Multi-Scale Interactive Network with Color Attention for Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing

Background:

• Low-light image enhancement (LLIE) is vital for computer vision.
• Existing methods often fail to balance light enhancement and color correction.
• Images contain multi-level information crucial for effective enhancement.

Purpose of the Study:

• To propose a novel network, MSINet, for effective low-light image enhancement.
• To address the challenge of balancing light enhancement and color correction.
• To leverage multi-scale features for improved LLIE.

Main Methods:

• Proposed MSINet, a multi-scale interactive network with color attention.
• Employed a CNN-based branch with residual channel attention blocks (RCABs) for local features.
• Utilized a Transformer-based branch with cross-scale attention (CSA) and multi-head self-attention (MHSA) for global features.
• Integrated a color correction branch (CCB) with self-attention (SA) for color fidelity.

Main Results:

• MSINet effectively explores local and global image features through interaction modules.
• The CCB ensures color fidelity in the enhanced images.
• Extensive experiments show MSINet outperforms state-of-the-art LLIE methods.

Conclusions:

• MSINet achieves superior performance in both light enhancement and color correction for low-light images.
• The proposed architecture effectively integrates multi-scale feature extraction and interaction.
• MSINet offers a promising solution for challenging low-light image enhancement tasks.