Skip to content

ankitpt/dog_breed_classifier

Repository files navigation

Dog Breed Classifier

Table of Contents

  1. Project Motivation
  2. File Descriptions
  3. Results
  4. Acknowledgements

Project Motivation

This is my Capstone Project submission for Udacity Nanodegree. A dog breed image data was made available with 8352 images spread across 133 categories and the goal was to train a classifier for identifying breed of a dog in a given image.

Following tasks were performed for the project:

  • Create and train a CNN from scratch for dog breed classification
  • Use bottleneck features from pre-trained VGG model as input and train a fully connected network
  • Use bottleneck features from pre-trained Resnet50 model as input and train a fully-connected network
  • Use the network trained with Resnet50 features for testing on new images (not part of original dataset)

Folder Descriptions

images folder consists of images for testing on trained models

haarcascades folder consists of pre-trained face detector based on haar feature

requirements folder consists of .yml and .txt files enlisting package requirements on different operating systems for running the jupyter notebook

saved_models folder consists of transfer learning models trained based on bottleneck features from VGG16 and Resnet50

Results

The performance of each of the CNNs on test set is summarized as follows:

  1. CNN trained from scratch: 5.5% accuracy
  2. Neural net trained with bottleneck features from VGG16: 41.87% accuracy
  3. Neural net trained with bottleneck features from Resnet50: 80.02% accuracy

The main findings of this work can be found on my Medium blog

Acknowledgements

I would also like to acknowledge Udacity Data Scientist Nanodegree instructors for creating and teaching the Data Scientist Nanodegree program and thus providing a holistic overview of data science.

About

No description, website, or topics provided.

Resources

License

MIT, MIT licenses found

Licenses found

MIT
LICENSE
MIT
LICENSE.txt

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published