Project description
In this project, I built a pipeline that can be used within a web or mobile app to process real-world, user-supplied images. Given an image of a dog, the algorithm will identify an estimate of the canine’s breed. If supplied an image of a human, the code will identify the resembling dog breed.
In the first part of the project, I worked in a Jupyter notebook to perform the following steps:
- Use Haar feature-based cascade classifiers to detect human faces in images
- Use a pre-trained (on ImageNet) ResNet-50 model to detect dogs in images
- Design a CNN architecture to identify dog breeds
- Use Transfer Learning from VGG16 to identify dog breeds
- Use Transfer Learning from GoogLeNet to identify dog breeds
My own CNN architecture (step 3) reached a 35.76% accuracy on the test set, well above the minimum requirements for the project (1%). It was trained for 4 hours on a GPU. However, using transfer learning from the Inception/GoogLeNet was very successful with a final accuracy of 80.5%.
Check the Jupyter notebook for more details.
Web application
In a second step, I built a Flask web application to serve the model through a Bootstrap/JQuery web interface. The full code is available on Github.
Technologies Used
