[Paper Review] Deep Inverse Tone Mapping Using LDR Based Learning for Estimating HDR Images with Absolute Luminance

Yuma Kinoshita, Hitoshi Kiya|arXiv (Cornell University)|Feb 28, 2019

Image Enhancement Techniques20 references4 citations

TL;DR

This paper proposes a deep learning-based inverse tone mapping method that trains a CNN using LDR images mapped via Reinhard’s global operator, enabling estimation of HDR images with absolute luminance. By learning the inverse transform of this operator, the method achieves superior HDR quality—outperforming state-of-the-art methods in HDR-VDP-2.2 and PU encoding + MS-SSIM metrics—without requiring HDR-specific loss functions or multi-exposure inputs.

ABSTRACT

In this paper, a novel inverse tone mapping method using a convolutional neural network (CNN) with LDR based learning is proposed. In conventional inverse tone mapping with CNNs, generated HDR images cannot have absolute luminance, although relative luminance can. Moreover, loss functions suitable for learning HDR images are problematic, so it is difficult to train CNNs by directly using HDR images. In contrast, the proposed method enables us not only to estimate absolute luminance, but also to train a CNN by using LDR images. The CNN used in the proposed method learns a transformation from various input LDR images to LDR images mapped by Reinhard's global operator. Experimental results show that HDR images generated by the proposed method have higher-quality than HDR ones generated by conventional inverse tone mapping methods,in terms of HDR-VDP-2.2 and PU encoding + MS-SSIM.

Motivation & Objective

• To address the limitation of existing CNN-based inverse tone mapping methods that cannot estimate absolute luminance in generated HDR images.
• To overcome the difficulty of designing effective loss functions for direct HDR image training, which is hindered by the wide dynamic range and non-linear response of HDR data.
• To enable end-to-end training of a CNN using only LDR images, avoiding the need for paired HDR-LDR data or multi-exposure inputs.
• To achieve high-quality HDR reconstruction with accurate absolute luminance while maintaining a simple network architecture.

Proposed method

• The CNN is trained to map input LDR images to LDR images that have been tone-mapped using Reinhard’s global operator, which is invertible when parameters are known.
• The method leverages the invertibility of Reinhard’s operator: after the CNN predicts tone-mapped LDR outputs, the inverse transform is applied to recover HDR images with absolute luminance.
• Key parameters—brightness factor $a$ and geometric mean $G(E)$—are estimated from the LDR output and input, enabling reconstruction of irradiance $E$ via $E'_{i,j} = \frac{G(E)}{a} \cdot \frac{I_{i,j}}{1 - I_{i,j}}$.
• The training loss is computed in the LDR domain using standard pixel-wise regression (e.g., L1 or L2), avoiding complex HDR-specific loss functions.
• The network architecture is a U-Net-like encoder-decoder with skip connections, optimized using Adam with $\alpha=0.002$, $\beta_1=0.5$, $\beta_2=0.999$.
• The method avoids the need for multi-exposure supervision or explicit HDR supervision, relying solely on LDR-based learning.

Experimental results

Research questions

• RQ1 Can a CNN trained on LDR images alone produce HDR images with accurate absolute luminance?
• RQ2 Does learning a transformation to Reinhard’s tone-mapped LDR outputs enable better HDR reconstruction than direct HDR learning or multi-exposure methods?
• RQ3 Can the proposed method outperform state-of-the-art inverse tone mapping techniques in both objective and subjective quality metrics?
• RQ4 Is the proposed method more efficient and simpler than existing CNN-based approaches that require complex architectures or multi-stage training?

Key findings

• The proposed method achieved the highest HDR-VDP-2.2 scores across all seven test images, with a maximum score of 61.49 (Image 7), significantly outperforming Direct ITMO (44.77) and DrTMO (43.82).
• In PU encoding + MS-SSIM evaluation, the proposed method achieved the highest scores on all images, with 0.9804 on Image 7, compared to 0.6997 for PMET and 0.6042 for DrTMO.
• Visual comparisons showed that the proposed method preserved fine details in dark regions better than Direct ITMO and PMET, and avoided the color distortion seen in DrTMO’s overexposed outputs.
• The method achieved higher quality than DrTMO despite having a simpler network architecture, as DrTMO used two 3D deconvolutional networks.
• The results confirm that LDR-based learning with inverse tone mapping via Reinhard’s operator enables accurate absolute luminance estimation without HDR supervision.
• The method demonstrates that training in the LDR domain with a known, invertible tone mapping function is a viable and effective alternative to direct HDR training.