Research On Image Colorization Method Based On Deep Color Transfer

Image colorization technology is a common image processing technology that is widely used in production and research in different fields.In field of medicine,most medical images are rendered as grayscale images due to imaging principles.The technology of medical image colorization can highlight tissues and organs,making it easier for doctors and patients to observe the physiological information.Traditional colorization algorithms of medical images can lose the image content and affect the doctor's diagnosis,and the tissue information of color medical images cannot be clearly displayed without combining the abstract deep features of the image.Therefore,this thesis proposes an algorithm based on deep neural networks and color transfer to achieve preservation of image content,and highlight or restore tissue information.In the field of natural image processing,the scenes of natural image are rich,and it is difficult to pick out reference images with the same content object.By selecting multiple reference images to relax the selection conditions,the flexibility and adaptability of the algorithm can be improved.Therefore,this thesis proposes a natural image colorization algorithm based on multiple reference images.The research work of image colorization algorithms in this thesis is as follows:1.Considering the problems that traditional medical image colorization algorithms only use hand-craft features without deep features,and the colored results can lose content information,this algorithm combines neural networks with color transfer to incorporate deep features into color images to fully display the information of image.During the coloring process,the content of the color image is kept intact through the losses of content,which ensure that the color medical image can still be used for medical research and diagnosis.The feasibility and effectiveness of the algorithm are verified through subjective and objective evaluations by experiments on multi-modal data sets.2.Considering the difficulty of reference image selection,this thesis adopts multi-reference technology to relax the selection conditions of reference images.First,multiple reference images are selected,and the target and the reference images are input to a semantic segmentation network to obtain semantic segmentation masks,and small area labels that affect the colorization result are trimmed.Then the word embedding model is used for semantic matching to obtain instance objects of reference images that match instance objects of the target image.Finally,the matching results are input into the improved colorization algorithm to generate the colored image and post-processing.A large number of experimental results verify the effectiveness of the multi-reference technology.3.Design and implement simulation systems for color perception of medical and natural images.Users can choose different systems according to different image types.Both systems are designed to can adjust the hyperparameters and select the reference images,which increase the flexibility of the system.The user inputs the grayscale medical or natural image into the system,and adjusts the parameters of system and runs,and the system outputs color images on the interface.