Lightweight Deep Exemplar Colorization via Semantic Attention-Guided Laplacian Pyramid.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Exemplar-based colorization relies on reference images for plausible color generation.
• Existing methods struggle with semantic correspondence, object-background color confusion, and feature fusion.
• Simple architectures often lead to loss of high-frequency information or high computational complexity.

Purpose of the Study:

• To propose a lightweight semantic attention-guided Laplacian pyramid network (SAGLP-Net) for deep exemplar-based colorization.
• To address challenges in semantic correspondence and feature fusion in existing colorization methods.
• To exploit multi-scale color properties and semantic guidance for improved colorization.

Main Methods:

• Developed a lightweight semantic attention-guided Laplacian pyramid network (SAGLP-Net).
• Introduced a semantic guided non-local attention fusion module for long-range dependency and feature fusion.
• Utilized a Laplacian pyramid fusion module with criss-cross attention for high-frequency component fusion.
• Employed an unsupervised multi-scale multi-loss training strategy including pixel loss, color histogram loss, total variance regularization, and adversarial loss.

Main Results:

• The SAGLP-Net effectively aligns object and background information using semantic guidance.
• The network successfully fuses local and global features, preserving high-frequency details.
• Experimental results show superior subjective and objective performance compared to state-of-the-art methods.
• The proposed method achieves better colorization results with lower computational complexity.

Conclusions:

• The SAGLP-Net offers an effective and efficient solution for deep exemplar-based image colorization.
• Leveraging multi-scale properties and semantic attention significantly enhances colorization quality.
• The proposed fusion modules and training strategy contribute to robust performance and reduced complexity.