A Comprehensive Study on Cat-Dog Image Classification Method Based on MobileNetV2 Transfer Learning

Deep learning has revolutionized the field of image classification, with Convolutional Neural Networks (CNNs) achieving remarkable success across various computer vision tasks. However, the deployment of deep learning models in resource-constrained environments remains challenging due to their substantial computational requirements. This paper presents an extensive investigation into the application of MobileNetV2, a lightweight CNN architecture, combined with transfer learning for cat-dog image classification. We conduct comprehensive experiments using the Kaggle Cats vs Dogs dataset to evaluate MobileNetV2’s performance in terms of accuracy, computational efficiency, and training dynamics. Our results demonstrate that MobileNetV2 achieves an exceptional classification accuracy of 98.08% on the test set while maintaining significantly lower computational complexity compared to traditional CNN architectures like VGG16 and ResNet50. The model’s efficiency is further evidenced by its rapid convergence during training and minimal parameter count of approximately 3.5 million. We provide detailed analysis of MobileNetV2’s architectural innovations, including inverted residual structures and linear bottlenecks, explaining how these design choices contribute to its superior performance. Additionally, we explore various transfer learning strategies and their impact on model performance. In addition, we selected the official Kaggle ’Dog Breed Identification’ dataset, which comprises 10,222 images across 13 common dog breeds, achieving a final test accuracy of 72.99%. This research not only validates MobileNetV2’s effectiveness for image classification tasks but also provides insights into the practical deployment of efficient deep learning models in real-world applications with limited computational resources.