Research On Image Colorization Technology Based On Deep Learning

Color plays an important role in many computer vision problems.Compared with gray images,colored images can provide additional color information for those problems,such as,image classification,object detection.Image colorization is a challenging task of adding colors to a given grayscale image and is widely used in academic and engineering fields.At the same time,deep learning has the characteristic of learning features of images automatically based on the development of computer,which can integrate features into the model effectively and reduce the intervention of human.Therefore,there is an inevitable trend of solving the image colorization problem by deep learning.According to the basic mapping relationship of image colorization and the knowledge of deep learning,a two-stage end-to-end image colorization framework based on deep learning is proposed in this paper,for mapping a given gray image to a colored image with three dimensions.The first stage of this framework is giving the basic color to the gray image by an initial colorization network.The second stage is correcting the image generated by the initial colorization network by the color palette,which forces the final result nears to ground truth.Therefore,the main contribution of this paper are followed:First,a two-stage gray image colorization network framework is proposed.In this framework,a color palette generation model based on texture features is designed.Since the model construction of generating colorized images directly is usually limited by the color diversity of images in the dataset,this paper generates the palette with five different colors for the given gray image according to the LBP texture pattern of it.In the first stage of the proposed framework,a conditional generative adversarial network is designed and the predicted color palette is be merged into the gray image in this network to obtain the initial colorization result.In the second stage,the palette correct model based on generative adversarial network is proposed.The predicted color palette strengthens the color of the initial result by this model to make the colorized result better.Secondly,as the color cast problem may occur when using the generative adversarial network for image colorization,a strong correlative loss function based on color cast and signal estimation is designed.This loss function can constrain the training direction of the colorization model and force the generated image be closer to the ground truth.The experiment results verify the effectiveness of the proposed method.The Experiments show that the image colorization algorithm based on deep learning proposed in this paper has certain application value.It can colorize large quantities of images at one time and has certain effectiveness and practical significance.