A computer vision driven ecosystem for cattle monitoring: Multi-disease classification with severity grading, multi-view individual identification, and weight estimation

Abstract

Effective cattle monitoring and management are essential for sustainable dairy and beef production in Bangladesh, where constant veterinary care and extensive monitoring are economically challenging. This thesis introduces a computer vision based cattle monitoring ecosystem that create a (i) multi-disease classification and severity grading, (ii) multi-view unique cattle identification, and (iii) body weight estimation with the use of four view RGB images, based on a custom-collected, curated dataset with publicly available sources. To analyze the disease, a hierarchical deep learning architecture is suggested to categorize Lumpy Skin Disease (LSD), Foot-and-Mouth Disease (FMD), Infectious Bovine Keratoconjunctivitis (IBK), and Healthy cattle along with grading the diseased cattle as Stage-1 (mild), Stage-2 (moderate), and Stage-3 (severe) cases. The severity grading is performed based on a clinically validated symptom profile and is supervised and confirmed by a District Livestock Officer (veterinarian). The cross-attentional multi-task architecture has the highest accuracy score of 89.96% in disease classification, 83.75% in severity staging, and 85.88% in hierarchical setup (predicts disease and then predicts severity). In the case of individual cattle recognition and weight estimation, a multi-view appearance-based recognition system is constructed based on left, right, front, and back views of cattle images where at first YOLOv8s is trained to extract the exact cattle region which have an IoU of 0.93, map@0.5 0.97. The accuracy of the identification system is determined at Rank-1: 96.40% and 74.56% while Rank-5: 100% and 89.80% for ConvNext-Tiny on two different protocols of Leave One View Out testing method and Cross-view-angle respectively which has been done to evaluate cattle identification from different angles or possess and consider the case of cattle with very similar patterns and colors. To estimate body weight, a regression model is used with multi-view RGB images and metadata that includes id, sex, breed, age and live weight where the best single-model baseline is the DN121-tuned regressor, achieving an MAE of 36.99 KG which improves to an MAE of 35.10 KG with ensemble strategy. In general, the suggested system proves that even without the use of invasive sensors, an RGB-based, low shot, non-invasive, and affordable vision-based system can be used successfully in the diagnosis of diseases and their severity grading, the unique identification of individuals among cattle, and the estimation of their live weight.