Identidog - A Dog Breed Identification App

This is the main repository for the app Identidog 🐶. As the screenshot shows, the app takes an input image showing a dog or a person and returns the top resembling breeds. One may try out and use the app in any web browser by visiting this page served by the AI/ML platform Hugging face spaces.

Repo Layout

The repo is organised as follows:

• The root folder contains the README.md, (MIT) Licence information, the configuration files: requirements.txt and environment.yml, a project notebook, walkthrough.ipynb , containing a detailed walkthrough of the development of the app , and the folder "walkthrough_imgs" containing the corresponding images.

• The folder "app" contains the main app script app.py, the required images and models, and a copy of the Huggingface Readme.

How Identidog came to be

Identidog was developed as a fun machine learning project for trying out identification of the mixed-breed dogs living in my home. Although dog identification apps have already been available since 2016, I found most give unsatisfactory results for mixed breeds - and figured I would obtain a better understanding of the methods and limitations by making my own app.

In addition, the project was meant to help cement techniques I learned from fastai's deep-learning course.

How Identidog works

The app essentially strings together three image classifiers:

• A classifier which detects human faces in image
• A classifier which detects whether a dog is in the image
• A classifier which identifies dog breeds from among 133 of some of the most well-known ones.

For the first two classifiers we load pretrained models: a front-facing haarcascades model from OpenCV for face detection and the VGG-16 Convolutional Neural Network (CNN) from PyTorch Hub, pretrained on the Imagenet database.

The breed identifier was created by fine-tuning a pretrained CNN ResNet (also trained on Imagenet) to our specific dataset of 133 dog breeds using the fastai's deep-learning library.

Acknowlegements

• This project was based on this repo from Udacity. In particular, the training data and template for the walkthrough were obtained there.
• Fastai's deep-learning course and book were invaluable resources in learning how to train and deploy a CNN classifier using transfer learning.
• The app is dedicated to Cosmo, Loki and Monty.

Author: Alexei Kudryashov