Research On DDPG Reinforcement Learning Algorithm Based Painting Simulation

As a common form of art,painting is a crucial visual communication tool in the fields such as journalism,prototyping,and film creation.Digital painting is easily disseminated in the Internet era and has significant research and applications value at present.Currently,there are several fields of research related to the synthesis or simulation of digital painting: neural style transfer and image translation are based on the latest deep learning technology to convert images between different styles or domains,and the painting content can be generated when the target style or domain is painting.because it uses neural networks to calculate pixel values directly,these methods cannot simulate the painting process and are poor in brushstroke details.Stroke Based Rendering(SBR)uses predefined strokes to synthesize paintings.Some of the research focuses on the design of stroke renderers to get a better result on stroke rendering,another line of research obtaining stroke control parameters by a greedy search algorithm or optimization methods to generate final painting,these methods can simulate the painting process and have better effect in brushstroke detail,however,the optimization process of these methods is slow.Existing methods based on reinforcement learning model the painting process as a Markovian decision process,and then generate stroke control parameters to synthesize paintings through a painting agent,which is fast and can achieve a human-like painting process simulation.However,these methods still have the problems of poor stroke rendering results and lack of stroke diversity.In this paper,we use the model-based DDPG(Deep Deterministic Policy Gradient)reinforcement learning algorithm as the main framework and combine it with the dual-path neural renderer to improve the rendering result of various brush strokes.Specifically,the work in this paper is summarized in the following three points.(1)This paper first defines various stroke types such as ink,watercolor,and oil paint,based on existing open-source library libmypaint,then integrates different strokes to build a unified renderer framework,and uses this framework to train a dual-path neural renderer to imitate the real renderers.ablation experiments prove the effectiveness of the training strategy proposed in this paper,and the trained neural renderer can get a similar stroke rendering result as the real one.(2)In this paper,the dual-path neural renderer is used to construct a painting simulation environment combined with the model-based DDPG algorithm framework to train painting agents for various strokes,to improve the quality and diversity of strokes.The comparative experiment in the quantitative and qualitative show that the painting results of this paper outperform existing methods.(3)This paper proposes a learnable stroke translator,which converts the control parameters of different strokes through neural networks.The translator can be used as an extension of the paint agent to convert the stroke control parameters it generated,to achieve transfer of painting with different stroke styles.The experiment results show that the painting content of various strokes can be obtained quickly using the translator,to saving training time and computational resources.