Dataset for image classification with knowledge

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels in image classification with large datasets.
• Performance declines significantly with limited data.
• Fine-grained classification is challenging for deep architectures.

Purpose of the Study:

• To address the limitations of deep learning in low-data and fine-grained image classification scenarios.
• To introduce novel datasets that integrate a priori knowledge.
• To facilitate research on leveraging prior knowledge in image classification.

Main Methods:

• Datasets were constructed from existing multilabel, multiclass classification, or object detection data.
• Frequent closed itemset mining was employed to generate classes and attributes.
• A priori knowledge was extracted in the form of rules based on these attributes.

Main Results:

• The developed datasets integrate a priori knowledge, enhancing image classification capabilities.
• The methodology enables the creation of structured knowledge from raw data.
• The rule generation algorithm is detailed for practical application.

Conclusions:

• Integrating a priori knowledge into datasets is crucial for improving deep learning performance in data-scarce and complex classification tasks.
• The proposed method offers a viable approach to generating such datasets.
• This work expands the available resources for research in knowledge-enhanced image classification.