Research On Colorization Method Of Grayscale Image Based On Deep Learning

Color plays a key role in the process of human cognition of the world.For this reason,grayscale image colorization is a challenging and research significance subject in the field of digital image processing.Current most of grayscale image colorization algorithms include color markers,reference images,mapping relationships,and deep learning.Although the first two algorithms have better coloring effects,they are poorly versatile and require high manual difficulty because plenty of manual intervention.The third algorithm aims at the problem of pseudo-colorization.The main purpose of this algorithm is to facilitate the human eyes to distinguish details,which is not suitable for colorization in natural scenes.With the development of deep learning and the rapid upgrade of current large-scale computing equipment,more deep learning-based colorization methods have been proposed,and the coloring effect has been continuously improved.This thesis uses deep learning method to study gray image colorization.The details are as follows:1.By summarizing the existing colorization algorithms,this thesis analyzes the advantages and disadvantages of each algorithm.There are three main problems in current image shading algorithms: border blurring,loss of detail,and singled coloring effects.2.Aiming at the problems of boundary blurring and loss of details,this thesis proposes a method for colorizing grayscale images based on a deep layer aggregation structure network.The deep layer aggregation structure network is introduced into the field of image colorization.Based on the original network,a combination of long connection and short connection is adopted to further improve the information utilization rate of the network and reduce the problem of gradient disappearance in network training.Experiment proves that the method alleviates the problems of boundary blurring and loss of detail during colorization.3.Aiming at the problem of singled coloring effects,this thesis improves a method for coloring grayscale images based on generating adversarial networks and setting up a discriminative network.Through the establishment of a discriminative network,the true color of the generated color image is determined dynamically.The experiment proves that the method can effectively alleviate the problem of singled coloring effect.4.This thesis conducts detailed tests on the proposed deep learning-based gray image colorization method on multiple data sets.Meanwhile,the old black and white photos are tested for colorization.Experiment proves that the method in this thesis reduces the problem of border leakage during coloring,restores more image details,and has richer image colors than the traditional colorization method.