Prediction of Chicken Diseases by Transfer Learning Method

Abstract

With the development of computing technologies, artificial intelligence is used in a wide range of areas, from engineering to healthcare. In this study, it was aimed to predict chicken diseases with transfer learning. For this purpose, a ready-made data set was studied. This data set contains fecal photographs of healthy chickens diagnosed with Coccidiosis, Newcastle and Salmonella diseases. The data set has been subjected to necessary pre-processing such as size readjustment. Subsequently, the data set, which was then subjected to pre-processing, was divided into 70% and 30% as training and testing. To solve the disease classification problem, a network was created by adding fully connected layers to ResNet50, InceptionV3, InceptionResNetV2, Xception and MobileNetV2 architectures. The weights of the architectures mentioned in these networks were selected as ImageNet and were not trained. Then, networks containing these architectures were trained using the training data set. The trained networks were validated with the test data set and accuracy rates of 32.7%, 80.6%, 85.2%, 89.2% and 90.7% were obtained, respectively. According to these results, MobileNetV2 was used in the proposed artificial neural network architecture since the best result was calculated using the MobileNetV2 architecture. The proposed artificial neural network architecture was trained with the same training set and validation was carried out with the same test data set. After these procedures, the true prediction rate of the proposed architecture for the test data set was calculated as 92.1%. Also, F1 score of the proposed architecture was measured 0.923. Additionally, thanks to the deconvolution layer used in the proposed architecture, network sizes have been reduced by approximately 50%. Thanks to this reduction, the training time is shortened and it becomes easier to implement it on embedded systems in future studies. As a result, the diseases of chickens were predicted largely accurately with the transfer learning method.