Abstract
Overexposure in imaging systems leads to irreversible loss of scene information due to sensor saturation, resulting in degraded perceptual quality and limited dynamic range. While recent deep learning approaches have shown promising results in correcting mild to moderate overexposure, their performance under severe overexposure and their perceptual quality in HDR image generation remain less understood. In this work, a comparative evaluation of four representative models—Multi-Scale Exposure Correction (MSEC), Illumination-Adaptive Transformer (IAT), LEDiff, and X2HDR—is conducted across multiple scenes and levels of overexposure, with particular emphasis on severe overexposure conditions. The evaluation focuses on the perceptual quality of the reconstructed HDR images, combining qualitative analysis, chromaticity-based assessment, effective dynamic range measurements, and a controlled psychophysical experiment using a two-alternative forced choice (2AFC) paradigm on an HDR display. The results show that conventional reconstruction-based methods exhibit limited performance in severely overexposed regions, often failing to recover meaningful structure or producing visible artifacts. Diffusion-based models demonstrate improved perceptual quality, with X2HDR consistently achieving higher observer preference and producing visually plausible reconstructions. However, despite their perceptual advantages, generative models do not guarantee faithful reconstruction of the original scene. The outputs may contain hallucinated structures and inaccurate colors, reflecting a reliance on learned priors rather than physical consistency. This limitation is particularly important in HDR imaging, where accurate representation of scene radiometry and color is critical. Overall, this work shows that while diffusion-based models represent a significant advancement in perceptual HDR image generation, their reliability under severe overexposure and their ability to preserve color fidelity remain limited. These findings highlight the need for future approaches that better balance perceptual quality with color accuracy in HDR reconstruction.
Publication Date
5-2026
Document Type
Thesis
Student Type
Graduate
Degree Name
Color Science (MS)
Department, Program, or Center
Color Science
College
College of Science
Advisor
Mekides A. Abebe
Advisor/Committee Member
Mark D. Fairchild
Advisor/Committee Member
Susan Farnand
Recommended Citation
Rabbanifar, Alireza, "Perceptual Evaluation of Deep Learning Models for HDR Reconstruction Under Severe Overexposure" (2026). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/12620
Campus
RIT – Main Campus
