基于多级连续编码与解码的图像超分辨率重建算法  

Image super-resolution reconstruction algorithm based on multi-level continuous encoding and decoding

作　　者：宋昭漾 赵小强[1,2,3] 惠永永 蒋红梅[1,2,3] SONG Zhao-yang;ZHAO Xiao-qiang;HUI Yong-yong;JIANG Hong-mei(College of Electrical Engineering and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Key Laboratory of Gansu Advanced Control for Industrial Processes,Lanzhou 730050,China;National Experimental Teaching Center of Electrical and Control Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学电气工程与信息工程学院,甘肃兰州730050 [2]甘肃省工业过程先进控制重点实验室,甘肃兰州730050 [3]兰州理工大学国家级电气与控制工程实验教学中心,甘肃兰州730050

出　　处：《浙江大学学报（工学版）》2023年第9期1885-1893,共9页Journal of Zhejiang University：Engineering Science

基　　金：国家重点研发计划资助项目(2020YFB1713600);国家自然科学基金资助项目(61763029);甘肃省科技计划资助项目(21YF5GA072,21JR7RA206);甘肃省教育厅产业支撑计划资助项目(2021CYZC-02)。

摘　　要：以卷积神经网络为模型框架的图像超分辨率重建算法难以提取低分辨率图像内部的多层次特征信息,导致重建图像缺少丰富细节,为此提出新的图像超分辨率重建算法.所提算法通过初始卷积层从低分辨率图像提取浅层特征;通过多个端对端连接的多级连续编码与解码的注意力残差模块获取低分辨率图像内部不同层级的图像特征,按照不同的提取难度生成这些特征的权重,重新校准不同层次的图像特征,获取图像内部丰富的细节特征;通过上采样模块和重建卷积层将提取到的丰富细节特征和浅层特征重建成高分辨率图像.在Set5、Set14、BSD100和Urban100测试集上进行的对比测试结果表明,使用所提算法重建的图像在客观评价指标、视觉效果上均优于使用主流算法重建的图像.A new image super-resolution reconstruction algorithm was proposed,aiming at the problem that the image super-resolution reconstruction algorithms by using convolutional neural network as the model framework was difficult to extract multi-level feature information inside a low-resolution image,resulting in the lack of rich details in the reconstructed image.The proposed algorithm extracted shallow features from low-resolution images through initial convolution layer.Image features of different levels in a low-resolution image were obtained through a plurality of end-to-end connected multi-level continuous encoding and decoding attention residual modules,the weights of these features were generated according to different extraction difficulties,and the image features of different levels were recalibrated to obtain rich detailed features in the image.Through the up-sampling module and reconstruction convolution layer,the extracted rich detailed features and shallow features were reconstructed into high-resolution images.The comparative test results on the test sets of Set5,Set14,BSD100 and Urban100 show that the image reconstructed by the proposed algorithm is superior to the image reconstructed by a mainstream algorithm in objective evaluation index and visual effect.

关 键 词：超分辨率重建 卷积神经网络 多级连续编码与解码 注意力 多层次特征信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]