Comparative Analysis of Machine Learning Algorithms for Image Classification Accuracy

Thesis Abstract

Machine learning algorithms have become integral to image classification systems across various domains, including medical diagnostics, autonomous vehicles, and multimedia retrieval. Despite the proliferation of diverse algorithms such as Convolutional Neural Networks (CNNs), Support Vector Machines (SVM), Random Forests (RF), and k-Nearest Neighbors (k-NN), there is a limited comprehensive comparative analysis regarding their classification accuracy, robustness, and computational efficiency in different application contexts. This study aims to systematically evaluate and compare the performance of these prominent machine learning algorithms in image classification tasks, with a focus on identifying their relative strengths, limitations, and optimal deployment scenarios. The primary objectives of this research are to (1) assess the classification accuracy of CNN, SVM, RF, and k-NN algorithms on standardized image datasets; (2) analyze the influence of image complexity and size on algorithm performance; (3) evaluate computational efficiency in terms of training and inference times; and (4) explore the robustness of each algorithm against common image distortions such as noise, occlusion, and rotation. The research adopts a quantitative, experimental research design employing a cross-sectional analysis framework. The study population comprises publicly available, high-quality image datasets, including the CIFAR-10, MNIST, and ImageNet subsets, totaling over 100,000 images across multiple categories. A stratified sampling technique is used to select representative subsets for training and testing, with a sample size of 10,000 images per dataset to ensure statistical validity. Data collection involves preprocessing steps such as normalization, augmentation, and feature extraction (where applicable), utilizing established image processing libraries. The algorithms are implemented using Python-based machine learning frameworks, including TensorFlow for CNN, scikit-learn for SVM and RF, and custom k-NN implementations. Data analysis involves descriptive statistics to summarize model performance metrics, including accuracy, precision, recall, and F1-score, analyzed through Analysis of Variance (ANOVA) to determine statistically significant differences among algorithms. Post-hoc tests such as Tukey’s Honest Significant Difference (HSD) are conducted to identify specific pairs with performance disparities. Additionally, computational efficiency is measured through time profiling, and robustness assessments involve applying controlled distortions such as Gaussian noise and affine transformations, followed by repeated accuracy measurements to evaluate stability. The anticipated findings suggest that CNNs outperform traditional algorithms like SVM, RF, and k-NN in accuracy, especially on complex, high-resolution datasets, while traditional algorithms demonstrate faster training times and less computational resource consumption on simpler images. The study expects to reveal that the choice of algorithm should be context-dependent, considering factors such as dataset complexity, available computational resources, and required robustness to distortions. These results will contribute to the theoretical understanding of the comparative performance of machine learning algorithms in image classification and provide empirically grounded guidelines for practitioners selecting appropriate models. The research will also extend existing frameworks by integrating performance and efficiency metrics into a holistic comparison model, grounded in the Theory of Pattern Recognition and the Computational Cost Model. The main conclusion emphasizes that no single algorithm universally outperforms others; instead, optimal selection hinges on specific application requirements. It is recommended that future work explore hybrid models combining strengths of different algorithms and incorporate emerging deep learning architectures such as Vision Transformers. The study aims to fill gaps in existing literature by offering a comprehensive, empirically validated assessment that informs both academic inquiry and practical implementation of machine learning in image classification.