Research And Implementation Of Gray Image Colorization Algorithm Based On Deep Neural Networks

Grayscale image colorization is a research hotspot in the computer vision field.Its goal is to assign a reasonable color to each pixel in the grayscale image so as to make the colored image has richer visual information.Currently,gray image colorization technology has a broad application prospect in many fields,such as machine vision,medical image,video making and so on.However,owing to many colors share the same gray value,image colorization is an ambiguous problem.Under the influence of this uncertainty,gray image colorization is still considered as a challenging and significant task.With the improvement of computer image processing ability,deep learning has also been rapidly developed.In the field of image colorization,because of generative adversarial network hold unique training method of confrontation,so that it has certain advantages in the quality and color saturation of the generated image.However,it still exist some problems such as color bleeding and lack of details.Therefore,on the basis of studying large number of advanced colorization algorithms at home and abroad,this thesis proposes two kinds of automatic colorization algorithms based on generative adversarial structure.The main research contents and innovation points are as follows:(1)Proposed an adversarial colorization algorithm based on global feature adaptive optimization.The algorithm uses the GAN as the main structure,and adopts paired data for learning color information under strict supervision mode.For the problem of color bleeding,a self-designed global feature fusion module is introduced into the generator network,and the global feature and local hierarchical features of different scales are adaptively fused to improve the model's ability for obtaining global semantic information,but this module will limit the input image size.For the problem of insufficient detail information,from the perspective of the generative model,a channel attention module is added to the generator network to weight the features to enhance the feature learning ability of the model.Because of the GAN exists gradient disappearance and mode collapse during training process,the WGAN-GP optimization idea is combined in the loss function to enhance the stability of training.Experimental results show that the algorithm can alleviate the problems of color bleeding and insufficient detail information.(2)Proposed an adversarial algorithm based on cycle consistency.The algorithm is improved on Cycle GAN which is suitable for unsupervised learning,It can realize colorization in an unsupervised mode with a small amount of unpaired data.First of all,for the phenomenon of color bleeding,a non-local module with no size limitation is used in the generator network to establish a long-distance dependency to obtain a representation of global information.Secondly,for the problem of insufficient detail information,from the perspective of the discriminator model,a multi-scale discriminator is used to imporve the details.The WGAN-GP objective function is also used in the loss function to enhance the stability of training.The experiment proved that the algorithm can generate colored images with strong sense of reality and complete details.(3)Compare the above two algorithms,analyze the training method,model structure,training time cost and the time required to color a single 256×256 image,then finally summarize the advantages and disadvantages of the two algorithms in this thesis.