Scalable and automated Zebu cattle breed identification system using low-complexity AI models

Abstract

The objective of study was to investigate computationally efficient, low-complexity ML models for automated image-based identification of important milch zebu cattle breeds, i.e. Sahiwal, Red Sindhi, Hariana and Tharparkar. For each breed, images were captured from frontal and left-side views to ensure a comprehensive visual representation of significant features. The dataset comprised images across eight classes of each breed. While deep learning models generally offer high representational capacity, they require substantial computational resources, memory, and training time, which may limit their deployment in resource-constrained agricultural environments. So, four classification models were studied and compared: CNN, Multi-Layer Perceptron (MLP), Random Forest (RF), and K-Nearest Neighbour (KNN).  To enhance computational efficiency, dimensionality reduction techniques were applied where appropriate. The results demonstrated that light-weight CNN model, found as top performer, achieving 95% accuracy with high precision, recall, and F1-scores across eight breed-view classes, closely followed by 90% accuracy for MLP, RF and 88% accuracy for KNN, demonstrating robust feature handling post-PCA dimensionality reduction while maintaining low computational demands suitable for field deployment. The study indicated that low-complexity machine learning models can serve as efficient and deployable alternatives to deep neural networks for cattle breed classification, enabling scalable implementation in smart dairy monitoring systems