Utilizing Vision Transformer-Based Analysis of Fecal Images for Poultry Disease Diagnosis

Abstract

In the poultry business, being able to detect and manage diseases as early as possible is crucial for ensuring productivity and minimizing economic losses. While large poultry farms can afford to have veterinary doctors on staff, small and medium scale farmers do not have access to specialized diagnostics. This shortcoming highlights the need for efficient technologies that allow for timely intervention and disease management. Our research proposes novel methodology for accurate and rapid disease diagnosis in poultry using deep learning. Our diagnostic model employs advanced deep learning and image processing techniques for automated poultry disease detection. The analysis of fecal images was done using a variety of convolutional neural networks (CNNs) GoogleNet, ResNet18, ShuffleNet, SqueezeNet and Vision Transformer with a maximal test accuracy of 97.62%. The dataset used contained more than 500,000 images and incorporated many real-life problems such as background noise, noise in the image, and different lighting intensities. The model proposed in this paper successfully detects Coccidiosis, Newcastle Disease, Salmonella and other poultry disorders from fecal images. This research provides a linkage between advanced AI-based diagnostic tools and poultry farmers’ needs to improve disease management, biosecurity and sustainability in the poultry industry. The Poultry Pathology Visual Dataset, used in this project, can be accessed by anyone at Kaggle (kaggle.com).