Pattern and structural detection in grayscale images through the application of quantile graphs in higher-dimensional...

Area of Science:

• Computer Vision
• Machine Learning
• Graph Theory

Background:

• Deep Learning (DL) and Machine Learning (ML) models, including Convolutional Neural Networks (CNNs) and Vision Transformers (VTs), face challenges with large datasets and extensive parameter tuning for image classification.
• Existing graph-based methods like Visibility Graphs (VGs) can be computationally intensive due to high node counts.
• Quantile Graphs (QGs) have shown success in time series analysis for pattern identification with reduced computational load.

Purpose of the Study:

• To extend the Quantile Graph (QG) framework from time series to two-dimensional image classification.
• To introduce a scalable, graph-based feature extraction method for ML and DL in computer vision.
• To evaluate the performance of QGs against established methods like CNNs and VTs, particularly in low-data scenarios.

Main Methods:

• Developed a novel method to transform 2D images into Quantile Graphs (QGs).
• Utilized benchmark datasets (MNIST, Fashion MNIST) for evaluating QG performance against CNNs and VTs.
• Applied the QG method to a medical imaging dataset to demonstrate relevance for disease detection.

Main Results:

• Quantile Graphs (QGs) demonstrated competitive performance, outperforming CNNs and VTs in scenarios with limited training data.
• QGs exhibited more consistent results across different training configurations compared to CNNs and VTs.
• The QG approach proved effective on a medical imaging dataset, highlighting potential for brain disease detection.

Conclusions:

• Extending Quantile Graphs (QGs) to image classification offers a computationally efficient and scalable alternative to traditional DL models, especially under data constraints.
• The QG framework provides a robust feature extraction technique applicable to diverse computer vision tasks and medical image analysis.
• This research introduces a valuable open-source tool for advancing graph-based methods in machine learning and artificial intelligence.