HPCDNet: Hybrid position coding and dual-frquency domain transform network for low-light image enhancement

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement (LLIE) aims to restore visibility in images captured under poor illumination.
• Existing LLIE methods often fail to fully leverage spatial positional and frequency domain information.
• This limits their ability to accurately reconstruct details and improve visual quality.

Purpose of the Study:

• To propose an end-to-end network, HPCDNet, for effective low-light image enhancement.
• To integrate hybrid positional coding and frequency domain recovery into a unified framework.
• To improve the visual quality, detail preservation, and texture fidelity of enhanced low-light images.

Main Methods:

• HPCDNet employs a hybrid positional coding technique within its self-attention mechanism to retain spatial information.
• Discrete wavelet and cosine transforms are used to recover lost frequency domain information.
• A dual-attention module adaptively weights and merges the recovered frequency domain features.

Main Results:

• The proposed HPCDNet significantly enhances visibility, contrast, and color properties of low-light images.
• The method demonstrates superior preservation of image details and textures compared to existing techniques.
• Experimental results validate the effectiveness of integrating positional and frequency domain information.

Conclusions:

• HPCDNet offers an advanced solution for low-light image enhancement by effectively utilizing both spatial and frequency domain features.
• The network's hybrid positional coding and dual-attention module contribute to improved image reconstruction.
• This approach advances the state-of-the-art in low-light image enhancement, yielding more natural and visually appealing results.