Single-Image HDR Reconstruction With Task-Specific Network Based on Channel Adaptive RDN

Guannan Chen, Lijie Zhang, Mengdi Sun, Yan Gao, Pablo Navarrete Michelini, Yanhong WuWorkshops @ CVPR 2021 · DOI

Abstract

High Dynamic Range (HDR) reconstruction from a single Low Dynamic Range (LDR) image is a challenging task due to the significant loss of information in over-exposed and under-exposed regions. This paper presents the Task-specific Network based on Channel Adaptive Residual Dense Network (TCRDN), a modular architecture designed to address this problem by decomposing it into three specialized sub-tasks: image reconstruction, detail restoration, and local contrast enhancement. By integrating Gate Channel Transformation (GCT) layers into a Residual Dense Network (RDN) backbone, the model effectively prioritizes feature channels critical for HDR recovery. TCRDN achieved 1st place in PSNR and 3rd place in muPSNR at the NTIRE 2021 Single-Image HDR Challenge, demonstrating state-of-the-art performance in both pixel fidelity and visual quality.

Introduction

Standard consumer cameras often fail to capture the full dynamic range of real-world scenes, resulting in LDR images with "blown-out" highlights or "crushed" shadows. While multi-exposure fusion is a common solution, it is prone to ghosting artifacts in dynamic scenes. Single-image HDR reconstruction aims to recover the missing dynamic range from a single frame, a mathematically ill-posed problem given the high variance of HDR values. Existing methods often struggle to balance high pixel-level accuracy (PSNR) with the perceptual quality required for tone-mapped displays (muPSNR).

Methodology

The core of the proposed approach is the Task-specific Network based on Channel Adaptive RDN (TCRDN). The architecture leverages two key innovations: task decomposition and a channel-adaptive backbone.

Task Decomposition

Rather than employing a single monolithic network, TCRDN utilizes three specialized subnets to handle different aspects of HDR reconstruction:

  1. Image Reconstruction (IR): Recovers the coarse, global HDR structure.
  2. Detail Restoration (DR): Focuses on fine-grained textures and edge details.
  3. Local Contrast Enhancement (LCE): Produces a contrast mask to improve the local dynamic range and visual "pop."
The TCRDN architecture decomposes HDR reconstruction into three specialized sub-tasks: Image Reconstruction (IR), Detail Restoration (DR), and Local Contrast Enhancement (LCE).

The outputs of these subnets are fused to generate the final HDR image. This "divide and conquer" strategy leads to more stable training and superior performance compared to single-purpose architectures.

Channel Adaptive RDN Backbone

Each subnet is built upon a Residual Dense Network (RDN) enhanced with Gate Channel Transformation (GCT) layers. The GCT layer serves as an attention mechanism that enables the network to adaptively weight feature channels based on their importance to the specific task. This is achieved through global context embedding and channel normalization, allowing the model to focus on critical features like highlight gradients while suppressing less relevant information.

Experiments and Results

The model was evaluated using the NTIRE 2021 HDR Challenge dataset, derived from the HdM HDR dataset. The dataset includes diverse indoor and outdoor scenes captured with professional Alexa Arri cameras. LDR inputs were synthetically generated using realistic degradation models, including noise, quantization, and clipping.

Quantitative and Qualitative Performance

TCRDN demonstrated exceptional performance in the NTIRE 2021 competition, securing 1st place in PSNR (33.49 dB) and 3rd place in muPSNR (34.41 dB). Ablation studies confirmed that the inclusion of DR and LCE modules significantly improved fidelity, raising the PSNR from 35.87 dB (IR only) to 36.41 dB on the validation set.

Visual comparison between the ground truth, medium-exposure input, and the TCRDN reconstruction, highlighting the model's ability to restore dynamic range and fidelity.

Qualitatively, TCRDN significantly outperformed baseline models like ExpandNet, providing sharper details and more accurate color reproduction in challenging lighting conditions.

Conclusion and Future Work

TCRDN provides a robust framework for single-image HDR reconstruction by combining task-specific sub-networks with a channel-adaptive backbone. While the model excels at denoising and contrast restoration, it faces limitations in regions of extreme saturation where LDR information is completely lost. Future work may explore the integration of Generative Adversarial Networks (GANs) to synthesize realistic textures in these clipped regions, potentially further improving visual quality at the cost of some pixel-level fidelity.

Summary updated

BibTeX
@inproceedings{chen2021single,
  author = {Guannan Chen and Lijie Zhang and Mengdi Sun and Yan Gao and Pablo Navarrete Michelini and Yanhong Wu},
  title = {{Single-Image HDR Reconstruction With Task-Specific Network Based on Channel Adaptive RDN}},
  booktitle = {Workshops @ CVPR},
  year = {2021},
  doi = {10.1109/cvprw53098.2021.00050},
}
Figure