Detection of Bovine Species on Image Using Machine Learning Classifiers

Abstract

There are too many cattle in the world and too many breeds of cattle. For someone who is new to cattle breeding, it may be difficult to tell which species their cattle are. In some cases, even an experienced person may not understand the breeds of two cattle that are similar in appearance. In this study, the aim is to classify the cattle species with image processing methods and mobile applications written in Flutter and TensorFlow Lite. For classifying breeds, The VGG-16 algorithm was used for feature extraction. XGBoost and Random Forest algorithms were used for classification and the combined versions of the two methods were compared. In addition, SMOTE algorithm and image augmentation algorithms were used to prevent the imbalance of the dataset, the performance results of the combined versions of the two methods were compared. Images of different cattle species from different farms were obtained and the dataset was prepared, different image processing models were trained, the trained models were tested and the performance analyses were made. As a result of performance tests, it is obtained that the best model is VGG16+Random Forest+SMOTE+Augmentation with 88.77% accuracy result for this study. In the mobile application, first the cattle is detected with a pre-trained object detection model, and then the breed classification of the cattle on the image is made with image classification model.