FM-DLM: A new method for image classification based on the fusion of multi-level deep learning models

Area of Science:

• Computer Science
• Artificial Intelligence

Background:

• Deep learning models face deployment challenges due to size and label limitations.
• Large models offer broad classification but are unsuitable for small devices.
• Small models fit devices but have restricted label capacities.

Purpose of the Study:

• To propose a novel classification method addressing limitations of existing deep learning models.
• To enhance classification accuracy and range, particularly for resource-constrained environments.

Main Methods:

• A Fusion of Multi-level Deep Learning Models (FM-DLM) approach is presented.
• Utilizes a Level 0 model (Baidu-AI) for wide-range sample classification.
• Employs Level 1 model differences for dataset prediction and Level 2 models for label classification.
• Leverages label distribution for improved accuracy.

Main Results:

• The FM-DLM method demonstrates superior accuracy compared to existing techniques.
• Achieves a wide range of classification capabilities.
• Successfully balances model size and classification performance.

Conclusions:

• The proposed FM-DLM method effectively overcomes the trade-offs between model size and classification range.
• Offers a viable solution for deploying advanced classification on diverse devices.
• Represents a significant advancement in multi-level deep learning for classification tasks.