Towards continual low-light image enhancement through causal inference.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Low-light image enhancement aims to improve image quality in poor lighting.
• Existing methods struggle with varying lightness distributions, limiting generalization.
• Continual learning presents a challenge due to catastrophic forgetting.

Purpose of the Study:

• To address the continual low-light image enhancement task for the first time.
• To challenge the generalization ability of existing low-light enhancement models.
• To mitigate catastrophic forgetting in models trained on sequential tasks with different lightness distributions.

Main Methods:

• A lightness enhancement network trained on a sequence of tasks with varying lightness distributions.
• Leveraging causal inference theory to construct a structural causal model.
• Employing backdoor adjustment to mitigate confounding variables and learn invariant representations.
• Utilizing a Rehearsal-based Invariant Structure Regularizer.
• Proposing a Channel Fourier Transform-based Self-Attention module.

Main Results:

• The proposed method effectively alleviates catastrophic forgetting in continual low-light enhancement.
• Demonstrated superior generalization performance on continuous low-light image enhancement benchmarks.
• The causal inference approach facilitates learning invariant representations across different lightness distributions.

Conclusions:

• The developed method significantly improves generalization for continual low-light image enhancement.
• Causal inference provides a theoretical framework to understand and address catastrophic forgetting.
• The novel attention module enhances the model's ability to process low-frequency information and estimate causal effects accurately.