dog-identification-CNN

Motivation of the project

This project aims to identify dog breeds from images using Convolutional Neural Networks so that I will gain experience in handling CNN algorithms and better grasp the concept. Throughout the project I employed 3 different CNN algorithms that are either fully or partially trained across the dataset provided by Udacity. If the image fed to the algorithm is recognized as a human or dog face, most resembling dog breed type is presented as the trained algorithm’s prediction on what is shown in the image. If the algorithm does not recognize a human or a dog, it states so.

Summary of the project

The project contains the following steps:

A sub-algorithm for detecting human faces are created.
A sub-algorithm for detecting dog faces are created.
Own CNN algorithm is created for dog breed identification, trained using the dataset provided by Udacity.
VGG16 CNN model is trained using transfer learning method for dog breed identification, trained using the dataset provided by Udacity.
ResNET-50 CNN model is trained using transfer learning method for dog breed identification, trained using the dataset provided by Udacity.

Summary of the results

The sub-algorithm called face detector used for detection of human faces was tested across 100 dog images and 100 human images with accuracy results listed below:

Human faces detected in 100 human face images provided in ratio: 100%, therefore it has 100% accuracy.
Human faces detected in 100 dog face images provided in ratio: 11%, therefore it has 89% accuracy.

The sub-algorithm called dog detector used for detection of dog faces was tested across 100 dog images and 100 human images with accuracy results listed below:

Dog faces detected in 100 human face images provided in ratio: 0%, therefore it has 100% accuracy.
Dog faces detected in 100 dog face images provided in ratio: 100%, therefore it has 100% accuracy.

Own model architecture has reached around 2.8% accuracy in correctly identifying the breed type of the test set of the dog images provided.

VGG16 model architecture has reached around 41.3% accuracy in correctly identifying the breed type of the test set of the dog images provided.

ResNET-50 model architecture has reached around 83.3% accuracy in correctly identifying the breed type of the test set of the dog images provided.

Files in repository

README.md : (this file) to explain the basics of the project
dog_app.ipynb : Jupyter Notebook where all the calculations and results could be attained.
dog_app.html : Html version of the finalized Jupyter Notebook for ease of access through a web browser application.

Python libraries used in project are as follows:

cv2
keras
re
tqdm
sklearn
numpy
matplotlib
random
PIL (implicitly)

Acknowledgments

License : MIT
Data source : Udacity

References

For more information please refer to blog post through this link.

qtepinkclou / dog-identification-cnn Goto Github PK