An Efficient and Low-Complexity Transformer-Based Deep Learning Framework for High-Dynamic-Range Image...

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• High-dynamic-range (HDR) image reconstruction enhances image quality by merging multiple low-dynamic-range images.
• Challenges include frame misalignment, overexposure, and motion, often tackled with deep learning.
• Existing Transformer architectures offer high quality but incur substantial computational expense.

Purpose of the Study:

• To propose a novel HDR reconstruction architecture that balances state-of-the-art quality with reduced computational cost.
• To leverage Transformer-based approaches for improved HDR image generation.
• To enhance feature refinement and prevent quality degradation during reconstruction.

Main Methods:

• Developed a Transformer-based architecture for HDR image reconstruction.
• Reduced the number of self-attention blocks for computational efficiency.
• Incorporated a Convolutional Block Attention Module for feature enhancement, using the central frame as a reference.

Main Results:

• Achieved competitive results with state-of-the-art methods on HDR reconstruction tasks.
• Demonstrated significant computational efficiency compared to other Transformer-based approaches.
• Obtained the best quality metrics on the Tel's dataset.

Conclusions:

• Low-complexity Transformer-based architectures are highly promising for HDR reconstruction.
• The proposed method offers a practical and efficient solution for HDR image enhancement.
• Potential applications extend beyond HDR to other image processing domains.