Deep Learning Approaches for Indian Cattle and Buffalo Breed Classification: A Comparative Analysis of CNN Architectures

Abstract

Correctly identifying cattle and buffalo breeds directly affects herd management decisions, selective breeding outcomes, and dairy-sector efficiency across India. This paper benchmarks five modern CNN architectures against one another for the task of automated breed classification. A custom dataset comprising 3,890 high- resolution images across 13 Indian cattle and buffalo breeds was curated through web scraping, YOLOv8-based filtering, and manual refinement. Five transfer learning models—ResNet50V2, InceptionResNetV2, MobileNetV2, ConvNeXtTiny, and EfficientNet-B0 were evaluated using a two- phase training strategy combining classifier warm-up and fine-tuning. EfficientNet-B0 outperformed the other four models, reaching 79.89% on the held-out test set and 82.11% on validation. The study reveals that compound scaling and adaptive feature extraction provide superior performance for phenotypically similar breeds, while parameter-efficient architectures like MobileNetV2 and ConvNeXtTiny offer deployment advantages with fewer than 500,000 trainable parameters. Key Words Convolutional Neural Networks, Transfer Learning, Cattle Breed Classification, Deep Learning, Indian Livestock, Image Recognition