删除或更新信息,请邮件至freekaoyan#163.com(#换成@)

梯度指导的快速轻型超分辨率重建密集残差网络

本站小编 Free考研考试/2022-01-16

李素梅,马 力,石永莲
AuthorsHTML:李素梅,马 力,石永莲
AuthorsListE:Li Sumei,Ma Li,Shi Yonglian
AuthorsHTMLE:Li Sumei,Ma Li,Shi Yonglian
Unit:天津大学电气自动化与信息工程学院,天津 300072
Unit_EngLish:School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China
Abstract_Chinese:近来,深度卷积神经网络在单图像超分辨率重建中取得了显著进展,但是,随着网络深度和宽度的增 加,计算复杂度和内存消耗也随之增加. 此外,超分辨率重建图像的边缘模糊以及伪影等问题也是目前方法不能很 好解决的难点之一. 针对上述问题,提出一种快速轻型超分辨率重建模型. 该模型由一个 3 层的浅通道和一个 29 层 的深通道构成,在结构的末端使用卷积层将深浅通道进行融合;浅通道主要用于恢复图像的整体轮廓,保留原图像 的轮廓信息;深通道主要用于恢复图像的高频细节信息,采用多尺度滤波器提取不同尺度的纹理,增加提取信息的 丰富度.模型结合密集块、多跨度残差连接降低了参数量、提高了网络收敛速度;在特征提取阶段,模型结合分组卷 积提出其增强模型;模型的损失函数基于梯度损失与 MAE 损失,能够有效改善重建图像的边缘细节以及伪影问题. 在基准数据集上的实验结果表明,所提模型在放大倍数为 2、3 任务上比现有的 IDN、MemNet、DRRN 等代表性模 型具有更好的性能,特别是在参数量、计算复杂度、重建速度及改善图像边缘、去伪影等方面均优于现有具有代表 性的算法;结合分组卷积的增强模型参数量仅有 330 210,是 IDN 参数量的 46%;在 Urban100 数据集上放大倍数为 2 的重建速度比 MemNet 快 420 倍,很好地满足了快速准确的重建要求.
Abstract_English:Recently, the deep convolutional neural network (CNN) has made remarkable progress in single image super-resolution (SISR). However,as the depth and width of the networks increase,the computational complexity and memory consumption also increase. Moreover,the problems of the blurred edge and artifacts in reconstructed images are difficulties that the current methods cannot adequately resolve. To address these problems,a fast and light weight super-resolution model was proposed. The model consisted of a 3-layer shallow channel and a 29-layer deep channel,fused by a convolutional layer. The shallow channel mainly restored the overall contour of the image and retained the outline information of the original image. The deep channel restored high-frequency detail information. It used multi-scale filters to extract textures at different scales and increase the richness of information. The model com\u0002bined dense blocks with multi-span residual connections to reduce the number of parameters and speed up the conver\u0002gence of network. In addition,the model was optimized,and vanilla convolutions were replaced with group convo\u0002lutions at the feature extraction stage in the enhanced model. The loss function of the model was based on gradient loss and MAE loss,which can effectively improve the edge details and artifacts in the reconstructed image. The experimental results on the benchmark datasets show that the proposed model outperforms representative models such as IDN,MemNet,and DRRN on ×2 and ×3 tasks. In particular,the proposed model is superior in terms of networkparameters, computational complexity, and reconstruction speed,and it is better at improving image edges and removing artifacts. In addition,the network parameters of the enhanced model are only 330 210 parameters,which is 46% of that of IDN. For the ×2 task,the reconstruction speed of the enhanced model on the Urban100 dataset is 420 times faster than that of MemNet,which satisfies the requirement of a fast and accurate reconstruction.
Keyword_Chinese:图像超分辨率重建;轻型结构;密集网络;残差网络;梯度损失
Keywords_English:image super-resolution reconstruction;lightweight structure;dense network;residual network; gradient loss

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6611
相关话题/网络 梯度