Optimizing CNN Architectures for Cat and Dog Classification with Focus on Generalization and Robustness

The accurate and efficient classification of images remains a fundamental challenge in computer vision. This paper presents an empirical study of Convolutional Neural Networks (CNNs) tailored for cat versus dog image classification, focusing on enhancing model performance through innovative architecture design, data preprocessing, and training strategies. We investigate the impact of hyperparameters, advanced data augmentation techniques, and transfer learning using pre-trained models like VGG16 and ResNet50 on classification accuracy. Our novel custom CNN architecture, featuring a unique combination of asymmetric filter sizes and adaptive dropout rates, combined with a tailored preprocessing pipeline incorporating histogram equalization, achieves competitive performance. The models are evaluated on the Kaggle "Cats vs. Dogs" dataset and the Oxford-IIIT Pet Dataset, achieving high accuracy, precision, recall, and F1-scores. Transfer learning significantly boosts performance, with fine-tuned ResNet50 achieving 92.8% accuracy on Kaggle and 90.1% on Oxford-IIIT, underscoring its efficacy for faster convergence and improved generalization. We analyze learned features to identify model strengths and limitations, particularly in handling variations in pose, lighting, and breed. The study also quantifies computational efficiency, reporting model size and inference time to assess edge deployment feasibility. CNNs’ potential for reliable image classification is highlighted, while addressing challenges like overfitting and adversarial vulnerabilities. Future research directions include integrating Vision Transformers, adversarial training, and model compression for edge devices, building on our findings to enhance robustness and scalability.