution reconstruction in Transformer network

作　　者：林坚普 吴镇城 王崑赋 林志贤[1,2,3] 郭太良 林珊玲 LIN Jianpu;WU Zhencheng;WANG Kunfu;LIN Zhixian;GUO Tailiang;LIN Shanling(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362252,China;Fujian Science and Technology Innovation Laboratory for Optoelectronic Information of China,Fuzhou 350116,China;College of Physics and Telecommunication Engineering,Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福州大学先进制造学院,福建泉州362252 [2]中国福建光电信息科学与技术创新实验室,福建福州350116 [3]福州大学物理与信息工程学院,福建福州350116

出　　处：《光学精密工程》2024年第12期1902-1914,共13页Optics and Precision Engineering

基　　金：国家重点研发计划资助(No.2023YFB3609400);国家自然科学基金青年科学基金(No.62101132)。

摘　　要：为了扩展图像超分辨率算法中卷积神经网络在多个尺度特征上的自适应学习能力,提升网络性能,本文提出一种基于级联残差方法的Transformer网络优化结构进行图像超分辨率重建。首先,该网络采用级联残差结构,增强了网络对低阶和中阶特征的迭代复用和信息共享能力;其次,将通道注意力机制引入Transformer结构中,增强网络的特征表达和自适应学习通道权重的能力;最后,优化Transformer网络结构中的感知模块为级联感知模块,扩展网络深度,增强模型的特征表达能力。在数据集Set5,Set14,BSD100,Urban100和Manga109上进行放大2倍、3倍和4倍的重建测试并与主流方法进行对比,客观评价结果表明,在4倍放大因子的Set5数据集下,本文方法所得图像的峰值信噪比对比其他主流方法平均值提升1.14 dB,结构相似度平均值提升0.019。结合主观评价结果表明,本文方法相比其他主流方法的图像重建效果更好,恢复得到的图像纹理细节更清晰。In order to expand the adaptive learning ability of convolutional neural network in image superresolution algorithm on multiple scale features and improve the network performance,this paper proposed an optimization structure of Transformer network based on cascade residual method for image super-resolution reconstruction.Firstly,the network adopted a cascaded residual structure,which enhanced the iterative reuse and information sharing ability of low and middle order features;Secondly,channel attention mechanism was introduced into Transformer structure to enhance network feature expression and adaptive learning capability of channel weights;Finally,the sensing module in Transformer network structure was optimized as a cascade sensing module to expand the network depth and enhance the feature expression capability of the model.Reconstruction tests of 2x,3x and 4x magnification were carried out on Set5,Set14,BSD100,Urban100 and Manga109 data sets and compared with mainstream methods.Objective evaluation results showed that under Set5 data set with 4x magnification factor,Compared with other mainstream methods,the peak signal-to-noise ratio of the image obtained in this paper is increased by 1.14 dB on average,and the average structural similarity is increased by 0.019.Combined with the subjective evaluation results,it is shown that the proposed method has better image reconstruction effect than other mainstream methods,and the restored image texture details are clearer.

关 键 词：卷积神经网络 图像超分辨率重建 残差网络 TRANSFORMER 注意力机制 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]