CUI-Net: a correcting uneven illumination net for low-light image enhancement

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Real-world images frequently exhibit uneven lighting, with both dark and overexposed regions.
• Conventional low-light enhancement methods often worsen overexposed areas, degrading overall image quality.

Purpose of the Study:

• To develop an image enhancement technique that selectively improves low-light areas while preserving high-light regions.
• To introduce a robust and practical solution for uneven illumination correction in diverse imaging scenarios.

Main Methods:

• Proposed the correcting uneven illumination network (CUI-Net), a hybrid architecture combining Convolutional Neural Networks (CNNs) and Transformers with sparse attention.
• Implemented a two-module design: an enhancement module for feature extraction and an auxiliary module for training convergence.
• Utilized zero-shot learning to enable adaptation to various lighting conditions without paired training data.

Main Results:

• CUI-Net demonstrated stable and robust performance across multiple diverse datasets.
• The network effectively enhanced low-light regions while constraining high-light features, improving visual quality.
• Evaluated on object detection, face detection, and semantic segmentation tasks, showcasing practical advantages over existing methods.

Conclusions:

• CUI-Net successfully addresses the challenge of uneven illumination by enabling differentiated enhancement of image regions.
• The proposed method offers a practical and effective solution for low-light image enhancement, applicable to various computer vision tasks.