Comparison of Training Strategies for Autoencoder-Based Monochromatic Image Denoising.

Area of Science:

• Image processing
• Machine learning
• Scientific imaging

Background:

• Monochromatic images are crucial for signal intensity analysis.
• Image noise significantly degrades object identification and intensity estimation.
• Deterministic denoising algorithms like Non-Local-Means and Block-Matching-3D are state-of-the-art.

Purpose of the Study:

• To explore machine learning for monochromatic image denoising.
• To evaluate ML performance across various data availability scenarios, including limited or no noise-free data.
• To assess the impact of training strategies and dataset characteristics on denoising efficacy.

Main Methods:

• Utilized a simple autoencoder architecture for image denoising.
• Trained and tested the autoencoder on MNIST and CIFAR-10 datasets.
• Investigated different training approaches for the autoencoder model.

Main Results:

• Machine learning denoising performance is influenced by training methods, autoencoder architecture, and dataset image similarity.
• The proposed ML methods achieve performance exceeding current state-of-the-art deterministic algorithms.
• Effective denoising is possible even when training data lacks noise-free examples.

Conclusions:

• Machine learning offers a promising and effective approach for monochromatic image denoising.
• ML-based denoising should be considered a viable alternative to traditional methods.
• Further research into ML training strategies can optimize denoising for specific applications.