Dog Breed Classification and Identification Using Convolutional Neural Networks

Abstract

This study aimed to assess the effectiveness of using pre-trained models to extract biometric information, specifically the dog breed and dog identity, from images of dogs. The study employed pre-trained models to extract feature vectors from the dog images. Multi-Layer Perceptron (MLP) models then used these vectors as input to train dog breed and identity classifiers. The dog breeds used in this study comprised two Thai breeds, Bangkaew and Ridgeback, and 120 foreign breeds. For dog breed classification, the results showed that, among the ImageNet classification models, the pre-trained NasNetLarge model has the highest dog breed classification accuracy (91%). The newly trained MLP model, which used feature vectors obtained by NasNetLarge, achieved higher accuracy at 93%. For dog identification, the results showed that, without data augmentation, the pre-trained ResNet50 model had the highest dog identification accuracy (75%). However, with data augmentation, MobileNetV2 could achieve a higher accuracy of 77%. When evaluating the identification performance of each breed, it is important to note that pugs achieved the lowest identification rate at 57.4%. Conversely, Bangkaew dogs demonstrated outstanding performance, with the highest identification rate at 98.6%.