InDeed: Interpretable image deep decomposition with grounded generalizability

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Image decomposition is crucial for downstream tasks, offering interpretability.
• Deep learning excels at image analysis but often lacks interpretability and generalizability.
• Combining deep learning with interpretable image decomposition is an underexplored area.

Purpose of the Study:

• To develop a novel framework for image decomposition into low-rank, sparse, and noise components.
• To integrate hierarchical Bayesian modeling with deep learning for improved interpretability and generalizability.
• To enhance deep neural networks (DNNs) for image analysis tasks.

Main Methods:

• Hierarchical Bayesian modeling of image decomposition.
• Transforming inference problems into optimization tasks.
• Deep inference using a modularized Bayesian DNN under a relaxed amortized formulation.
• Analysis of generalization error bounds using PAC-Bayes theory.

Main Results:

• The proposed framework successfully decomposes images into low-rank, sparse, and noise components.
• Demonstrated improved generalizability and interpretability in image denoising and unsupervised anomaly detection tasks.
• Developed a test-time adaptation approach for out-of-distribution scenarios.

Conclusions:

• The novel framework effectively combines Bayesian modeling and deep learning for interpretable image decomposition.
• The method offers enhanced generalizability and interpretability for various computer vision tasks.
• The approach provides a foundation for more robust and understandable deep learning models in image analysis.