Convolutional Neural Networks (CNNs) have become fundamental to modern computer vision, delivering state-of-the-art results in a wide range of image classification tasks. This study investigates the binary classification of dog and cat images using custom-designed CNN architectures, trained entirely from scratch without leveraging pretrained weights. Three distinct models are developed, varying in depth, convolutional block structure, and regularization techniques, to systematically evaluate the impact of architectural design choices on classification performance. Preprocessing strategies, including data augmentation and pixel value normalization, are employed to enhance model robustness and prevent overfitting. Extensive experimental evaluation on the Dogs vs. Cats dataset demonstrates that deeper architectures, when combined with batch normalization and dropout, achieve superior generalization and convergence behavior. The best-performing model attains a validation accuracy of approximately \(92\%\) , confirming the effectiveness of the proposed approach. These findings underscore the importance of thoughtful CNN design and regularization, offering valuable insights for future applications in automated animal recognition and real-world visual classification systems.