Enhanced uncertainty sampling with category information for improved active learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traditional uncertainty sampling methods in active learning often neglect category information.
• This leads to imbalanced sample selection in multi-class computer vision tasks, hindering model performance and generalizability.

Purpose of the Study:

• To develop a novel active learning framework that integrates category information with uncertainty sampling.
• To address the limitation of imbalanced sample selection in multi-class computer vision tasks.
• To improve the efficiency and representativeness of data annotation.

Main Methods:

• Employed a pre-trained VGG16 architecture for efficient category feature extraction.
• Utilized cosine similarity metrics to capture category information without additional model training.
• Combined category features with traditional uncertainty measures for balanced sampling.

Main Results:

• Achieved competitive mean average precision (mAP) scores in object detection with balanced category representation.
• Attained accuracy comparable to state-of-the-art methods in image classification.
• Reduced computational overhead by up to 80% in image classification tasks.

Conclusions:

• The proposed framework effectively balances sampling efficiency with dataset representativeness across various computer vision tasks.
• Offers a practical and efficient solution for large-scale data annotation, especially in domains with limited labeled data and diverse class distributions.