Research On Model Rendering Optimization Method Based On Machine Learning

With the rapid development of metaverse,game,film and television special effects,the requirements for rendering optimization are becoming higher and higher.How to ensure the quality of rendering while reducing the cost of hardware has brought great challenges to traditional rendering methods.In this case,rendering optimization using machine learning provides a valuable solution.Compared with the traditional complex and time-consuming rendering pipeline,the machine learning-based rendering method can effectively retain the rendering information,thus avoiding high hardware costs.On the other hand,machine learning-based rendering methods can often achieve good rendering effects while reducing rendering time.In this thesis,proposes a model rendering optimization method based on machine learning.The content involved is volume rendering method,which uses neural networks to learn the features of volume rendering scenes and render rendered images with near native quality.Most downstream processes do not need to complete the rendering process,and only need to obtain scene parameters from the method to render the final results,thus saving rendering costs and shortening processing time.In this thesis,the following research work is mainly carried out:(1)The principles and main methods of each part of the rendering pipeline are studied and analyzed.The traditional rendering method has the problems of slow rendering speed,low efficiency and high hardware cost.A model rendering optimization method based on machine learning is proposed to reduce the rendering cost while preserving the rendering quality as much as possible.(2)Volume based rendering optimization method: Combine volume rendering method with machine learning to improve on this basis.By using the method of neural network to extract the characteristics of the rendering scene and apply it to the volume rendering,the expected effect is achieved in the rendering effect.(3)Multi view rendering optimization method: The improved 3D-Unet network terminal was used for rendering optimization,and the insufficient side view optimization was improved by extracting the features of the reference rendering image under multiview.(4)The proposed method was analyzed experimentally and compared horizontally,and the expected experimental results were achieved on the rendering optimization results from several different perspectives,which verified the feasibility and effectiveness of the method.