A deep convolutional neural network based on U-Net to predict annual luminance maps

Abstract

Studying annual luminance maps during the design process provides architects with insight into the space's spatial quality and occupants’ visual comfort. Simulating annual luminance maps is computationally expensive, especially if the objective is to render the scene for multiple viewpoints. This paper proposes a method based on deep learning that accelerates these simulations by predicting the annual luminance maps using only a limited number of rendered high-dynamic-range images. Our proposed model predicts HDR images that are comparable to the rendered ones. Using the transfer learning approach, our model can robustly predict HDR images from other viewpoints in the space with less rendered images and less training required. We evaluated our method using various evaluation metrics, such as MSE, RER, PSNR, SSIM, and runtime duration. Our method shows improvements in all metrics compared to the previous work, especially 33% better MSE loss, 48% more accurate DGP values, and 50% faster runtime.

KEYWORDS:

• Daylighting simulation
• luminance maps
• Radiance rendering
• HDR imagery
• convolutional neural network
• machine learning

Data availability

After the publication of this paper, we will release our data set and code at this address: https://github.com/maqorbani/DCNU_Lighting. Additionally, an application with a user-friendly UI will be provided to make this method available to architects.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 (-ab n) is an argument for the Radiance software RPICT command that determines n number of maximum ambient bounces to be calculated for the desired rendering output.

2 (-vtv) is an argument for the Radiance software RPICT command that sets the rendering view type to perspective.

3 -af is an argument for the Radiance software RPICT command that determines where the program should store or retrieve its calculated indirect illuminance values.