Few-shot trained single-lens diffractive neural network for all-optical image denoising.

Area of Science:

• Optics
• Artificial Intelligence
• Image Processing

Background:

• Diffractive neural networks (DNNs) offer high-speed analog image denoising.
• Existing all-optical DNNs suffer from large data requirements, poor generalization, and complex multi-layer structures.

Purpose of the Study:

• To develop a novel few-shot trained single-lens DNN (SL-DNN) for efficient and generalized image denoising.
• To overcome the limitations of existing all-optical DNNs in terms of data needs and structural complexity.

Main Methods:

• Incorporating a DNN into a single-lens imaging system to leverage spatial frequency filtering priors.
• Utilizing a single diffractive layer architecture (SL-DNN) for denoising.
• Employing few-shot learning techniques for training the SL-DNN.

Main Results:

• The SL-DNN achieved superior image denoising performance compared to a five-layer DNN using only one diffractive layer.
• The SL-DNN demonstrated significantly improved generalization capabilities on diverse datasets (Quick Draw, ECSSD) despite few-shot training.
• The proposed method avoided alignment difficulties and diffraction efficiency losses associated with multi-layer structures.

Conclusions:

• The SL-DNN presents a more efficient and generalized approach to optical image denoising.
• This physics-inspired method offers a promising strategy for designing task-specific visual processors for applications like image display and projection.