Research Of Key Colorization Technology For Image's Multi-Regional Based On Deep Learning

Recently,in each area computer technology is increasingly developed especially in image.Image Colorization is a work that take time and effort.As for an image,color scheme is key factor to impact the quality of image.Therefore,image auto colorization is a task has research significance and application value.Deep learning can work well in image auto colorization with the development of computer hardware.The methods of image colorization can divide into data-based,reference-based,user hints-hased methods according to the source of color informationAlthough the above methods achieve great results,they still have some limitations For example,user cannot change one target color only and keep other object color remain unchanged in one image.It is not conducive to create secondarily.To solve this problem,this paper proposes a multi-region hybrid colorization based on instance segmentation and poisson editing.The method of this paper is divided into data-based background colorization and reference-based foreground colorization.As follows are the main work and contributions of this paper1?In order to identify the multiple targets and background in the image and the merge the result,this paper study Mask R-CNN and poisson image editing.For foreground colorization,given a gray image and a color reference image.This paper transfers the color of the specified area of reference image to the specified area of the gray image using Mask R-CNN to extracts the respective segmentation.During the transfer,image and its segmentation as the input of VGG.The output of VGG is gray content feature and reference style feature.And then train a noise close to gray content feature and reference style feature respectively.Moreover,adding a photorealism regularization during training to keep result real.Meanwhile,we color the background using U-net.Finally using poisson editing to merge the background and foreground2?This paper has designed a method to automatically recommended reference image that similar to gray image.The image recommended extract image feature using VGG at first.And then calculate the cosine of the feature as semantic distance.Finally use the hash to calculate the similarity between the image and select an image from databased as reference.3?Color transfer based on VGG feature extraction will take times during iteration.To overcame this shortcoming,this paper proposes a new color transfer method based on image transfer network.The new method can enhance the speed of colorization by adding image transfer network before VGG.Given a reference image and train the transfer network on COCO datasets to make transfer network save reference image information.