Development of water buffalo face recognition models using deep learning and traditional machine learning techniques

In this study, an artificial intelligence-supported automatic facial recognition system has been developed as an alternative to traditional identification methods for large livestock. Facial images of eight buffalos from a facility in the Yozgat province were used in the dataset created to train the models. The images were captured from various angles to make the raw dataset. The raw dataset was augmented using some commonly used augmentation techniques in the literature, and ultimately, a new dataset named BUFFALO-22 was created. In the study, the results were compared by employing both traditional and modern methods in the training of the BUFFALO-22 dataset. On the traditional side, common feature extraction techniques (Gray Level Co-occurrence Matrix and Local Binary Patterns) were utilized on images pre-processed with certain techniques to extract the desired features. The extracted features were classified using the Support Vector Machine as the traditional classifier. In the conducted tests, a test accuracy of 87.7% was achieved with the LBP-HSV model. On the other hand, in the realm of modern techniques, the YOLOv5 model, the fifth version of the YOLO (You Only Look Once) deep learning model widely recommended for image classification tasks such as object recognition, was employed alongside pre-trained deep learning models like InceptionV3. Experimental results have demonstrated the successful utility of both pre-trained CNN classifiers like InceptionV3 and real-time detection approaches like YOLOv5 in animal identification tasks based on facial recognition. According to the results, the InceptionV3 approach achieved an average test accuracy of 98.5% in facial detection. Training the YOLOv5 algorithm resulted in a precision of 0.98, a recall of 0.99, and a mean average precision (mAP) value of 0.99. Furthermore, in addition to training, validation, and test accuracy metrics, other criteria have also supported the notion that the models employed in this study can be comfortably used for animal identification tasks. The study suggests that animal identification may pave the way for the facilitation of various research areas such as performance monitoring, reproduction, health and welfare tracking, and behavior analysis.