Core Concepts
FastHDRNet, a lightweight and efficient deep learning framework, achieves state-of-the-art performance in converting standard dynamic range (SDR) television content to high dynamic range (HDR) television, while significantly reducing computational complexity compared to previous methods.
Abstract
The paper introduces FastHDRNet, a novel deep learning framework for converting standard dynamic range (SDR) television content to high dynamic range (HDR) television. The framework consists of two key components:
Adaptive Universal Color Transformation (AUCT):
The AUCT network is designed to perform global color mapping from SDR to HDR, comprising a base network and a conditioning network.
The base network uses a fully convolutional architecture with 1x1 convolutions to emulate a 3D lookup table for efficient color mapping.
The conditioning network extracts global priors, such as color harmony and feature consistency, to adaptively modulate the base network.
Local Enhancement (LE):
The LE network, based on a U-Net architecture, refines the output of the AUCT network to further improve visual quality and address spatially variant mapping.
The LE network leverages spatial feature transformation (SFT) layers to effectively modulate the intermediate features.
The authors construct a new dataset, HDRTV1K, to train and evaluate the proposed method. Extensive experiments demonstrate that FastHDRNet achieves state-of-the-art performance in both quantitative and visual quality metrics, while significantly reducing the computational complexity compared to previous methods.
Stats
The HDRTV1K dataset used in the experiments consists of 22 HDR video sequences and their corresponding SDR versions, encoded in the PQ-OETF within the Rec.2020 color space. 18 pairs were used for training, and the remaining 4 pairs were used for testing.
Quotes
"Our method has the fastest reference time in all the algorithms mentioned."
"FastHDRNet significantly reduces the computational cost and runs much faster compared to HDRTVNet while achieving better performance."