Invertible Color-to-gray With L0 Regularized Residuals

Image decolorization and image colorization are important technologies in the field of computer vision and image processing.Decolorization for colorful images has many applications in real life,such as monochrome printing,single-channel image processing,and stylization.Colorization for grayscale image is significant in history image repainting and stylization.But grayscale obtained by traditional decolorization method could not be recovered to the original colorful image precisely by state-of-the-art colorization method.To realize faithful decolorization and faithful colorization,invertible color-to-gray with L0 regularized residuals are proposed in this paper to solve this problem.This method combines decolorization processing with colorization processing in a generative network.Main works of this paper are illustrated as follow:At first,the thesis describes the significance and the applications of image decolorization and image colorization.Then some representative researches are introduced while their advantages and disadvantages are discussed as well.Next,this paper summarizes the applications of sparsity representation in the field of image processing and analyses some convolutional networks on how to extract and refine features.After that,we proposed invertible color-to-gray with L0 regularized residuals.The encoder network in algorithm encodes color information from colorful image to a residual image.Invertible grayscale image are obtained by the summation between residual image and grayscale with lossy compression.Decoder network in algorithm decodes color information from invertible grayscale image to recovered colorful image.Furthermore,we design L0 regularization to enhance its sparsity and suppress redundant values in pixels.In this way,the residual image can just occupy less storage but contributes to high recovered quality with less color hints.To improve the efficiency of above networks,we make some optimizations in networks and design dual features ensemble modules which make invertible grayscale same as traditional grayscale image while making recovered color image same as the original color image.Finally,sufficient experimental results show that,combined with L0 regularization,our dual features ensemble modules are advantageous to color recovery.In general,the main contributions of this thesis are: 1)This thesis analyzes the disadvantages of existing methods in image decolorization and image colorization and then proposes invertible color-to-gray with L0 regularized residuals.L0 regularization encourages pixels to be zero so that this method can make residual image sparsely in invertible grayscalewhich occupy less space but obtain accurate colorization;2)Dual features ensemble skillfully combines advantages of existing convolutional network architectures.Its strength on feature extraction improves color recovery results;3)As for proposed architectures above,more experiments,qualitatively and quantitatively,prove that the thesis' s method has faithful ability in image decolorization and image colorization.