基于并行映射卷积网络的超分辨率重建算法  

Super-resolution algorithm based on parallel mapping convolution network

作　　者：毕笃彦[1] 王世平 刘坤 何林远[1] BI Duyan;WANG Shiping;LIU Kun;HE Linyuan(Aeronautics and Aeronautics Engineering college,Air Force Engineering University,Xi’an 710038,China)

机构地区：[1]空军工程大学航空工程学院,陕西西安710038

出　　处：《系统工程与电子技术》2018年第8期1873-1880,共8页Systems Engineering and Electronics

基　　金：国家自然科学基金(61372167;61701524)资助课题

摘　　要：针对基于卷积网络的超分辨率重建算法对不同场景下的图像存在复原质量不佳、细节信息丢失的问题,对卷积网络结构详细分析,结合重建模块和损失函数约束条件存在的问题,提出了基于并行映射卷积网络的超分辨率重建模型。该模型基于端到端的思想,构建并行映射网络及正则化约束条件,能对图像特征进行层次化自主提取,在高分辨率图像重建时极大地丰富图像特征的维数;并且将全变分正则化引入到重建模块,有效地克服了超分辨率的病态问题,从而获得鲁棒、丰富的图像信息,提升了重建图像的质量。实验结果表明,所提出的网络模型具有更优异的性能,其超分辨率算法在视觉评价和量化指标上取得了更好的重建效果。The traditional superresolution algorithm based on convolutional network is difficult to recover highresolution images and fuse edge information in different scenes. In order to solve this problem and based on the detailed analysis of the network of the typical models, the parallel mapping convolutional network proposed model relies on the analysis of the problem of reconstruction module inputs and loss function constraints. The model based on endtoend manner, constructing parallel mapping convolutional network and regularization constraints, can be hierarchically independent of image features extracted and greatly enrich image in polymerization of high resolution image feature dimension. Meanwhile, the network introduces the total variation regularization after the convolution layer and constraints illposed problem, which extract accurate and robust image features from the network, and enrich the edge information of image. The experimental results on typical databases show that the proposed algorithm achieves better superresolution results, the subjective visual effect and objective evaluation indices are improved significantly, and the resolution of the image is enhanced.

关 键 词：图像复原 超分辨率重建 并行映射卷积网络 全变分正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]