Poultry Disease Identification In Fecal Images Using Vision Transformer

Abstract

Early diagnosis and control of diseases assume critical importance in the poultry farming industry. While the large sized scale poultry farms may be able to afford in-house veterinary support, the small and medium-sized operations may not be able to have theirown diagnosis services, making it all the more critical for such farms to make use of the emerging technologies to assist them in the early diagnosis and control of their poultry farms. Prompt and accurate identification of poultry diseases without delay assumes paramount importance, as the delays can be devastating, resulting in significant economic losses. To confront this pressing issue, our research presents a novel and highly practical method with the use of Deep learning models. We harness the capabilities of Deep learning and image analysis to facilitate quick diagnosis of diseases. The paper has attempted to analyze the fecal images and train a range of models such as GoogleNet, Resenet18, ShuffleNet, SqueezeNet, and Vision Transform, achieving a peak test accuracy of 97.62%. Towards this, the study has made use of an extensive dataset of over half a million images and in the process accounted for the challenging environmental conditions such as dust on cameras and lower quality of images. The model can automatically identify common poultry diseases such as Coccidiosis, New Castle Disease, and Salmonella through the analysis of fecal images. The study is an attempt to bridge the gap between technological advancements and the day-to-day requirements in disease detection requirements of the poultry farming community, which may in turn support the poultry industry in better sustainability, economics, and enhanced welfare of the birds. The dataset used in this study has been made available Poultry Pathology Visual Dataset (kaggle.com).