Ranking-Based Color Constancy With Limited Training Samples.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Computational color constancy is crucial for white balancing in imaging devices.
• Deep convolutional neural networks (CNNs) show promise but are computationally expensive and large for real-time applications.
• Existing methods struggle with resource-constrained environments.

Purpose of the Study:

• To develop an efficient color constancy method suitable for low-resource Image Signal Processors (ISPs).
• To achieve performance comparable to CNN-based methods without their computational burden.
• To propose a novel ranking-based approach for selecting optimal simple statistics.

Main Methods:

• Proposed a novel ranking-based color constancy (RCC) method formulating SM selection as a label ranking problem.
• Designed a specific ranking loss function, incorporating low rank and grouped sparse constraints.
• Extended RCC to RCC_NO, enabling learning without ground truth illumination using partial binary preference annotations.

Main Results:

• RCC significantly outperforms shallow learning methods and matches/exceeds CNN-based methods in performance.
• RCC achieves this with drastically reduced model size (1/2000) and training time.
• RCC_NO demonstrates strong performance with reduced annotation costs and no reliance on ground truth illumination.

Conclusions:

• The proposed RCC method offers an efficient and effective solution for real-time color constancy on low-resource devices.
• RCC provides a viable alternative to computationally intensive deep learning models.
• RCC_NO presents a practical approach for color constancy when ground truth data is unavailable or costly to obtain.