Research On Colorization Algorithm Of Black And White Image Based On Deep Neural Network

In recent years,with the appearance of deep convolutional neural networks,the colorization of grayscale images has become a research hotspot once again.The goal of grayscale image coloring is to assign color to each pixel of the grayscale image,which has high research and application value.There are two main methods of image colorization,they are image coloring based on color transfer and image coloring based on color extension.Those two methods require the interaction between human and computer,and the operation is complicated and difficult in the process of processing images.what's more,the versatility is not strong,and cannot be used to complete all the coloring tasks with the same image coloring algorithm,and it is difficult to promote in a large-scale.With the improvement of processing ability on graphic images,deep learning has been applied to many fields and can be perfectly combined with digital image processing.So it has important practical significance and value to make research on black and white image colorization algorithm based on deep learning.In view of the current shortcomings of this method,we propose two automatic coloring algorithms in our paper.The specific research contents are as follows:(1)At the beginning of this paper,two traditional coloring algorithms are introduced,which are image colorization algorithm based on local color expansion and image colorization algorithm based on color transfer.Then we focus on the history,the advantages and disadvantages in the application of two traditional colorization algorithms(2)After analyzing the shortcomings of the two main traditional colorization algorithms,we propose a new image colorization algorithm based on convolutional neural network,and applies U-Net network in this algorithm.Our algorithm uses the SE-Inception-ResNet-v2 network to extract high-level features of the image,and replaces the linear rectification function(ReLU)with Power Linear Units(PoLU)functions.The experimental results show that our proposed algorithm can effectively colorize the grayscale image.(3)We propose a new colorization algorithm based on Self-Attention Generative Adversarial Networks.In this algorithm,the self-attention network is used to model the longdistance dependence relationship of the image generation task,and use the clues from all feature locations to generate image details.In addition,a spectral normalization method is added to improve the image quality that generated by the generator in the network,the discriminator in the improved algorithm can check whether the detail features in the far part of the image are Consistent.Experiments show that the proposed algorithm can effectively colorize black and white photos in batches.