These exercises were cloned from this Udacity github location.
The project runs through the following steps:
- Import datasets for pictures of human faces and of dogs of various kinds.
- Detect and extract faces from the human pictures using a Haar Cascade.
- Detect dogs using a ResNet50 model pretrained on ImageNet data, which is available in keras.
- Create a CNN to classify dog breeds from scratch using keras functions.
- Now use a pretrained CNN to classify dog breeds using transfer learning from the VGG-16 model.
- Pick a different pretrained CNN to classify dog breeds using transfer learning. I picked the VGG-19 model.
- Write an algorithm for a draft application that receives an input image, determines whether a dog or a human face is present, and then classify the picture according to what dog the picture most resembles.
- Test the algorithm - feed multiple images of one's own choosing to test how the algorithm performs.
Below are two pictures, one of a celebrity, and another of a dog, and the output of the network in terms of what dog breed is resembled in those pictures. Additional such pictures along with their network classifications can be found in this notebook.
