卷積神經網路降噪技術加速全域照明之探討

Denoising Path Tracing Renderings using Convolutional Neural Networks

杜宜家(Tu Yi-Chia)

指導教授 ： 張鈞法

國立臺灣師範大學/理學院/資訊工程學系/碩士(2019年)

https://doi.org/10.6345/NTNU201900615

摘要

近年GPU硬體技術進步，光線追蹤即時繪製有了開端，在複雜的場景繪製效能仍然有限，因此本論文將使用人工智慧輔助路徑追蹤，以卷積神經網路降噪技術代替部分的路徑追蹤計算，加速全域照明場景的產生。
蒙地卡羅方法高頻率取樣，會耗費相當高的時間成本在計算上，透過路徑追蹤低取樣頻率產生的影像，以人工智慧的方法去除蒙地卡羅方法產生的雜訊，提升影像品質。
論文中主要探討降噪技術，透過調整卷積神經網路結構，達到降噪效果，並保持一定程度的穩定性，與不同的場景變換之下廣泛的適用性，比較預測結果與實際場景影像的差異，討論即時降噪光線追蹤遇到的問題與未來趨勢。

關鍵字

路徑追蹤 ； 全域照明 ； OptiX光線追蹤引擎 ； 卷積神經網路降噪

並列摘要

In this paper, we use artificial intelligence to support path tracing. We replace part of the rendering calculations with image denoising which is implemented by convolutional neural network. This method effectively reduces rendering time to global illumination.
Monte Carlo method takes a lot of time to render a scene. While the number of samples increases, the noise decreases. In order to generate a high quality image and reduce sampling time, we use convolutional neural network to rebuild the image, which is based on low frequency sampling. The result is almost the same as Monte Carlo rendering with higher frequency sampling image.
Our primary focus is on the offline denoising technique. We use this technique to improve the stability and capability of the network. To process noisy images of different viewpoints, scenes and illumination, we adjust the network layers and training data. We compare a higher frequency sampling image with a low frequency sampling image whose noise is reduced. Eventually, we discuss about real time denoise rendering.

並列關鍵字

path tracing ； global illumination ； OptiX ray tracing engine ； Convolutional Neural Networks denoising

參考文獻

[1]Bako, S., Vogels, T., McWilliams, B., Meyer, M., Novák, J., Harvill, A., ... & Rousselle, F. (2017). Kernel-predicting convolutional networks for denoising Monte Carlo renderings. ACM Transactions on Graphics (TOG), 36(4), 97.

[2]Buades, A., Coll, B., & Morel, J. M. (2005, June). A non-local algorithm for image denoising. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) (Vol. 2, pp. 60-65). IEEE.

[3]Chaitanya, C. R. A., Kaplanyan, A. S., Schied, C., Salvi, M., Lefohn, A., Nowrouzezahrai, D., & Aila, T. (2017). Interactive reconstruction of Monte Carlo image sequences using a recurrent denoising autoencoder. ACM Transactions on Graphics (TOG), 36(4), 98.

[4]Kajiya, J. T. (1986, August). The rendering equation. In ACM SIGGRAPH computer graphics (Vol. 20, No. 4, pp. 143-150). ACM.

[5]Li, T. M., Wu, Y. T., & Chuang, Y. Y. (2012). SURE-based optimization for adaptive sampling and reconstruction. ACM Transactions on Graphics (TOG), 31(6), 194.