Deep learning-based modeling and recognition of poultry diseases

In this paper, various architectures of neural networks of deep learning technologies were created to model viral diseases of poultry. Models for recognizing coccidiosis, salmonellosis, and Newcastle disease, which are common diseases in poultry farming, were built. The constructed models for detecting poultry diseases are aimed at diagnosing sick birds at an early stage. Convolutional neural networks and deep learning models were created, which are based on predicting poultry diseases by classifying healthy and unhealthy feces images of four types. Unhealthy feces images that can be symptoms of coccidiosis, salmonellosis, and Newcastle disease were identified. A model was built using the base model of convolutional neural networks of various architectures. The models were trained using feces images labeled by the farm and the laboratory, and then their hyperparameters of the neural network were fine-tuned. The test set used images labeled with our own data taken from poultry farms. The test accuracy results without fine-tuning are obtained, which amounted to 85.06% for the base CNN, 87.85% for the updated CNN. Fine-tuning while freezing the batch normalization layer improved the accuracy of the model to 95.01% with F1 scores for all classifiers above 83% in all four classes. Considering the smaller weight of the model trained with the CNN and its better generalization ability, we recommend using this model for early detection of poultry diseases at the farm level. The main result of the work is the implementation of models based on transfer learning of VGG 16, ResNet 18 and Efficient Net models. The efficiency of the transfer approach to building models and their comparative analysis are shown. Error matrices and reports on the classification of poultry diseases based on transfer learning up to 99% accuracy are built.