Incorporating Deep Learning and Optimization Techniques with Data Augmentation for Improved Image Analysis and Classification

BOUDOUH, Nouara (2025) Incorporating Deep Learning and Optimization Techniques with Data Augmentation for Improved Image Analysis and Classification. Doctoral thesis, Université Mohamed Khider (Biskra - Algérie).

Text Thesis.pdf Download (28MB)

Abstract

Deep learning methods often face challenges due to unbalanced or non-representative data, and in many cases, data scarcity limits model effectiveness. We advocate that improving data quality can lead to significant performance enhancements. This thesis presents new methods for data augmentation. Our first method involves randomly create filters to remove certain rows and columns from the original image to generate smaller, more informative images. This method was applied to the Cats vs. Dogs dataset to train the Basic CNN and ResNet50 models, showing improved results compared to the original dataset. However, random filter generation can sometimes produce images that are too similar to the originals, reducing diversity. To address this, we developed a secondary technique incorporating a random optimization algorithm to select optimal generated images based on entropy, yielding promising results when applied to the VGG16 model. Nevertheless, image selection remains dependent on filter quality, potentially limiting diversity. Therefore, our third method employs a genetic algorithm to enhance filter generation and ensure greater diversity. Additionally, we improved the architectures of the VGG16 and VGG19 models. When applied to the Cats vs. Dogs and Chest X-ray datasets and used to train a set of seven models (VGG16, VGG19, their enhanced versions, EfficientNet-B0, Inception-V3, and Vision Transformer), we observed promising improvements in model performance compared to the second method. Since optimization techniques require considerable time and resources, we proposed an alternative method to enhance model performance without increasing data size. This approach leverages the unique capabilities of each model to extract features by merging their outputs into a unified representation used to train a single classifier. The integrated models using VGG16, VGG19, EfficientNet-B0, and Inception-V3 showed clear performance superiority compared to each model’s individual performance.

Actions (login required)

View Item