Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HDR Image Reconstruction using an Unsupervised Fusion Model

Kumbha Nagaswetha

TL;DR

This work tackles HDR reconstruction by combining traditional multi-exposure fusion with an unsupervised deep learning framework. An encoder–decoder CNN predicts per-pixel weights for two-exposure inputs, optimizing a weighted SSIM loss that blends local variance, gradient, and well-exposedness cues; the fusion follows $I_{\mathrm{fused}}(x,y) = \sum_{n=1}^{N} W_n(x,y) I_n(x,y)$, with a hybrid variance–gradient $\gamma$ achieving the best performance ($$MEF-SSIM = 0.9921$$). An adaptive MEF baseline using well-exposedness and gradient maps is also developed, achieving $MEF$-SSIM of $0.9207$ on five sequences. The approach removes the need for ground-truth HDR data and demonstrates practical, high-quality HDR reconstruction for static scenes in real-world settings by leveraging per-pixel adaptive fusion weights guided by perceptual cues.

Abstract

High Dynamic Range (HDR) imaging aims to reproduce the wide range of brightness levels present in natural scenes, which the human visual system can perceive but conventional digital cameras often fail to capture due to their limited dynamic range. To address this limitation, we propose a deep learning-based multi-exposure fusion approach for HDR image generation. The method takes a set of differently exposed Low Dynamic Range (LDR) images, typically an underexposed and an overexposed image, and learns to fuse their complementary information using a convolutional neural network (CNN). The underexposed image preserves details in bright regions, while the overexposed image retains information in dark regions; the network effectively combines these to reconstruct a high-quality HDR output. The model is trained in an unsupervised manner, without relying on ground-truth HDR images, making it practical for real-world applications where such data is unavailable. We evaluate our results using the Multi-Exposure Fusion Structural Similarity Index Measure (MEF-SSIM) and demonstrate that our approach achieves superior visual quality compared to existing fusion methods. A customized loss function is further introduced to improve reconstruction fidelity and optimize model performance.

HDR Image Reconstruction using an Unsupervised Fusion Model

TL;DR

This work tackles HDR reconstruction by combining traditional multi-exposure fusion with an unsupervised deep learning framework. An encoder–decoder CNN predicts per-pixel weights for two-exposure inputs, optimizing a weighted SSIM loss that blends local variance, gradient, and well-exposedness cues; the fusion follows , with a hybrid variance–gradient achieving the best performance (). An adaptive MEF baseline using well-exposedness and gradient maps is also developed, achieving -SSIM of on five sequences. The approach removes the need for ground-truth HDR data and demonstrates practical, high-quality HDR reconstruction for static scenes in real-world settings by leveraging per-pixel adaptive fusion weights guided by perceptual cues.

Abstract

High Dynamic Range (HDR) imaging aims to reproduce the wide range of brightness levels present in natural scenes, which the human visual system can perceive but conventional digital cameras often fail to capture due to their limited dynamic range. To address this limitation, we propose a deep learning-based multi-exposure fusion approach for HDR image generation. The method takes a set of differently exposed Low Dynamic Range (LDR) images, typically an underexposed and an overexposed image, and learns to fuse their complementary information using a convolutional neural network (CNN). The underexposed image preserves details in bright regions, while the overexposed image retains information in dark regions; the network effectively combines these to reconstruct a high-quality HDR output. The model is trained in an unsupervised manner, without relying on ground-truth HDR images, making it practical for real-world applications where such data is unavailable. We evaluate our results using the Multi-Exposure Fusion Structural Similarity Index Measure (MEF-SSIM) and demonstrate that our approach achieves superior visual quality compared to existing fusion methods. A customized loss function is further introduced to improve reconstruction fidelity and optimize model performance.
Paper Structure (15 sections, 8 equations, 5 figures, 2 tables)

This paper contains 15 sections, 8 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Overview of Methods
  3. Using Multiple SDR Images
  4. Using a Single SDR Image
  5. Using Deep Learning Methods
  6. Multi-Exposure Image Fusion Based on Adaptive Weights
  7. Experimental Results
  8. HDR Image Reconstruction Using an Unsupervised Learning Model
  9. Model Architecture
  10. Encoder
  11. Decoder
  12. Proposed Loss Function
  13. Experiments and Results
  14. Dataset and Training
  15. Conclusion

Figures (5)

  • Figure 1: Comparison of MEF outputs: combined weights produce balanced HDR images with improved detail preservation.
  • Figure 2: Overall CNN architecture for unsupervised HDR reconstruction. The encoder extracts feature maps, and the decoder reconstructs the weight maps $W_n$ used for HDR fusion.
  • Figure 3: Encoder and Decoder sub-modules of the network. Encoder extracts features, and decoder reconstructs weight maps used for HDR fusion.
  • Figure 4: Gamma visualization on under and over-exposed images. The combination of variance and gradient extracts fine structural details, while other combinations yield smoother but less textured regions.
  • Figure 5: HDR reconstructions for different $\gamma$ definitions on a test image. The hybrid variance–gradient formulation yields the most balanced and detailed reconstruction.