Research On Key Problems Of SVBRDF Material Efficient Modeling Based On CNN

In the field of computer graphics,the pursuit of the trueness of rendering results is a core and traditional topic,and this is inseparable from the accurate depiction of light sources,material surfaces,and interactions between the two above,so how to model the surface appearance efficiently and precisely has become the goal pursued by related researchers.This article mainly introduces an efficient surfaces material modeling method for the surface of an object.It requires only a single photograph of a planar material sample under unknown natural lighting conditions and a convolutional neural network that has been studied and trained to model physical and reasonable conditions.The SVBRDF results in the apparent diffuse reflection,normal,high-light reflectance,and roughness components.First of all,we must establish a rendering model suitable for these four kinds of markers in order to train the establishment of data sets.Second,we must have a large amount of texture material attribute data,and collecting markups that consist of the corresponding reflection-related parameters of the SVBRDF sample's photos is very difficult for the dataset.Therefore,in order to reduce the amount of marker training data required,this paper has adopted a self-enhanced training strategy:By using a set of labeled training pairs to obtain an initial network,as a start,next send another set of unlabeled data to the network to predict,and use the prediction,i.e.,parameters re-render a corresponding set of images in a new light environment and form a new set of labeled training pairs with this set of prediction parameters.After refining the network with this new set of labeled training pairs,re-predict the unlabeled data,re-estimate the material parameters,and this self-augmented process is repeated until convergence.This article describes the proposed network of three kinds of materials in wood,metal,plastic practice,and the effect is good.