Semantic Guidance Based Cycle-Consistent Adversarial Automatic Image Colorization Network

The purpose of the image coloring task is to convert each pixel of the grayscale image into suitable color pixels,so that the resulted image is perceptually meaningful and visually appealing.Image colorization can assist high-level tasks like visual understanding and object tracking.In our everyday life,no matter for rekindling dusty memories or for expressing artistic creativity,colorization technique plays important role.Since colorization has so many practical applications as a helpful technology on image editing,it has been an active topic in image processing and computer graphic research field.Image colorization is an ill-conditioned and inherently ambiguous problem.There are potentially many colors that can be assigned to a gray pixel.The multi-modality existing in solution space therefore makes image colorization a highly challenging task that has no unique correct solution.As an image colorization technology that can automatically color grayscale images without manual,automatic coloring has been an active branch in the field of image colorization.Unlike most existing methods,which focus on supervised learning and color space conversion techniques to restore the colors of the original image,this article considers that a reasonable solution is to generate some coloring results that look natural and should avoid paired training Data to avoid data collection inconvenience.In addition,this article believes that no matter what color area is to be assigned,the colored area should be consistent in semantics and space.Since the color value corresponding to the gray value is not unique,strict supervised learning may lead to unsaturated coloring.GAN-based generation scheme needs to be highly consistent in the semantic space.To this end,this paper proposes an unsupervised automatic coloring scheme ACCycleGAN based on unpaired samples.In particular,inspired by CycleGAN,our treats the shading process as image transformation,and uses the idea of cyclic consistency to train the model.During the training process,high-level identity loss modification and low-level grayscale loss are introduced into the optimization model.In this paper,a small amount of data randomly selected in PASCAL VOC 2007 is used as the training set,because the architecture of this paper does not require orders of magnitude data set to train.The results of its coloring experiments prove the effectiveness of the coloring scheme based on cyclic consistent adversarial in this paper.In order to ensure the consistency of colored regions in semantics and color feature space,our further proposes a generator guided by the semantic segmentation task to predict the color distribution,and optimizes the proposed gray loss from?₁ norm to?₂ norm.Our use PASCAL VOC 2012 to randomly train a small number of data sets to train further proposed models.To verify the improved performance of the model,we uses some popular advanced schemes for comparative experiments and explores the contribution of model components to performance.Experimental results show that the proposed scheme can generate convincing colors while maintaining the consistency of image content.