深度光线跟踪技术研究

Ray tracing is a core technology in realistic rendering. It provides underlying support for various lighting effects to produce highly realistic images. However, due to the trace and intersection calculation of a large number rays involved, it suffers from slow speed that result from heavy load on computing. Though people have proposed many accelerating technologies in recent years, most of the improvement on speed is obtained by the development of hardware. The improvement brought by algorithms themselves is limited. With the development of application, it demands higher realistic lighting effects and higher rendering speed, resulting in an urgent need for introducing new ideas to break the development bottleneck. For this, in this project we tends to draw support from the method of deep learning in artificial intelligence and introduce the deep neural network into the ray tracing framework. By taking advantage of the ability of neural network to approximate non-linear function, we will establish direct mapping between rays and intersections and improve the efficiency of the whole ray tracing framework by completing the calculation of intersection quickly. We will select different neural networks (e.g. full connected neural network, recurrent deep neural network), embed them in different levels of the framework (local / global) by the feature of scenes and applications (static / dynamic) and organize a flexible deep ray tracing framework to push the development of realistic rendering in interactive or even real-time applications.

光线跟踪技术是真实感绘制中的核心技术，它可为各种光照效果提供底层框架支撑，产生高度真实的图像。然而，由于涉及大量光线的跟踪与求交计算，其计算负载大、速度慢。尽管近年来人们提出了多项加速技术，但多通过硬件的发展取得进展，由算法本身带来的速度提升有限。随着应用的发展，其对真实光照效果和绘制速度的要求越来越高，迫切需要引入新思路突破发展瓶颈。对此，本项目拟借助人工智能技术中深度学习方法，将深度神经网络引入光线跟踪框架，利用其非线性的函数近似功能，建立光线-交点的直接映射，通过快速完成光线路径上的交点求取操作提高整个光线跟踪框架的计算效率。我们将根据场景和应用特点（静态、动态），选择将不同的深度神经网络（如：全连接网络、循环深度网络等）嵌入到框架中的不同层次（全局、局部），组建一种灵活多变的深度光线跟踪构架，从而促进真实感绘制在交互乃至实时应用中的发展。

光线跟踪作为真实感绘制中的核心技术能生成很高质量的真实感图像，但其高昂的光线跟踪与求交计算开销阻碍了其应用，迫切需要引入新技术来突破发展瓶颈。在此背景下，我们将人工智能技术中深度学习方法引入光线跟踪框架，使用深度神经网络拟合光线求交函数，建立光线-交点间的直接映射，由此提高光线跟踪中交点求取速度。具体地，本课题对静态场景中基于深度神经网络的光线-交点映射拟合方法进行了研究。同时，我们对可嵌入深度神经网络的基础框架结构进行了研究，以通过局部化降低求交函数的拟合难度和提高拟合精度。它们包括网格划分结构、凸体划分结构、球面六边形层次网格结构等。多项技术取得了进展，提高了处理速度和质量。这些进展的取得能够很好的促进光线跟踪技术的效率提升，推动其在业界的更广泛应用。