Research On Quality Assessment And Content Generation For Two-Dimensional Cartoons

Cartoon has been popular for its concise,vivid and special artistic expression.From a global perspective,cartoon animation industry has become a huge and growing emerging industry.However,the production of cartoon animation is a very tedious process and a cartoon animation often contains thousands of in-between frames,thus drawing and coloring them are time-consuming and labor-intensive.Moreover,early cartoons contain scratches,damaged aeras,poor color,low resolution,compression artifacts,noises and other problems,hence they cannot meet the visual experience requirements of current users.In recent years,methods in the fields of computer vision and image processing have developed rapidly,and especially with the booming of deep learning,the research of related tasks such as generation,enhancement and quality evaluation of natural images has made breakthrough progress.Unfortunately,because characteristics of cartoon images are not sufficiently considered,the effect is often not ideal to directly apply these methods focusing on natural images to cartoon processing.Hence,this thesis focuses on cartoon contents generation,including how to promote cartoon animation production by artificial intelligence technology and high-definition reproduction of existing black-andwhite old cartoons.However,in the practical research of this thesis,it can be found that due to the large differences between cartoon images and natural images,the quality evaluation algorithms of natural images often cannot accurately measure the quality of cartoon images,and even give higher scores to low-quality cartoon images,which also restricts the research of cartoon enhancement and generation technology.In order to provide a feasible evaluation metric for subsequent generation task and related research on cartoon image processing,this thesis first carries out research on blind image quality assessment for cartoon images.Hence,this thesis conducts research and discussions related to the image quality assessment and content generation for two-dimensional cartoons,and the main work is as follows:1)Blind image quality assessment method is proposed for cartoon images.Aiming at the problem that the blind image quality assessment algorithms for natural images are often ineffective for cartoon images,this thesis analyses the statistical characteristics of cartoon images,a blind cartoon image quality evaluation method is designed based on edge and texture statistical priors.A cartoon image is first divided into edge regions and non-edge regions,and then the quality of edge regions and non-edge regions are measured by edge sharpness model and local texture complexity model,respectively.2)Blind image quality assessment method based on convolutional neural network for cartoon images is proposed.To improve the robustness of evaluation method,a largescale cartoon image quality assessment dataset is established,and a full-reference image quality assessment method is proposed to generate pseudo-labels for cartoon images in training set.On this basis,a blind cartoon image quality evaluation network is proposed and trained.Experimental results on both artificial degradation and real-world cartoon image datasets have shown good performance of the proposed method.3)High-definition reproduction framework is proposed for early black-and-white cartoon videos.Aiming at multiple distortions of early black-and-white cartoons,the recursive alignment network-based method is designed to improve the visual quality of early cartoons,and later a semi-interactive network is proposed to color key frames,and then the remaining intermediate frames are colorized by automatic video colorization method based on key frame reference.Experimental results on both synthetic black-andwhite cartoon dataset and real-world early black-and-white cartoon dataset show that the proposed method can realize the high-definition restoration of the early black-and-white cartoons.4)Explore a method for cartoon video automatic generation.This thesis focuses on the tedious and time-consuming ‘inbetweens’ and ‘colorization’ in cartoon animation creation process and proposes a three-stage cartoon generation network by imitating the cartoon production process,including sketch color inferring,intermediate frames prediction and animation sequence generation,which is aimed at improving the efficiency of cartoon animation production,so as to ultimately promote the development of animation industry.Experimental results show that the proposed method can initially realize the automatic generation of simple animation sequences from a small number of hand-drawn sketches.