Convolutional Neural Network Dog Classifier

Overview

Given an image of a dog, this notebook will estimate the canine’s breed. If supplied an image of a human, the code will identify the resembling dog breed.

Details

Training Data

Training data consists of over 8000 dog images tagged with their breeds. The images are split into training, test and validation sets.

The data contains instances of minimal inter-class variation, such as between retrievers and spaniels.

Retriever	Spaniel

There are also cases of high intra-class variation, such as within Labradors.

Golden Labrador	Brown Labrador	Black Labrador

Import Human Face Detector

OpenCV's human face detector using Haar Cascades is used here.

Import Dog Detector

The ResNet-50 model is used to identify images containing dogs.

The images are preprocessed according to ResNet-50's specifications (normalized and resized).

Train Dog Classifier

Transfer learning is used to create a CNN using Xception bottleneck features.

A 500 paramter layer is added to focus on the current dog image dataset, followed by a 133 parameter fully-connected layer.

Training was performed on an EC2 GPU instance.