Empirical Analysis of the Performance of Machine Learning Algorithms in Classifying 2D MR Images from PCA Reduced HOG and LBP Features
This study investigates the role of feature extraction and dimensionality reduction techniques in addressing high-dimensional image data, with a particular focus on Alzheimer’s disease classification using 2D magnetic resonance imaging (MRI). Histogram of Oriented Gradients (HOG) and Local Binary Pa...
| Authors: | , , , , |
|---|---|
| Format: | article |
| Status: | Published version |
| Publication Date: | 2025 |
| Country: | España |
| Institution: | Universidad de Sevilla (US) |
| Repository: | idUS. Depósito de Investigación de la Universidad de Sevilla |
| OAI Identifier: | oai:dnet:idus________::50bf8d220b0bd23a070b41474eddb1ed |
| Online Access: | https://hdl.handle.net/11441/186313 http://dx.doi.org/10.62762/BISH.2025.993395 |
| Access Level: | Open access |
| Keyword: | medical imaging feature extraction ensemble techniques histogram of oriented gradients local binary patterns PCA. |
| Summary: | This study investigates the role of feature extraction and dimensionality reduction techniques in addressing high-dimensional image data, with a particular focus on Alzheimer’s disease classification using 2D magnetic resonance imaging (MRI). Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) are employed to extract discriminative features from MRI images; however, due to the high dimensionality of the extracted features, dimensionality reduction is required. Principal Component Analysis (PCA) is utilized to reduce feature dimensionality while preserving most of the relevant information, as reflected in the improved performance of the underlying machine learning (ML) classifiers. Two feature extraction pipelines are evaluated: (i) HOG combined with PCA, and (ii) LBP combined with PCA. The reduced feature sets are subsequently used for classification. Experimental results demonstrate that ML algorithms consistently achieve superior performance using features derived from the HOG+PCA pipeline compared to those obtained from the LBP+PCA pipeline. Although the LBP+PCA approach exhibits certain advantages, HOG+PCA proves to be more effective for the problem under consideration, while acknowledging that performance may vary across applications. Furthermore, the study confirms that ensemble learning methods generally outperform individual classifiers by leveraging complementary strengths, and that larger datasets tend to enhance model performance by enabling the learning of richer patterns. In contrast, memory-intensive algorithms such as k-nearest neighbors (KNN) may be suitable for smaller datasets but are typically less scalable for large-scale applications. |
|---|