A Reduced-Precision Network for Image Reconstruction

Published: 12/03/2020

Detailed Description

By Manu Mathew Thomas (Intel Corporation), Karthik Vaidyanathan (Intel Corporation), Gabor Liktor (Intel Corporation), Angus G. Forbes (University of California, Santa Cruz)


A comparison of different reconstruction techniques on a frame from the INFILTRATOR scene with PSNR (top row) and SSIM (bottom row) metrics. TAA has the worst reconstruction quality with excessive blurring followed by direct prediction with U-Net. QW-Net produces the best results with quality comparable to brute force sampling with 256 samples per pixel.

Abstract

Neural networks are often quantized to use reduced-precision arithmetic, as it greatly improves their storage and computational costs. This approach is commonly used in image classification and natural language processing applications. However, using a quantized network for the reconstruction of HDR images can lead to a significant loss in image quality. In this paper, we introduce QW-Net, a neural network for image reconstruction, in which close to 95% of the computations can be implemented with 4-bit integers. This is achieved using a combination of two U-shaped networks that are specialized for different tasks, a feature extraction network based on the UNet architecture, coupled to a filtering network that reconstructs the output image. The feature extraction network has more computational complexity but is more resilient to quantization errors. The filtering network, on the other hand, has significantly fewer computations but requires higher precision. Our network recurrently warps and accumulates previous frames using motion vectors, producing temporally stable results with significantly better quality than TAA, a widely used technique in current games.
Research Area: Neural networks, Rendering, kernel prediction, antialiasing.
Published in SIGGRAPH ASIA 2020

Notice: This software is subject to the U.S. Export Administration Regulations and other U.S. law, and may not be exported or re-exported to certain countries (Cuba, Iran, North Korea, Sudan, and Syria) or to persons or entities prohibited from receiving U.S. exports (including Denied Parties, Specially Designated Nationals, and entities on the Bureau of Export Administration Entity List or involved with missile technology or nuclear, chemical or biological weapons).

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.