Unsupervised Illumination Adaptation for Low-Light Vision.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Insufficient lighting hinders both human and machine visual analysis.
• Current low-light enhancement methods often overlook machine vision requirements and semantic understanding.

Purpose of the Study:

• To develop a novel illumination enhancement model specifically for high-level machine vision tasks.
• To address the limitations of existing methods that prioritize human perception over machine analytics.

Main Methods:

• Developed a lightweight illumination enhancement model inspired by camera response functions.
• Introduced two training approaches using base enhancement curves and self-supervised pretext tasks for normal-to-low-light adaptation.
• The framework operates without requiring labeled low-light data.

Main Results:

• The proposed model enhances images from a machine vision perspective.
• Achieved effective illumination restoration and feature alignment.
• Significantly improved performance in downstream tasks like classification, face detection, and action recognition.

Conclusions:

• The research pioneers a machine-centric approach to low-light image enhancement.
• The plug-and-play framework offers a versatile solution for diverse high-level vision applications.
• Advances the field of low-light machine analytics by overcoming existing algorithmic limitations.