npalaska / cnn_dog_breed_classifier

Dog Breed Classifier using Convolutional Neural Networks

Data Science Nanodegree project on building a dog breed image classifier with keras.

Overview

This project is part of Udacity's Data Science Nanodegree program. The aim of this project is to create a robust CNN model for detecting dog breeds based on the given image.

I have created a web application using Flask that is able to identify a breed of dog if given an image as input. The web application gives top 5 results based on the probability of dog belonging to a certain breed. If the image contains a human face, then the application will say that its a human.

Final model for identifying dog breeds achieved around 83% accuracy level on the test dataset.

Project Motivation

The reason I chose this capstone project is because I am currently working on a lane detection algorithms using opencv and this project gave a great boost in understanding different image processing, ImageNet dataset, and of course Convolutional neural nets and different types of training it.

This project allowed me to explore state-of-the-art CNN models for classification and helped me understand the nuances of building a CNN from scratch as well as using transfer learning.

Data

The training data for this project is located here. This dataset contains 133 different breeds of dogs and is already split into train, test, and validation sets.

General Dependancies

• Flask==1.0.2

• Flask-Cors==3.0.7

• Jinja2==2.10

• Keras==2.0.2

• Keras-Applications==1.0.7

• Keras-Preprocessing==1.0.9

• numpy==1.14.4

• opencv-python==4.1.0.25

• pandas==0.24.2

• python-dateutil==2.8.0

• scipy==1.1.0

• tensorflow==1.13.1

• tensorflow-estimator==1.13.0

• Theano==1.0.2

Please take a look at requirement.txt file for more detailed list

Instructions to run the web application

The Convolutional Neural Network is trained on 133 different dog breeds:

• If a dog image is selected:
  • Determine top 5 dog breed it most likely is based on probabilities.
• If a human face image is selected:
  • It displays that the chosen image is of human
• If neither a dog or human face detected it asks to choose another image

Steps to run the web app

1. Clone the repository.

git clone https://github.com/npalaska/CNN_Dog_Breed_Classifier.git
cd CNN_Dog_Breed_Classifier

2. Create and activate a new virtual environment.

python3 -m virtualenv cnn
source cnn/bin/activate

3. Download the dependencies using pip.

pip install -r requirements.txt

4. Start the app locally

python dog_breed_classification.py

Overview of dog_app jupyter notebook

The exploration of the state of the art CNN models for dog breed detection/classification happened in following steps:

• Step 0: Import Datasets

• Step 1: Detect Humans

  • We used OpenCV's implementation of Haar feature-based cascade classifiers to detect humans

• Step 2: Detect Dog using ImageNet

  • We used a pre-trained ResNet-50 model to detect dogs in images. the model has been trained on ImageNet, a very large, dataset used for image classification and other computer vision tasks

• Step 3: Pre-processing the imported data that can be used by the CNN models built by the Keras

  • When using TensorFlow as backend, Keras CNNs require a 4D array as input (tensors), with shape (nb_samples,rows,columns,channels)
  • Where nb_samples corresponds to the total number of images, and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.

• Step 3: Create a CNN to Classify Dog Breeds (from Scratch)

  • I used multiple 2D convolutional layers with first layer having an input shape =(224, 224, 3) which is a size of an individual image in our dataset.
  • I used the relu activation function and dropout rate of 40% on all the layers except the output layer.
  • On the output layer I have used the softmax activation function
  • The accuracy I got is around 11%
  • I have also experimented using data augmentation to see if it helps improve the accuracy, however data augmentation only assisted with reducing the loss function values but did not helped in improving the accuracy itself.

• Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)

  • To improve the accuracy I experimented with multiple bottleneck features such as VGG16, Inception and Xception, however, I found out that bottleneck features from Xception greatly improves our accuracy to around 83%
  • I have pre-trained the model and already saved the model weights in saved_models folder for later use in the web application

• Step 5: Written an algorithm for dog breed classification

• Step 6: Algorithm testing

Acknowledgements

This dataset is provided by Udacity and this work is done under Udacity guidelines as a part of data science capstone project.