APSIPA Transactions on Signal and Information Processing > Vol 13 > Issue 1

Deep-Learning for Objective Quality Assessment of Tone Mapped Images

Ishtiaq Rasool Khan, Abu Dhabi School of Management, United Arab Emirates AND College of Computer Science and Engineering, University of Jeddah, Saudi Arabia, Romil Imtiaz, Pakistan Institute of Engineering and Technology, Pakistan, romilimtiaz302@gmail.com
 
Suggested Citation
Ishtiaq Rasool Khan and Romil Imtiaz (2024), "Deep-Learning for Objective Quality Assessment of Tone Mapped Images", APSIPA Transactions on Signal and Information Processing: Vol. 13: No. 1, e25. http://dx.doi.org/10.1561/116.20240007

Publication Date: 21 Oct 2024
© 2024 I. R. Khan and R. Imtiaz
 
Subjects
 
Keywords
Image quality assessmentdeep learningimage datasetstone mappinghigh dynamic range
 

Share

Open Access

This is published under the terms of CC BY-NC.

Downloaded: 98 times

In this article:
Introduction 
Literature Review 
Proposed Dataset 
The Proposed Deep-Learning Model 
Performance Validation 
Conclusion 
References 

Abstract

High dynamic range (HDR) images capture real-world luminance values which cannot be directly displayed on the screen and require tone mapping to be shown on low dynamic range (LDR) hardware. During this transformation, tone mapping algorithms are expected to preserve the naturalness and structural details of the image. In this regard, the performance of a tone mapping algorithm can be evaluated through a subjective study where participants rank or score tone mapped images based on their preferences. However, such subjective evaluations can be time-consuming and cannot be repeated for every tone mapped image. To address this issue, numerous quantitative metrics have been proposed for objective evaluation. This paper presents a robust objective metric based on deep learning to quantify image quality. We assess the performance of our proposed metric by comparing it to 20 existing state-of-the-art metrics using two subjective datasets, including one benchmark dataset and a novel proposed dataset of 666 tone mapped images comprising a variety of scenes and labeled by 20 users. Our approach exhibits the highest correlation with subjective scores in both evaluations, confirming its effectiveness and potential to be a reliable alternative to laborious subjective studies.

DOI:10.1561/116.20240007