Neural Network Based Reconstruction Of Realistic Images Of Virtual Scene

Image-based modeling and rendering(IBMR) techniques render novel images directly from input images. Comparing with traditional Graphics methods IBMR techniques generate images quickly and realistically. But there are some difficult issues need to be studied further, including how to recontruct dynamic scene images and images with specular reflection, and how to change reflectance properties of object in images.No effective IBMR method has been presented to reconstruct dynamic scene images under changing lighting conditions till now. However, by studying the relationship of image color and the lighting conditions, a method of generating realistic images at hypothesizing lighting conditions is presented in the paper. The inherent relations of color and outside lighting conditions are discovered by the use of neural networks. This method doesn't need any space geometry informations and the arithmetic is simple. Using the techniques presented here, dynamic virtual images can be obtained. And if combining with other techniques, the method can be farther applied into real-time dynamic virtual reality.Reconstruction using IBMR can't accurately recover the illumination caused by the specular reflection. In this paper, by studying the relationship between positions of viewpoints and colors of images, a method is proposed to reconstruct novel images using neural network. The reconstruction method presented here is simple, and it can obtain novel viewpoint images with photo-realistic property, especially can accurately generate images with obvious specular reflection. In image reconstructing, the reflectance properties of object surfaces are often modulated in order to obtain content visual effects. In this paper an image-based rendering method is presented to separate specular and diffuse reflection elements in an image. Appointed new specular and diffuse reflectance ratio to the separated images, scene images with different reflectance properties can be obtained and novel images with different ratio of specular and diffuse reflection energy distribution can be reconstructed based on IBMR techniques . General regression neural networks(GRNNs) is adopted in the paper. Since the current neural networks techniques can't adapt to variational distance between input vectors and samples in regard of sparse samples problem, a new model based on GRNNs is presented, namely increment addition model. In the paper, realistic images under virtual lighting condition are generated using the model, which can also be applied in any situation of asymmetrical samples or sparse samples.Introducing neural networks, methods of reconstructing realistic images of virtual scene are presented. A new technique to solve problem, using neural networks, is exploited in IBMR. And effective methods are put forward to reconstruct images with obvious specular reflection, a simple arithmetic to reach dynamic virtual reality at changing lighting conditions, are provided. The model based on GRNNs presented in the paper is the enhancements to neural networks.