Object Detection Model for Poultry Diseases Diagnostics

Poultry diseases lead to low productivity, economic losses, and high veterinary expenses for smallholder farmers in Tanzania. Since most poultry farmers rely on professionals to identify the diseases, they usually face a challenge of getting timely responses whenever an incident occurs due to the limited number of veterinary officers. In this study, we developed a deep learning model for object detection tasks using YOLOv5 architecture to accurately and fast predict diseases in poultry using their fecal images. The model was trained to detect three poultry diseases namely Salmonella, Coccidiosis, and Newcastle disease, and the healthy class. An open-source dataset of 1117 fecal images annotated for object detection task in JSON format was used for training the model. The annotation heatmap across all classes (Salmonella, Coccidiosis, Newcastle, and healthy) indicated a good representation of all four classes of the dataset. Model performance during training was evaluated based on the following metrics: precision, recall, and mean average precision with IOU (Intersection Over Union) of 0.5 (mAP_0.5). The best training results were obtained at epoch 506. The model was deployed on a web application using Streamlit framework. The application is suitable for the timely detection of three diseases namely Salmonella, Coccidiosis, and Newcastle disease, and the healthy class. The object detection task indicates where in the fecal image the disease-causing sample is located. The object detection task improves the model explainability for target users which in this case are smallholder poultry farmers and agricultural experts.