Research On Grayscale Image Colorization Method Based On Deep Learning

Color is of great significance to human beings.It is the core of visual aesthetics,as well as the embodiment of the physical and physiological attributes of the natural world,and it has a profound impact on our life and culture.Colorization is a shading procedure in the image or video,which is done to give detail and lucidity to the image or video.The color image is more conducive to the extraction and expression of image features.Image colorization is relevant in numerous regions,for example,colorization of old highly contrasting photographs,old films,cultural relic restoration,medical image processing and space exploration.Colorization techniques can be generally separated into two classes: user-guided colorization and data-driven colorization.Among the different color methods,data-driven color method has become the research target of current colorization technology because of its ability of self-learning from image data sets.This paper mainly studies the deep learning-based gray image colorization algorithm,including the following three aspects:Firstly,a deep convolutional neural network based image colorization is presented.On the basis of analyzing the deficiency of traditional non-deep learning methods in image colorization,the possibility of color colorization method based on deep learning is explored,and a "U" shaped color neural network is realized.Second,a generative adversarial network with dual feature extractor for image colorization is presented.To solve the problems of context confusion and detail information loss in color images,the U-Net-like network was used as the trunk network in the generator.The encoder was used to extract the local features of a grayscale image.The branch extractor used the Res Ne Xt network that was added to the SE module,as a high-level feature extractor to extract the global features of the grayscale image.It improves the problem of color leakage and detail information loss in the process of color coloring,and the addition of adversarial loss also makes the tone tend to "average" to be improved.Finally,a new method of image colorization based on a generative adversarial network combined with the semantic segmentation task is proposed.For color image is not clear in the edge of an object,on the choice of color difference loss is limited,this paper use a deep semantic integration of the structure of the generator to infer a given color of gray image under the condition of semantic clue,and predict the color information and semantic label information to learn gray image to color image mapping relationship.A color difference loss with human visual observation characteristics is used to calculate the loss,and the model is trained by combining loss,so as to improve the overall effect of colorization.