摘要
基于深度学习的真实图像超分辨率(super-resolution, SR)重建算法目前存在参数量过大的问题,为解决该问题,提出了一种多尺度残差特征融合的轻量级真实图像SR重建算法。首先利用深度可分离卷积和复用卷积针对多尺度特征提取块进行改进,在提取特多尺度特征的同时实现了模块的轻量化,参数量仅为改进前的7.5%。其次使用残差特征融合操作将4个多尺度深度可分离特征提取块(multi-scale depthwise separable block, MSDSB)聚合成一个残差特征融合块,以减少残差路径长度。然后使用增强型注意力模块从通道和空间维度进行自适应调整以提升算法性能。最后使用自适应上采样模块获得SR重建图像。在消融实验中,本文算法重建性能超过原始算法,且参数量仅为3.53×106,是原始算法的34.5%。在对比实验中,其重建性能超过了当前主流算法,与组件分而治之(component divide-and-conquer, CDC)算法相比,PSNR和SSIM指标分别提升了0.01 dB与0.001 0,且参数量仅为组件CDC算法的8.84%,在保证重建性能的同时实现了算法的轻量化。
At present, there exists too many parameters amount in the real-world image super-resolution(SR) reconstruction algorithm based on deep learning.To solve this problem, a lightweight real-world image SR reconstruction algorithm based on multi-scale residual feature aggregation is proposed.First, the depthwise separable convolution and multiplexing convolution are used to improve the existing multi-scale feature extraction block, which achieves the lightweight of the module while extracting the extra-multi-scale feature, with only 7.5% of the parameters amount before the improvement.Next, the residual feature aggregation is exploited to aggregate the 4 multi-scale depthwise separable blocks amount(MSDSB) into a residual feature aggregation block to reduce the length of the residual path.Then, the enhanced attention module is utilized to adaptively adjust the channel and spatial dimensions to improve the performance of the algorithm.Finally, the adaptive upsampling module is used to obtain SR reconstructed images.In ablation experiments, the reconstruction performance of the algorithm is better than that of the original algorithm, and the parameters amount is only 3.53×10~6,which is 34.5% of the original algorithm.In the comparative experiments, the reconstruction performance of the proposed algorithm is better than the current mainstream algorithm. Compared with the component divide-and-conquer(CDC) algorithm, the PSNR and SSIM indexes of the presented algorithm are increased by 0.01dB and 0.001 0,respectively, and the parameters amount is only 8.84% of that of the CDC algorithm. The lightweight of the algorithm is realized while ensuring the reconstruction performance.
作者
吕佳
许鹏程
LV Jia;XU Pengcheng(College of Computer and Information Sciences,Chongqing Normal University,Chongqing 401331,China;Chongqing Center of Engineering Technology Research on Digital Agriculture Service,Chongqing Normal University,Chongqing 401331,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2023年第2期120-131,共12页
Journal of Optoelectronics·Laser
基金
国家自然科学基金重大项目(11971084)
重庆市高校创新研究群体资助(CXQT20015)
重庆市教委科研项目重点项目(KJZD-K202200511)
重庆市科技局技术预见与制度创新项目(2022TFII-OFX0265)资助项目。
关键词
真实图像
图像超分辨率(SR)重建
卷积神经网络
深度可分离卷积
残差特征融合
real-world image
image super-resolution(SR)reconstruction
convolutional neural network
depthwise separable convolution
residual feature aggregation
作者简介
吕佳(1978-),女,博士,教授,硕士生导师,主要从事机器学习、数据挖掘及其在医学图像处理等方面的研究,E-mail:lvjia@cqnu.edu.cn。
