Color Method Of Animation Line Drawing Based On Gated Residual And Semantic Transfer

In anime creation,coloring anime line drawings is usually one of the most timeconsuming aspects.In the traditional grayscale image coloring method,the grayscale image ontain features in shades of gray from black to white,and the regions with the same grayscale information are divided into several regions to obtain a better coloring effect by using the luminance information.However,due to the lack of information of anime line drawings,only contour features exist without grayscale region information,so the coloring method of grayscale images cannot be directly applied to black and white line drawings images.Against the background of researchers’ increasingly mature research on deep learning technology,automatic coloring technology for anime line drawing images brings new opportunities for fast and high-quality production of anime.Most of the existing coloring methods have problems such as coloring boundary penetration,color confusion,and unrealistic color lines.In this thesis,we propose an automatic coloring method for anime line drawings based on generative adversarial networks.It can color the manga line drawing image according to the reference image.In order to alleviate the problems of color confusion and unreasonable color lines,a dualconditional generative adversarial network is designed in this thesis.The dual conditions are the structure condition and the color condition,respectively,which prompt the coloring of the generated image to be more in line with the color style of the reference image,and the contour of its generated image is similar to that of the anime line drawing image.Because the network has the problem of feature loss during the training process,the semantic migration module is proposed in this thesis.It can reduce the loss of color information and generate colorful and clear textured color images.And the gated residual module is proposed in this thesis,which can better solve the problem of network gradient disappearance and deep network degradation compared with the traditional residual module.In order to make the effect of coloring the anime line drawing better,a series of optimization of the network is also carried out.The experimental results show that the method has better coloring effect compared with the existing methods.For the difficulty of acquiring input image-real image pairs,this thesis proposes an automatic coloring method for anime line drawings based on augmented self-reference.In this thesis,we adopt the learning method of enhanced self-referencing,in which the reference image is derived from the original color image through color transformation and spatial transformation.Not only the training process can be conducted in pairs,but also the size of the training set is expanded to a certain extent to improve the generalization ability of the model.In order to better guide the generator to learn the color style of the reference image,this thesis proposes the Progressive Style Migration Network(PSTN),which can conveniently embed the color into the reconstruction of the final image.The experimental results demonstrate the effectiveness of the method.