Smartphone based detection and classification of poultry diseases from chicken fecal images using deep learning techniques

Abstract

Human demand for animal products is increasing, forcing the agricultural industry, particularly poultry farming, to increase the quantity of its output. Increased poultry farming can lead to increased transmission of infectious diseases, resulting in widespread poultry death and significant economic losses. Traditional techniques for detecting diseases in poultry involve manual methods that are labor-intensive, time-consuming, and error-prone. Furthermore, the interpretation of the results often requires the expertise of trained professionals. These limitations can impede timely disease detection and increase the risk of the disease spreading throughout the flock, which can have severe consequences. This paper presents a detection and classification system for poultry diseases. The system was developed using two core algorithms: YOLO-V3 object detection algorithm and ResNet50 image classification model. YOLO-V3 was used to segment region of interest (ROI) from faecal images while ResNet50 was used for classification of the segmented image into four health conditions: Health, Coccidiosis, Salmonella, and New Castle Disease. The models were trained on 10,500 chicken faecal images collected from Zenodo open database. Oversampling and image augmentation techniques were applied to the dataset to handle class imbalance prior to training the ResNet50 model. The YOLO-V3 object detection model, implemented in Darknet, achieved a mean average precision of 87.48% for detecting regions of interest (ROI), while the ResNet50 image model demonstrated a classification accuracy of 98.7%. Based on our experimental findings, the proposed chicken disease detection and classification system exhibits the ability to accurately identify three prevalent poultry diseases. Therefore, this system can prove to be a valuable tool for assisting poultry farmers and veterinarians in farm settings.