Comparative Analysis of Deep Learning Models for Image Classification Accuracy

Thesis Abstract

The rapid advancement of deep learning has significantly transformed the field of image classification, yet the relative efficacy of various models under differing conditions remains inadequately understood, posing a challenge for practitioners seeking optimal solutions for specific applications. This study aims to conduct a comprehensive comparative analysis of prominent deep learning models, including Convolutional Neural Networks (CNNs), Residual Networks (ResNets), DenseNet, and Inception-v3, with a focus on their classification accuracy, computational efficiency, and robustness across varied datasets. The specific objectives are to evaluate and compare the performance of these models using the ImageNet, CIFAR-10, and MNIST datasets, identify the optimal models for different image classification tasks, and analyze the impact of hyperparameter tuning on model performance. The research adopts a quantitative, experimental research design involving a systematic evaluation of four selected deep learning architectures. The population comprises open-source deep learning models implemented within TensorFlow and PyTorch frameworks. A purposive sampling technique was applied to select the models based on their prevalence and reported performance in the literature. The sample size includes three pre-trained models for each architecture, resulting in a total of 12 models evaluated across three datasets. Data collection was performed through the utilization of these datasets, which were partitioned into training, validation, and test subsets following established protocols. Model training was conducted under controlled computational environments, ensuring consistency in hardware (NVIDIA Tesla V100 GPU) and software configurations. The study employed transfer learning techniques to fine-tune models for each dataset, with hyperparameters such as learning rate, batch size, and number of epochs systematically optimized through grid search. Data analysis involved statistical comparison of classification accuracy, precision, recall, and F1-score using repeated measures ANOVA to determine significant differences among the models. Post hoc tests were conducted to identify specific pairs of models with statistically significant performance variations. Additionally, computational efficiency was assessed via training time and model parameter count, while robustness was measured by performance under added noise and occlusion conditions. The study further incorporated model interpretability analysis using Grad-CAM visualizations to understand decision-making processes. The research is underpinned by the Representation Learning Theory, which emphasizes how models learn and abstract features from images, and the Bias-Variance Tradeoff framework, guiding the hyperparameter tuning process. Expected findings include notable performance disparities among the evaluated models, with ResNets potentially demonstrating superior accuracy and robustness, and Inception-v3 offering better computational efficiency. It is anticipated that hyperparameter optimization will significantly enhance model performance, especially for deeper networks. The study aims to delineate clear guidelines for selecting appropriate deep learning architectures tailored to specific image classification contexts, considering accuracy, efficiency, and robustness. The contribution to knowledge lies in providing a systematic, empirical comparison of state-of-the-art deep learning models across multiple datasets, thereby informing best practices for researchers and practitioners in computer vision. The findings will advance understanding of the trade-offs associated with different architectures and hyperparameters, and contribute to theoretical frameworks concerning model generalizability and interpretability. The main conclusion underscores that no single model is universally optimal; instead, their effectiveness depends on the specific application requirements and environmental constraints. Recommendations include adopting hybrid strategies combining models, further exploration of model interpretability techniques, and extending the analysis to include emerging architectures such as Vision Transformers. Suggestions for future research involve longitudinal studies on model scalability and adaptation in real-world deployment scenarios, as well as exploring unsupervised and semi-supervised learning approaches within the context of image classification accuracy.