A Computer Vision System for Early Detection of Sick Birds in a Poultry Farm Using Convolution Neural Network on Shape and Edge Information

Abstract

Poultry farming is increasingly becoming popular because it provides cheap alternatives to animal-based proteins in the form of eggs and meat. It is also easy to practice poultry farming because it does not require very large pieces of land to do poultry farming. However, large-scale poultry farming faces the challenge of disease management. Most of the time, routine vaccination, and healthy feeding have been relied on to promote healthy growth and production of poultry products. However, these practices are not sufficient to fully address the problem of disease management in poultry farming. One of the most assured ways of managing disease is close monitoring and detecting sick hens early enough before the disease spreads catastrophically. This study aimed at developing a computer vision system for the early detection of sick birds in a poultry farm using a convolution neural network on shape and edge information. This was achieved by creating a labeled dataset of sick and healthy birds based on edges, ridges, and Harris corners. Four convolution neural network models were trained, one based on full feature sets of the images while the other three were based on edges, ridges, and Harris corners. The purpose of training the different models was to establish which descriptors can best predict the health status of a hen based on its appearance of the hens. The assumption was that sick hens generally have their wings and tails stoop downwards with a weak neck that is bent downwards. Images were captured using a camera and used in four convolution neural network models. Three models were based on the features extracted (Ridges, edges, and Harris corners) while the other model was trained based on the full image without removing a single feature. The models were compared in terms of their accuracy of prediction. The models were evaluated based on performance during training and performance on unseen data. The model based on Harris corners was found to perform best at an accuracy of 94.14% while the model based on full feature sets was found to perform least with an accuracy of 46.66%. The model based on Harris corners was used to develop a web-based system that was deployed to predict the health status of hens in a poultry farm. The study was able to achieve its objective and proved that it is possible to classify healthy and sick birds based on a single feature.