Chart Classification Using Siamese CNN.

Area of Science:

• Computer Science
• Machine Learning
• Data Visualization

Background:

• Chart type classification is crucial for information retrieval from visual data.
• Existing methods, like Support Vector Machines (SVM) and Convolutional Neural Networks (CNNs), struggle with limited or synthetic datasets.
• Real-world chart image datasets are scarce, hindering the development of robust classification models.

Purpose of the Study:

• To introduce and evaluate a Siamese Convolutional Neural Network (CNN) architecture for chart type classification.
• To address the challenge of small datasets in chart image classification.
• To compare the performance of Siamese CNNs against traditional CNNs using Few-shot learning (FSL).

Main Methods:

• Implementation and testing of multiple Siamese CNN architectures.
• Comparative analysis of performance across various dataset sizes.
• Verification using Few-shot learning (FSL) methodologies.
• Evaluation of Siamese CNNs' ability to classify with minimal data per class.

Main Results:

• Siamese CNNs demonstrate significant advantages over traditional CNNs, especially with limited data.
• Achieved 100% average classification accuracy with only 50 images per class.
• Traditional CNNs achieved only 43% average classification accuracy on the same dataset.
• Siamese CNNs proved effective even with a single image per class.

Conclusions:

• Siamese CNN architecture is a highly effective solution for chart type classification, particularly when dealing with small or imbalanced datasets.
• The proposed method significantly improves classification accuracy compared to conventional approaches in Few-shot learning scenarios.
• This research pioneers the use of Siamese CNNs for chart classification, offering a promising direction for future research and applications.