IllumiSIFT: A Cascade Framework for DoG Pyramid Learning in Darkness.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Low light conditions severely degrade image quality, impacting critical applications like navigation and surveillance.
• Existing deep learning methods enhance pixel values but often fail to preserve essential gradient structures for feature detection.
• Reliable feature detection is crucial for downstream machine vision tasks.

Purpose of the Study:

• To develop a task-driven dark image enhancement framework, IllumiSIFT, specifically designed to preserve Scale-Invariant Feature Transform (SIFT) key points.
• To improve machine vision performance in low-light conditions by focusing on gradient-domain enhancement.

Main Methods:

• IllumiSIFT directly learns the Difference-of-Gaussian (DoG) pyramid from low-light inputs.
• A cascaded residual learning architecture predicts multi-scale Gaussian-blurred representations.
• The framework generates enhanced DoG images aligned with the SIFT detection process.

Main Results:

• IllumiSIFT consistently outperforms state-of-the-art enhancement methods in SIFT matching performance under severe low-light conditions.
• Experiments on CDVS, Oxford Buildings, and Paris datasets validate the approach.
• The method demonstrates superior preservation of SIFT key points compared to pixel-level enhancement.

Conclusions:

• Gradient-domain, task-aligned enhancement is more effective for recognition-centric low-light imaging.
• IllumiSIFT offers a practical solution for improving feature detection in challenging lighting.
• The proposed method enhances machine vision capabilities in low-light surveillance and retrieval applications.