Proceedings of the 1st International Conference on Artificial Intelligence, Communication, IoT, Data Engineering and Security, IACIDS 2023, 23-25 November 2023, Lavasa, Pune, India

Research Article

Mitigating Training Bias in Cattle Breed Identification through Orientation-Aware Framework

Download114 downloads
  • @INPROCEEDINGS{10.4108/eai.23-11-2023.2343289,
        author={Vijayalakshmi  A and Shanmugavadivu  P and Vijayalakshmi  S and Shreyansh  Padarha and Sivaranjani  R},
        title={Mitigating Training Bias in Cattle Breed Identification through Orientation-Aware Framework},
        proceedings={Proceedings of the 1st International Conference on Artificial Intelligence, Communication, IoT, Data Engineering and Security, IACIDS 2023, 23-25 November 2023, Lavasa, Pune, India},
        publisher={EAI},
        proceedings_a={IACIDS},
        year={2024},
        month={3},
        keywords={orientation labelling data augmentation convolutional neural networks (cnn) cattle breed identification},
        doi={10.4108/eai.23-11-2023.2343289}
    }
    
  • Vijayalakshmi A
    Shanmugavadivu P
    Vijayalakshmi S
    Shreyansh Padarha
    Sivaranjani R
    Year: 2024
    Mitigating Training Bias in Cattle Breed Identification through Orientation-Aware Framework
    IACIDS
    EAI
    DOI: 10.4108/eai.23-11-2023.2343289
Vijayalakshmi A1,*, Shanmugavadivu P1, Vijayalakshmi S2, Shreyansh Padarha2, Sivaranjani R3
  • 1: Gandhigram Rural Institute (Deemed to be University), Dindigul
  • 2: Christ (Deemed to be University), Pune Lavasa Campus
  • 3: Veterinary College and Research Institute, Namakkal
*Contact email: vijiranjanis@gmail.com

Abstract

This paper introduces a comprehensive strategy to mitigate training bias in cattle breed identification models, focusing on cow orientation in still frames. Leveraging YOLOv7 for face detection, our methodology incorporates a novel orientation-aware pre-processing step that categorizes images into right-oriented, left-oriented, original, and original with inverted orientations. The breed identification model, based on Convolutional Neural Networks (CNN), is trained on this augmented dataset. While the left-oriented approach achieves the highest accuracy, the original and inverted orientation strategy demonstrates superior validation accuracy, showcasing its effectiveness in addressing bias during real-world applications with varied cow orientations. These findings underscore the importance of orientation-aware training for robust cattle breed identification, providing a solution for an overlooked niche