Open-source dog breed identification using CNN: explanation of the development & underlying technological specifications

Abstract

This thesis aims to develop a deep learning model that uses the Fast.ai library and transfer learning to create a convolutional neural network model. The model will identify dog breeds in digital images and evaluate the developed model’s performance with regards to loss and accuracy. The thesis also aims to explain the underlying technological specifications of a deep learning model. The model has been trained on dog breed images from the Stanford Dogs dataset and tested against similar dog breed identification models and apps. Open source is a cornerstone of the thesis, and the code of the developed model will be hosted publicly on the GitHub platform under an open-source license for further expansion by others. The motivation for this research comes from the lack of accurate open-source dog breed identification tools alongside the growing popularity of dogs, as well as the author's passion for deep learning and computer vision. The research questions include the development of an accurate CNN model, the suitability of Fast.ai for image classification tasks, the selection of a pre-trained model to use with transfer learning, and the comparison of the developed model against similar models and apps. Overall, this thesis aims to create a well-functioning, well-documented and modern model that accurately identifies dog breeds, and the thesis and developed code are meant to inspire others to learn and work in the fields of computer science such as artificial intelligence, and deep learning.