Enhancing CNN Performance for Image Classification

This study explores various architectural modifications and hyperparameter tuning strategies to enhance the performance of Convolutional Neural Networks (CNNs) in image classification. We systematically analyze the effects of varying convolutional kernel sizes, dropout layers, and fully connected layer configurations. Our experiments demonstrate that while dropout improves generalization, increasing convolutional layers can lead to overfitting. Comparisons with alternative approaches, including the Bag of Visual Words (BoW) model and a Multi-Layer Perceptron (MLP), reveal that BoW achieves the highest accuracy but at the cost of increased computational complexity, while MLP exhibits significant overfitting. The findings highlight the trade-offs between network depth, regularization techniques, and computational efficiency. Future improvements may focus on data augmentation, refined region-of-interest segmentation, and leveraging pre-trained models like VGG16 for fine-tuning. This study provides insights into optimizing CNN architectures for improved image classification performance.