Development of a lightweight deep learning model for the identification and classification of Indigenous cattle breeds

Abstract

This study aimed to develop a lightweight deep learning model for the identification and classification of Tharparkar and Hariana cattle breeds as they are phenotypically similar-looking and have subtle differences in visual appearance. Images were collected from 115 cows of each breed under natural conditions. A CNN-based semantic segmentation model was developed to accurately identify the cow as a Region of Interest in the given image. The IoU value of 84.15% and F1-Score of 87 % of the segmentation model for the cow region suggested that the model was capable in segmenting the cow pixels. The masked image as output from the segmentation model was used as input for the final breed classifier model. The recall value of 86 % and precision value of 88 % of the segmentation model for the cow region indicated that the model effectively identified cow regions with high accuracy, minimizing false positives. The model requires approximately 618 ms and 3.27 million parameters to perform segmentation for one image. The accuracy of the classification model for the Tharparkar and Hariana class was found to be 72.5%. Precision, recall value, and F1-Score for the Hariana breed were 73.7%, 70.0%, and 71.8% respectively. Whereas precision was 71.4%, recall value was 75.0%, and F1-Score was 73.2% for Tharparkar. This study attempted to differentiate white-coloured breeds using a deep learning method without the help of manual help and experts. Further research on robust datasets and fine-tuning of the model parameters may lead to better accuracy in breed classification.