Globally and Locally Semantic Colorization via Exemplar-Based Broad-GAN

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Exemplar-based image colorization transfers color from a reference to a target image.
• Challenges arise from semantic content differences between images.
• Existing methods struggle with semantic and structural consistency.

Purpose of the Study:

• To address limitations in exemplar-based image colorization caused by semantic discrepancies.
• To propose a novel method for generating natural-looking color images from grayscale targets using reference images.
• To improve both perceptual quality and semantic accuracy in image colorization.

Main Methods:

• Developed an exemplar-based conditional broad generative adversarial network (GAN).
• Employed a two-sub-network framework: match sub-net and colorization sub-net.
• Utilized dictionary-based sparse representation and global-local affinity energy for matching consistency.
• Integrated a broad learning system into the GAN for enhanced feature extraction and training stability.

Main Results:

• The proposed method achieves superior performance compared to state-of-the-art approaches.
• Demonstrated significant improvements in perceptual quality and semantic similarity.
• The integration of a broad learning system enhanced training stability and semantic accuracy.

Conclusions:

• The exemplar-based conditional broad-GAN offers a robust solution for image colorization challenges.
• The method effectively handles semantic differences between target and reference images.
• This approach advances the field of image colorization with improved accuracy and visual fidelity.