Accelerated Self-Supervised Multi-Illumination Color Constancy With Hybrid Knowledge Distillation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human Visual System

Background:

• Color constancy is vital for accurate color perception under changing illumination.
• Deep learning methods show promise but are limited by dataset scale and model size.
• Existing approaches struggle with effective discriminative feature learning and practical camera deployment.

Purpose of the Study:

• To propose a multi-illumination color constancy approach overcoming current limitations.
• To enhance feature learning and model efficiency for practical applications.
• To improve color constancy performance on both multi-illumination and single-illumination benchmarks.

Main Methods:

• A three-phase approach: self-supervised pre-training, supervised fine-tuning, and knowledge distillation.
• Pre-training utilizes Transformer and U-Net encoders with light normalization and grayscale colorization pretext tasks.
• Knowledge distillation aligns CNN features with Transformer and U-Net features using a hybrid technique.

Main Results:

• The proposed method outperforms state-of-the-art techniques on multi-illumination and single-illumination datasets.
• A lightweight decoder achieves better illumination distributions with fewer parameters.
• Ablation studies and visualizations validate the model's effectiveness.

Conclusions:

• The self-supervised learning and knowledge distillation approach significantly advances color constancy.
• The method offers improved feature learning and model efficiency for real-world deployment.
• This research provides a robust solution for accurate color perception across diverse lighting conditions.