Research And Application On Ambient Occlusion Algorithm Based On Convolutional Neural Network

Ambient Occlusion(AO)is a very critical component in the field of photorealistic rendering.The algorithm calculates and determines the occlusion of the object to the ambient light by acquiring the geometric information around the object in the virtual scene,which can render pretty real contact shadow effect of the world.In recent years,with the extensive application of artificial intelligence technology in many fields of computer science,some ideas of ambient light occlusion algorithms combined with neural networks have emerged.Neural Network Ambient Occlusion(NNAO)is the first algorithm that introduces neural network technology into real-time ambient occlusion calculation,and has achieved certain research results.It has more obvious performance than other real-time ambient occlusion algorithms,becoming a benchmark in the field of research.However,there are some problems exit in this kind of algorithms,such as excessive sampling and loss of high-frequency shadow details.These problems have not been solved in later research results,and become a bottleneck and constraint in the research of ambient occlusion algorithms based on Neural Network.In response to these research problems,we propose Accelerated Detail-enhanced Ambient Occlusion(ADAO)algorithm to achieve performance and detail enhancement based on Neural Network.This paper first analyzes and studies the algorithm structure and implementation ideas of the NNAO algorithm,and conducts experimental reproduction of the proposed neural network structure and the idea of using the training model to calculate AO.The results prove that it is feasible to use the Neural Network model instead of the empirical model to implement AO calculation,achieving higher rendering efficiency and accuracy.And then,in response to the problem that the relatively large number of samples in this algorithm leads to unnecessary calculation burden,this paper introduces Spiral Sampling and Level of Detail(LOD)techniques to propose a new type of spiral sampling based on visual depth algorithm.The sample point distribution of the Spiral Sampling algorithm has higher representativeness and accuracy than the original random sampling algorithm.According to the distance of the sample point from the human eye,we can further adjust the number of samples,reducing the overall number of samples in the scene.This article also uses the image enhancement technology in the field of graphics processing to enhance the contrast of AO calculation results,restore high-frequency details,and make the sense of shadow layering stronger.Finally,this paper proposes a new bilateral filtering algorithm,which adds the normal information of the object to the weight calculation of the filtering operation,and the shadow edge generated by the sudden change of the object normal can be well restored.Experimental results show that,compared to NNAO and other real-time ambient occlusion algorithms,the ADAO algorithm can improve the rendering efficiency while enhancing the realism of shadow details.