基于密集连接与激励模块的图像超分辨网络  

Image Super-Resolution Network Based on Dense Connection and Squeeze Module

作　　者：胡诗语 王国栋 赵毅 王岩杰 潘振宽 Hu Shiyu;Wang Guodong;Zhao Yi;Wang Yanjie;Pan Zhenkuan(College of Computer Science and Technology,Qingdao University,Qingdao,Shandong 266071,China)

机构地区：[1]青岛大学计算机科学技术学院

出　　处：《激光与光电子学进展》2019年第20期261-270,共10页Laser & Optoelectronics Progress

基　　金：山东省自然科学基金（ZR2019MF050）;"十二五"国家科技支撑计划（2014BAG03B05)

摘　　要：针对基于卷积神经网络的超分辨技术在纹理恢复时存在信息丢失、边缘模糊的现象,将密集块和激励模块结合,对图像进行低分辨率到高分辨率的端对端处理。密集连接融合后组成的密集块结构使图像区域的上下文信息得到有效利用。激励模块将有价值的全局信息选择性放大并将无用特征加以抑制。图像重建部分中的多个1×1卷积层结构减小了前一层的尺寸,在加速计算的同时减少了信息丢失。直接处理的原始图像缩短了训练时间,卷积层和滤波器的优化显著降低了计算复杂度。Aiming at the loss of information and edge blurring during texture recovery using super-resolution technology based on convolution neural networks,we combine dense block and squeeze module to learn the mapping from low-resolution to high-resolution in an end-to-end manner.The dense block structure formed by the fusion of dense connection utilizes context information of image region effectively.The squeeze module amplifies valuable global information selectively and suppresses the useless features.The multiple 1×1convolution layer structures in the image reconstruction section reduce the dimension of the previous layers,and speed up the calculation while reducing the loss of information.Processing the original image directly shortens the training time,and the optimization of convolution layers and filters reduces the computational complexity significantly.

关 键 词：图像处理 超分辨技术 密集连接 激励模块 1×1卷积层 计算复杂度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]