基于多频特征和纹理增强的轻量化图像超分辨率重建  

Lightweight image super-resolution reconstruction with multi-frequency feature and texture enhancement

作　　者：刘媛媛[1] 张雨欣 王晓燕[1] 朱路[1] Liu Yuanyuan;Zhang Yuxin;Wang Xiaoyan;Zhu Lu(College of Information Engineering,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学信息工程学院,南昌330013

出　　处：《计算机应用研究》2024年第8期2515-2520,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61967007,61963016);江西省重点研发计划重点资助项目(20201BBF61012)。

摘　　要：现有基于卷积神经网络主要关注图像重构的精度,忽略了过度参数化、特征提取不充分以及计算资源浪费等问题。针对上述问题,提出了一种轻量级多频率特征提取网络(MFEN),设计了轻量化晶格信息交互结构,利用通道分割和多模式卷积组合减少参数量;通过分离图像的低频、中频以及高频率信息后进行特征异构提取,提高网络的表达能力和特征区分性,使其更注重纹理细节特征的复原,并合理分配计算资源。此外,在网络内部融合局部二值模式(LBP)算法用于增强网络对纹理感知的敏感度,旨在进一步提高网络对细节的提取能力。经验证,该方法在复杂度和性能之间取得了良好的权衡,即实现轻量有效提取图像特征的同时重建出高分辨率图像。在Set5数据集上的2倍放大实验结果最终表明,相比较于基于卷积神经网络的图像超分辨率经典算法(SRCNN)和较新算法(MADNet),所提方法的峰值信噪比(PSNR)分别提升了1.31 dB和0.12 dB,参数量相比MADNet减少了55%。Existing studies based on convolutional neural networks mainly focus on the accuracy of image reconstruction,ignoring problems such as excessive parameters,insufficient feature extraction,and resource waste.In response to the above,this paper proposed multi-frequency feature extraction network(MFEN),which designed a lightweight lattice information interaction structure and used channel segmentation with multi-mode convolution combination to reduce the number of parameters.By separating the low-frequency,mid-frequency,and high-frequency information of the image and extracting the feature heterogeneity,it improved the expressiveness and feature differentiation of the network,made the network pay more attention to the restoration of texture detail features,and reasonably allocated the computational resources.In addition,it integrated the local binary pattern(LBP)algorithm into the network to enhance texture sensitivity,which further improved the network’s ability to extract details.It experimentally verifies that the proposed method balances complexity and performance well.In the 2X zooming experiments on the Set5 dataset,compared to the conventional image super-resolution algorithm(SRCNN)based on convolutional neural network and the newer algorithm(MADNet),the peak signal-to-noise ratio(PSNR)of the proposed method is improved by 1.31 dB and 0.12 dB respectively,and the number of parameters is reduced by 55%compared to MADNet.

关 键 词：图像超分辨率重建 卷积神经网络 轻量化 多频率特征提取 局部二值模式算法 

分 类 号：TP911.73[自动化与计算机技术]