Deep Learning Based Interactive Image Colorization

Recently,deep learning has shown strong problem-solving capabilities in the field of image processing,such as gray-scale coloring,image enhancement,image classification,image detection,image completion,and style conversion for image.Because of the higher information richness and greater value,gray-scale coloring has always been a hot research spot in the image field.The current colorization methods are mainly divided into user-guide colorization,example-based semi-automatic coloring,and fully automatic coloring.However,traditional algorithm coloring has problems such as inefficiency and unsatisfactory coloring effect.Moreover,the existing coloring interactive methods have difficulty in operation mode,difficulty in controlling coloring results,difficulty in finding a suitable reference picture,and so on,which are difficult to put into practical applications to bring practical significance to users.Therefore,with the continuous improvement of human requirements for user experience and the influence and role of deep learning in image processing,it is very practical to do researches on image coloring algorithms based on deep learning.In view of the current development of deep learning,existing theories and research results,and the advantages and disadvantages and limitations of the existing coloring methods,this paper proposes a novel interactive image colorization method based on deep learning.The main work contents are:(1)Proposing a interactive image colorization method combining global input and local input,and construct a network model that can train two kinds of shading inputs at the same time,and a suitable loss function is designed to constrain two user inputs.Implements a coloring interaction method that can use two user inputs separately or simultaneously,and ensures high quality of coloring results by inputting related auxiliary information.At the same time,the rationality and effectiveness of the algorithm are verified by experiments.(2)Using the idea of residual learning,adds a residual network module to the colorization model,and the output of the original model is corrected by learning a residual map,which further improves the coloring effect.(3)Embedded the deep learning coloring algorithm into a coloring system prototype.On this prototype,the visual operation of global input coloring and localinput coloring can be performed.Through the prototype,some user researches on the coloring mode and coloring result are carried out.Combined with theory and practice,the effectiveness and practical application of the coloring method in this paper are verified.(4)Use a wide range of image datasets,including object images such as butterflies,birds,and landscape images such as mountains,buildings,and pre-processing input information so that network models can perform well on most categories of images Coloring.