BTF Material Expression And Rendering Method Based On Neural Network

In the real world,many materials are not flat,and there are complex physical phenomena such as self-shadowing and self-reflection at the mesoscale.The warm texture of knitted sweaters and the texture of felt fleece are the effects of light on the complex physical process of the surface at the mesoscale.In the application field of graphics,how to render realistic scenes enough to deceive the naked eye has always been the focus of research.Photorealistic rendering is currently the goal pursued by many industries,such as large-scale games,special effects for film and television dramas,and advertising.To achieve realistic rendering,there are usually two types of methods:One is the mathematical modeling of complex light propagation by computer.In order to achieve precise modeling,complex models and huge computing overhead are required,and the hardware requirements of computing equipment are high;another type of method directly measures and collects the reflectivity list of different materials under different light source perspectives from the real world for rendering,and does not have to consider the complex light propagation process.The classic method has a data-driven bidirectional reflection distribution function(BRDF)model and data-driven bidirectional texture function(BTF)model,the measurement data has a high degree of realism,but this method requires GB-level storage and is difficult to apply in actual production.Therefore,the dimensionality reduction technology of BTF material has become an important topic.The innovations of this thesis are as follows:1.When using discrete BTF material data for photorealistic rendering,it is necessary to query the light source direction,viewing angle direction and position direction,resulting in a large number of interpolation requirements.In addition,during the rendering process,the data of most of the items in the measured BTF material will not be accessed,but it is required to be read into the memory during rendering,resulting in a lot of space waste.Therefore,a BTF compression method that can reduce both storage and interpolation needs to be proposed.The work in this thesis is inspired by work on neural BTF compression and interpolation.These architectures use encoders to convert six-dimensional BTF textures into potentially low-dimensional manifolds.In these works,the ABRDFs at different positions of each BTF data are processed separately,and the BTF materials are similar in position,so there is still a lot of redundant information in the position direction of the compressed data.In contrast,our method based on Resnet encodes the BTF position,viewing direction,and lighting direction information together into the underlying vector space,enabling smooth interpolation and extrapolation during rendering,simplifying the rendering process.The method in this thesis can recover the real details of the measured material,improve the compression ratio of the BTF material by 134%,and shorten the rendering time by 12%.2.The collection of BTF materials requires a lot of time and storage costs.The editable compression method can generate a variety of realistic materials in the limited BTF material library,which is helpful for practitioners in secondary creation.Under the premise of high-quality compression,maintaining the editability of BTF materials is also a research difficulty.In this regard,this thesis also proposes a SVBRDF generation network based on measuring BTF materials.This method maps ABRDF to a set of SVBRDF parameter vectors through an encoder based on residual network,and the SVBRDF parameter vectors at each position are concatenated to form SVBRDF.Diffuse reflection,normal,specular reflection and roughness four maps,this method discusses the possibility of mapping BTF materials to SVBRDF maps through neural networks,and has achieved certain progress.