Deep Learning for Horse Breed Recognition

Recognition of specific types of objects and entities in images has been among the research interests for a long time. In animal species recognition in still images, horse breeds identification has not yet been investigated. In this paper, the identification of horse breeds in natural scene images is addressed. Since a publicly available horse breed dataset of images was not found, a dataset is collected by searching on the Internet. The dataset contains 1693 images of 6 horse breeds. Deep learning methods, specifically Convolutional Neural Networks (CNNs) are dominant approaches today, especially for classifying objects in natural scenes. In this work, Transfer Learning of well-known deep CNN architectures, pre-trained on the ImageNet dataset and fine-tuned on the proposed dataset, is used as the method of choice. Here VGG architectures with 16 and 19 layers, InceptionV3, ResNet50, and Xception are applied as the pre-trained CNN classifiers. The average classification accuracy of ResNet50, the most accurate classifier on the collected dataset, is 95.90 which provides a baseline for further research in the field of horse breed recognition. The results are investigated to specify which horse breed is similar to which, from the perspective of a CNN.