Comparative Deep Learning Models for Cow and Buffalo Classification Using CNN and Attention Mechanisms

This study set out to answer a straightforward question: can a deep learning model reliably tell a cow from a buffalo in an ordinary outdoor photograph, and which architecture does it best? We curated a 1,713-image dataset from two Kaggle sources covering eleven Indian breeds, cleaned it through a rigorous multi-stage pipeline, and trained five different architectures under identical conditions. The answer is clear. ResNet50 achieved the best performance with 99.22% accu- racy and 99.96% AUC-ROC, making it the strongest model for this task and practical enough for real deployment. The ablation study showed that adding CBAM attention on top of ResNet50 did not improve results — accuracy dropped slightly to 98.83% — confirming that deep pretrained features are already sufficient for this binary classification problem. Grad-CAM maps confirmed the model attends to horn shape, the dewlap region, and body silhouette rather than background elements, showing the learned representation is anatomically grounded. Three concrete outputs came out of this work. First, a cleaned and documented 1,713-image dataset for binary cow-vs-buffalo classification. Second, a controlled benchmark showing what five architectures achieve on this task. Third, a clear finding that transfer learning with a deep pretrained backbone is sufficient for this problem. The next steps include breed-level classification, video-based analysis, and deploy- ment on low-power edge devices.

[1]     Y. Pan, H. Jin, J. Gao, and H. T. Rauf, “Identification of Buffalo Breeds Using Self-Activated-Based Improved Convolutional Neural Networks,” Agriculture, vol. 12, p. 1386, 2022. Available at: https://doi.org/10.3390/ agriculture12091386 [2]     N. Naik, N. Jagadisha, and S. Ramu, “Buffalo Identification in Mixed- Species Environments: A Comparative Deep Learning Approach Using ResNet50 and EfficientNetB3,” Journal of Buffalo Science, vol. 14, 2025. Available at: https://doi.org/10.6000/1927-520X.2025.14.08 [3]     R. Warhade, I. Devi, N. Singh et al., “Attention Module Incorporated Transfer Learning for Classification of Phenotypically Similar Tropical Cattle Breeds (Bos indicus),” Tropical Animal Health and Production, 2024. Available at: https://doi.org/10.1007/s11250-024-04050-7 [4]     S. Li, L. Fu, Y. Sun, Y. Mu, and L. Chen, “Individual Dairy Cow Identification Based on Lightweight CNN with Attention Mechanism,” PLOS ONE, vol. 17, e0260510, 2022. Available at: https://doi.org/10. 1371/journal.pone.0260510 [5]     N. R. Duraiswami et al., “Cattle Breed Detection and Categoriza- tion Using Image Processing and Machine Learning (ResNet50 vs MobileNetV2),” in Proc. IEEE ASSIC, 2022. Available at: 10.1109/ ASSIC55218.2022.10088381 [6]     M. H. Kang and S. H. Oh, “Research Trends in Livestock Facial Identification: A Review of CNN-Based Deep Learning,” Journal of Animal Science and Technology, 2025. Available at: https://doi.org/10. 5187/jast.2025 [7]     X. Zhang et al., “CNN-Bi-LSTM: A Complex Environment-Oriented Cattle Behavior Classification Network,” Animals, vol. 13, p. 3280, 2023. Available at: https://doi.org/10.3390/ani13203280 [8]     X. Du, Y. Qi, J. Zhu, Y. Li, and L. Liu, “AI-Powered Cow Detection in Complex Farm Environments Using YOLOv8-CBAM,” Smart Agri- cultural Technology, 2025. Available at: https://doi.org/10.1016/j.atech. 2025.100xxx [9]     H. Chen, G. Zhou, W. He et al., “Classification and Identifica- tion of Agricultural Products Based on Improved MobileNetV2,” Sci- entific Reports, vol. 14, 2024. Available at: https://doi.org/10.1038/ s41598-024-53349-w [10]    Multiple Authors, “Transfer Learning in Agriculture: A Systematic Review (2022–2024),” Artificial Intelligence Review, 2025. Available at: https://doi.org/10.1007/s10462-024-11081-x