Colorization Using Neural Network Ensemble.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Image colorization converts grayscale images to color, a task traditionally requiring manual intervention.
• Existing learning-based methods often struggle with diverse color styles using a single neural network.

Purpose of the Study:

• To develop an improved automatic image colorization method.
• To address the challenge of diverse color styles in image datasets.
• To achieve artifact-free, high-quality colorization results.

Main Methods:

• Proposed a mixture learning model to capture sub-color styles within datasets.
• Employed an ensemble of multiple neural networks for enhanced color estimation.
• Utilized a two-step strategy: adaptive color style clustering followed by neural network ensembling.
• Implemented a joint bilateral filtering post-processing step for artifact reduction.

Main Results:

• The proposed method demonstrates superior color estimation performance compared to individual neural networks.
• Achieved high-quality colorization results comparable to state-of-the-art algorithms.
• Successfully generated artifact-free outputs through the proposed post-processing technique.

Conclusions:

• Mixture learning and neural network ensembling effectively handle diverse color styles in image colorization.
• The proposed two-step strategy with joint bilateral filtering significantly improves colorization quality.
• This method offers a robust solution for automatic, high-fidelity image colorization.