The Research On Colorization Algorithm Of Image Based On Conditional Generative Adversarial Networks With Three Channels

In actual learning and scientific research,in many cases,it is necessary to render grayscale images into color images,such as coloring black and white photos,colorizing night vision imaging effects,and so on.The more popular gray-scale image coloring method is that professional technicians use professional image processing software such as Photoshop to partition the picture,and then adjust the hue and color saturation of each partition to complete the coloring.As we can see,the entire process is quite complicated.For non-professionals,there are certain difficulties for personnel.In the field of computer vision,there are also traditional digital image processing techniques such as color transfer,color marking and other methods to achieve gray-scale image coloring,but this type of technology often requires the processing personnel to have a wealth of computer vision expertise and processing and the speed of this kind of method is slow.With the rapid growth of computer vision and deep learning technology,deep learning technology is widely used in image processing technology,including gray-scale image coloring.The current mainstream gray-scale image coloring technology based on deep learning mostly uses deep convolutional neural networks(such as Res Net,VGG,etc.)to extract image features,color fill different regions of the image according to the extracted features,and when training the model generally,multi-channel color images are regarded as a whole,that is,converted into a one-dimensional tensor input neural network model for training.The characteristic of this type of model is that the depth of the network is deep,it is easy to overfit,and a color image is taken as a whole to consider the lossy coloring effect.Considering that the color of each pixel of a multi-channel color image is determined by the gray values corresponding to all single-channel images,and the same color image can be described by different color spaces.Therefore,from the perspective of the multi-channel nature of color images and the superiority of the effect of generating adversarial networks,I propose a hybrid three-channel generative adversarial network model for coloring grayscale images.The main work of this article is as follows:(1)A mixed three-channel integrated model based on RGB color model and Lab color model is designed,which is a single model in color mode can restore more realistic colors.(2)Adopt a generative adversarial network instead of the traditional deep convolutional neural network as the basic network of the model,and obtain a more excellent model through game training.At the same time,a hollow convolution is added to the generator to assist the general convolution to obtain multi-scale images features,can restore the details of the image color.(3)Continue to make progress for the model designed in(1),use the semantic segmentation graph of the original image to constrain the neural network model,construct a conditional generation adversarial network based on the semantic segmentation graph,improve the convergence speed of the network model,and reduce the occurrence of probability of gradient dispersion.