Learning Image-Adaptive 3D Lookup Tables for High Performance Photo Enhancement in Real-Time.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Learning-based photo enhancement methods are popular but often resource-intensive.
• Existing methods struggle with high-resolution images due to computational and memory demands.
• Traditional 3D LUTs are manually tuned and fixed, limiting their adaptability.

Purpose of the Study:

• To develop a fast and robust photo enhancement technique for high-resolution images.
• To introduce image-adaptive 3-dimensional lookup tables (3D LUTs) learned from data.
• To overcome the limitations of existing methods in terms of efficiency and performance.

Main Methods:

• Learning image-adaptive 3D LUTs from annotated data using pairwise or unpaired learning.
• Simultaneously training multiple basis 3D LUTs and a small convolutional neural network (CNN) end-to-end.
• Utilizing a CNN on down-sampled images to predict weights for fusing basis 3D LUTs into an adaptive one.

Main Results:

• The proposed model achieves high efficiency, processing 4K images in under 2 ms with minimal parameters (<600K).
• Outperforms state-of-the-art methods in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and color difference metrics.
• Demonstrates superior performance on two public benchmark datasets.

Conclusions:

• Learned image-adaptive 3D LUTs offer a highly efficient and effective solution for photo enhancement.
• The proposed CNN-based fusion method enables flexible and content-dependent color and tone transformations.
• This approach significantly advances the practical application of photo enhancement for high-resolution imagery.