Broad Colorization.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Traditional image colorization methods, such as scribble- and example-based approaches, demand significant user input and extensive training times.
• Deep neural networks (DNNs) commonly used for colorization are computationally expensive and time-consuming to train.

Purpose of the Study:

• To develop an automatic image colorization approach that eliminates user dependence and reduces training time.
• To enhance the efficiency and accessibility of image colorization techniques.

Main Methods:

• A novel framework combining local and global features from grayscale images for automatic colorization.
• Utilizing a local broad learning system to extract pixel chrominance values from low-, mid-, and high-level features.
• Employing a global broad learning system to refine the generated chrominance map, guided by global image features.

Main Results:

• The proposed method achieves automatic colorization without requiring user input.
• Demonstrates a significant reduction in training time, an order of magnitude faster than traditional DNN-based methods.
• Allows for user-driven data augmentation without full system retraining, enhancing user initiative.

Conclusions:

• The developed automatic colorization approach offers a faster and more user-friendly alternative to existing methods.
• The combination of local and global features, along with broad learning systems, proves effective for efficient image colorization.
• Experimental results validate the superiority of this approach over state-of-the-art colorization techniques.