Multi-Exposure Image Fusion via Deformable Self-Attention.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing multi-exposure image fusion (MEF) methods often use unidirectional alignment, neglecting global features and augmented location effects.
• Limited local region alignment in conventional methods leads to deficient global feature preservation.

Purpose of the Study:

• To propose a multi-scale bidirectional alignment network for adaptive image fusion.
• To address limitations of unidirectional alignment in MEF by incorporating global feature considerations.

Main Methods:

• Developed a novel deformable self-attention module for variant long-distance attention and interaction.
• Implemented bidirectional alignment to adaptively fuse differently exposed images towards normal exposure.
• Utilized multi-scale feature extraction for complementary scale features, fine details, and contextual information.
• Employed learnable weighted summation and predicted offsets for adaptive feature alignment and model generalization.

Main Results:

• The proposed network effectively aligns differently exposed images to a normal exposure degree.
• Multi-scale features provide complementary information, enhancing both fine details and contextual understanding.
• The deformable self-attention module facilitates adaptive feature alignment across various scenes.
• Extensive experiments show superior performance compared to state-of-the-art MEF methods.

Conclusions:

• The proposed multi-scale bidirectional alignment network offers an effective solution for adaptive image fusion.
• The novel deformable self-attention mechanism significantly improves feature alignment and fusion quality.
• This approach enhances global feature preservation and detail complementarity in multi-exposure image fusion.