Modern imaging systems struggle to capture both very bright and dark regions within a single shot due to limited dynamic range.

This project presents a Human-Perception-Based Deep Learning Model that uses Convolutional Neural Networks (CNNs) to generate perceptually guided weight maps for multi-exposure image fusion (MEF).

Unlike traditional fusion methods, this model integrates perceptual cues such as contrast, color saturation, well-exposedness, local entropy, and visibility — enabling image fusion that mimics human visual perception.

Traditional MEF methods fail under extreme exposure variations or in real-time scenarios due to manual weight computation and poor perceptual adaptation.

Inspired by how the human brain balances brightness and color under diverse lighting conditions, this model seeks to integrate AI-based perceptual learning into exposure fusion.

• Develop a deep learning-based framework for weight map generation guided by human perceptual features.
• Implement and analyze five perceptual quality parameters:
  • Contrast
  • Saturation
  • Well-exposedness
  • Local entropy
  • Visibility
• Train on a dataset of over 10,000 HDR images covering diverse illumination conditions.
• Achieve real-time and artifact-free exposure fusion comparable with or superior to existing MEF methods.

1. Input
   Multiple images of the same scene captured at different exposure levels (under, normal, over-exposed).

2. Feature Extraction
   For each image, the CNN extracts five perceptual features:

   • Contrast — via Laplacian gradients.
   • Saturation — via RGB standard deviation.
   • Well-exposedness — via Gaussian distribution centered at mid-intensity.
   • Local Entropy — via local texture complexity.
   • Visibility — via gradient magnitude.

3. Weight Map Generation
   The CNN learns to assign pixel-wise perceptual weights balancing brightness and contrast intelligently.

4. Fusion Process
   Multi-scale representation using Gaussian and Laplacian pyramids ensures natural blending and accurate edge preservation.

5. Output
   A single perceptually optimized fused image maintaining visibility, contrast, and natural illumination.

• Python 3.x
• TensorFlow / PyTorch (Deep Learning Framework)
• OpenCV (Image Processing)
• NumPy, Matplotlib (Computation and Visualization)
• Google Colab / Jupyter Notebook (Model Training & Testing)

• Generated perceptual scores closely align with human visual judgment.
• Fused images demonstrate improved clarity, contrast, and realism.
• Effective for photography, surveillance, and medical imaging.

• Contains ~10,000 HDR images with varied illumination, textures, and contrast.
• Used for future model training, evaluation, and optimization.
• Dataset planned for public release after model validation.

• Indigenous Perception Model — Tailoring fusion for specific lighting conditions or cultural visual biases.
• Integration into Unified CNN Model — Streamlined, perception-aware real-time fusion.
• End-to-End Deep Fusion Network — Replace pyramid-based fusion with trainable CNN/Transformer model for efficiency.

The work builds upon seminal research in multi-exposure image fusion and perceptual modeling, including:

• Mertens et al., Exposure Fusion: A Practical Alternative to HDR (2007)

• Prabhakar et al., DeepFuse: A Deep Unsupervised Approach for Exposure Fusion (ICCV 2017)

• Yang et al., GANFuse: Generative Adversarial Fusion Framework (2021)

• Qu et al., TransMEF: Transformer-Based Self-Supervised Fusion (ICIP 2020)

📄 Developed as part of a research project on human perception-guided deep learning for exposure fusion.