Active Learning Plus Deep Learning Can Establish Cost-Effective and Robust Model for Multichannel Image: A Case on...

Area of Science:

• Agricultural and biological engineering
• Computer vision
• Machine learning

Background:

• Deep learning excels in image classification but requires extensive labeled data, which is costly and labor-intensive in specialized fields like agricultural and biological engineering.
• Multichannel images (e.g., CT, MRI, HSI) are common in these domains, posing unique annotation challenges.
• Current annotation methods are time-consuming, expensive, and require domain-specific expertise.

Discussion:

• A novel framework combining active learning and deep learning is proposed for multichannel image classification.
• Three active learning strategies (least confidence, margin sampling, entropy) are employed for efficient data selection.
• An 'image pool' strategy leverages data augmentation to maximize the utility of generated images.

Key Insights:

• The proposed framework significantly reduces the manual labeling effort required for training deep learning models.
• Active learning with entropy selection and an image pool achieved accuracy comparable to full manual annotation using only a fraction of the data.
• This approach offers a practical solution for developing and updating models in multichannel image classification tasks.

Outlook:

• The framework demonstrates potential for broad application in multichannel image classification within agricultural and biological engineering.
• Future work could explore additional active learning algorithms and data augmentation techniques.
• Further validation across diverse multichannel image datasets is recommended to assess generalizability.