Multi-task feature integration and interactive active learning for scene image resizing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image retargeting is crucial for adapting complex scenes across different displays.
• Existing methods struggle with intricate spatial configurations and capturing human perception.

Purpose of the Study:

• To develop an AI-driven adaptive image retargeting method using multi-channel visual features.
• To precisely capture human gaze dynamics for improved scene understanding and recomposition.

Main Methods:

• Utilized BING objectness metric for semantic and visual patch acquisition.
• Introduced a multi-task feature selector for integrating disparate scene patch features.
• Developed locality-preserved and interactive active learning (LIAL) to generate gaze shift paths (GSP).
• Employed a multi-layer aggregating algorithm for deep feature calculation and Gaussian Mixture Model (GMM) encoding.

Main Results:

• The proposed method significantly outperforms five counterpart retargeting techniques.
• Achieved 3% higher precision than the second-best visual recognizer among 17 tested.
• Demonstrated reduced testing time consumption, achieving 49.8% of the second-best performer's time.

Conclusions:

• The developed AI methodology effectively achieves adaptive image retargeting for complex scenes.
• The LIAL strategy enhances scene coherence and incorporates human interaction for gaze path generation.
• The approach offers superior precision and efficiency compared to existing methods.