Resumen de Transfer Learning Study for Horses Breeds Images Datasets Using Pre-trained ResNet Networks
Enrique Antonio de la Cal Marín, Enol García González, José Ramón Villar Flecha
In previous work, we have carried out an academic study of the automatic classification of horse breed images by pre-trained DL models.In the present paper, we continue that line of research by extending the former results considering a new dataset including known and unknown breeds. Thus, two main goals are tackled here: i) new experiments of transfer learning considering the known breeds of both former and new datasets, and ii) a study of similarity between the known and unknown breeds. When trying to classify unknown breeds, it is expected that the models obtained in goal i) can be used to analyze the morphological similarity between unknown breeds and known breeds. In order to “evaluate” the results of this analysis, we have relied on the advice of an expert in the field of horses.From the experts’ point of view, the horses’ morphology defines some of the typical uses: riding, draught, multi-purpose. Thus, as most of the comparisons agreed with the expert’s assessment, the research line into morphological similarities using pre-trained DL models is reliable. Future work will be proposed to carry out similarity studies with other datasets and similarity studies using parts of the horse’s body instead of taking full photos.
