Learning Converged Propagations with Deep Prior Ensemble for Image Enhancement

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Image enhancement is crucial for vision and learning applications.
• Existing methods include knowledge-driven (MAP) and data-driven (CNN) approaches.
• These methods have limitations in robustness and theoretical guarantees.

Purpose of the Study:

• To propose a unified framework, Deep Prior Ensemble (DPE), integrating knowledge-based and data-driven techniques.
• To enhance image visual qualities across various applications.
• To provide a theoretically sound and robust image enhancement solution.

Main Methods:

• Developed a unified framework (DPE) combining prior modeling and residual Convolutional Neural Networks (CNNs).
• Established a basic propagation scheme using image modeling cues and CNNs for direction prediction.
• Incorporated prior projections for feedback control to ensure convergence and constraint satisfaction.

Main Results:

• Theoretically proved DPE converges and satisfies fundamental task constraints.
• Demonstrated DPE's robustness against local minimums compared to MAP approaches.
• Showcased DPE's superior performance over state-of-the-art methods on various image enhancement tasks.

Conclusions:

• DPE offers a generic ensemble methodology integrating domain knowledge and data-driven cues.
• The framework provides theoretical guarantees for feedforward propagation control.
• Experimental results confirm DPE's effectiveness in improving quantitative measures and visual perception quality.