An Untrained Neural Network Prior for Light Field Compression.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Deep generative models are effective for image processing but typically require large datasets.
• The Deep Image Prior (DIP) concept shows untrained neural networks can serve as image priors.
• Light field data presents unique challenges for image processing and compression.

Purpose of the Study:

• To propose a compact deep generative model for light fields that does not require external training data.
• To develop a complete light field compression scheme utilizing the proposed generative model.
• To evaluate the compression performance against state-of-the-art methods.

Main Methods:

• A novel, compact deep generative model tailored for light fields was developed.
• The model was integrated into a light field compression framework.
• Quantization-aware learning and entropy coding of weights were employed for efficient compression.

Main Results:

• The proposed method achieved competitive results compared to existing light field compression techniques.
• Performance was evaluated using standard metrics such as Peak Signal-to-Noise Ratio (PSNR) and Multi-Scale Structural Similarity (MS-SSIM).

Conclusions:

• The proposed compact deep generative model is a viable approach for light field prior representation.
• The developed compression scheme demonstrates the practical utility and effectiveness of the generative model.
• This method offers a promising direction for efficient light field data compression without large training datasets.