Identification of some avian diseases based on images obtained from a camera by convolutional neural networks

Abstract

The research paper introduced several results achieved in the research and construction of an automatic system for early detection of various poultry diseases by computer vision technology, using artificial intelligence to analyse images from surveillance cameras. The experimental system used the deep learning neural network RepVGG as an inference model for detecting poultry diseases. Training data set and model evaluation for 3 diseases ORT, CRD, MAREK, including 500 images for each disease, were directly collected from surveillance cameras and from the internet. The training process used 80% of the sample images; the model reached an error of 0.02 after 5000 rounds of training. Evaluation of the model with the remaining 10% of samples achieved an accuracy of 99.94%. In the actual test with the surveillance camera (resolution 1902x1080 pixels), the system can detect and classify correctly diseases over 90%, especially for diseases with images involving multiple distinguishing symptoms, the accuracy was more than 98%. The obtained results showed that the system can be used to support early warning of diseases for poultry. The system can be trained to detect other poultry diseases such as Newcastle disease, fowl pox, pullorum disease, typhoid...