High-Resolution Photo Enhancement in Real-Time: A Laplacian Pyramid Network.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Current photo enhancement methods face a trade-off between performance and computational efficiency.
• High-performance models are often too resource-intensive for edge devices.
• Efficiency-focused models frequently deliver insufficient quality for practical use.

Purpose of the Study:

• Introduce LLF-LUT++, a novel pyramid network for efficient and high-performance photo enhancement.
• Address the limitations of existing methods by integrating global and local image processing techniques.
• Enable fast processing of high-resolution images on resource-constrained devices.

Main Methods:

• Utilized closed-form Laplacian pyramid decomposition and reconstruction to integrate global and local operators.
• Employed an image-adaptive 3D Look-Up Table (LUT) for global tonal enhancement based on downsampled image characteristics.
• Incorporated a spatial-frequency transformer weight predictor and local Laplacian filters for adaptive detail refinement.

Main Results:

• Achieved a 2.64 dB improvement in Peak Signal-to-Noise Ratio (PSNR) on the HDR+ dataset.
• Processed 4K resolution images in 13 milliseconds on a single GPU, demonstrating significant speed improvements.
• Outperformed state-of-the-art methods in extensive experiments on two benchmark datasets.

Conclusions:

• LLF-LUT++ successfully balances enhancement performance with computational efficiency.
• The proposed network architecture and transformer model enable rapid, high-quality image enhancement.
• The method offers a viable solution for real-world applications requiring fast and effective photo enhancement.