Skip to content

Deeplearning application from a CNN trained on Stanford dataset, to identify dog breed from a photo,

Notifications You must be signed in to change notification settings

opsabarsec/Dog-pictures-classification-by-Deep-Learning

Repository files navigation

Dogs breed image recognition CNN

1. Background

A dog shelter needs an automatic system to identify dog breeds and possibly match hybrids to the most probable breeds.

Cane

The project goal is to build a ML pipeline that predicts the breeds and associated probabilities from a photo of the dog. The algorithm has to be deployed to an API, easy to use for non technical customers.

In a further step an Android app can be built based on the same.

2. The data

Having a good training dataset is a huge step towards the robust model. There is Stanford Dogs Dataset with ~20K images of dogs of 120 breeds. Every image in the dataset is annotated with the breed of a dog displayed on it.

3. The model

Full code Based on Xception convolutional neural network, this model

training

provides 85% accuracy.

The model has been saved to a file that can be loaded in a matter of seconds. It can be run on Google Colabs to identify the breed from an uploaded dog picture. The same has been deployed to an API, the is result illustrated in this video.

The Android app was built using TensorFlow Lite and importing the model files into Android Studio.

dognet

The full explanation of the model and of the results is provided in a presentation.

4. Conclusions

  • Efficient breed recognition can be obtained through a web API, posting a picture and obtaining the most probable 5 breeds in matter of seconds.

  • An app for Android devices, OpenDogNet, has been developed and deployed to Google Play .

About

Deeplearning application from a CNN trained on Stanford dataset, to identify dog breed from a photo,

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published