一种恢复图像自然色彩的重构方法  

Reconstruction method for restoring the natural color of an image

作　　者：陶霖密[1] 袁春[1] 王浩达 TAO Linmi;YUAN Chun;WANG Haoda(Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学计算机科学与技术系

出　　处：《智能系统学报》2019年第5期877-881,共5页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61672017)

摘　　要：现代数码相机是通过颜色过滤矩阵在每个像素位置采集一个颜色分量,重构出全彩色数字图像。压缩感知理论证明了该重构是误差有界的,但在实际应用时却隐含着一个问题:重构图像所需的稀疏编码字典是从图像数据库学习出来的,而目前数字图像都是重构出来的,因此存在着从重构的图像学习字典去重构图像的循环悖论。针对这个问题,提出并构建了新的完全采样彩色图像的Sandwich图像数据集,打破了压缩感知理论在应用于图像重构时的循环悖论,使得压缩感知方法能够真正地重建自然彩色图像。Sandwich图像数据集的构建及其训练得到的字典可以应用于如图像超分辨率重构、去噪、修复等领域。深入的图像重建实验表明,使用sandwich图像集训练的字典不论是字典原子特性还是由其重构得到的图像质量均好于基于传统数据集的结果。Currently,the digital camera captures one color component at each pixel location through a color filter array and reconstructs a full-color digital image.The compressed sensing theory has proven that this reconstruction is based on the error bound with sparse coding dictionary.However,in practice,there is a hidden problem:the sparse coding dictionary needed for image reconstruction is learned from the image database while the current digital images are reconstructed.Thus,there is a cyclic paradox of reconstructing images from the reconstructed image learning dictionary.To solve this problem,this study proposes and constructs a fully sampled color image dataset called Sandwich to break the cyclic paradox of compressed sensing theory in the application of image reconstruction,enabling the compressed sensing method to reconstruct truly natural color images.The dictionary trained from the constructed Sandwich dataset can be applied to several domains,such as image super-resolution reconstruction,denoising,and restoration.The in-depth image reconstruction experiments show that the dictionary trained using the Sandwich image dataset is better than the dictionary trained using the traditional dataset in terms of both dictionary atomic characteristics and the reconstructed image quality.

关 键 词：颜色过滤矩阵 压缩感知 稀疏编码 字典学习 图像重构 自然色彩 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]