michaelgsuttles / dog_breed_identification

Using a convolutional neural network to identify a dog breed from a picture.

Dog Breed Identification - Michael Suttles

Introduction

• Animal shelters around the country could benefit from a system that easily identifies a dog by its breed—especially in cases where the dog is a mixed breed.
• I used a ResNet-50 convolutional neural network for image classification of dog breed. (See future work section.)
• I used TensorFlow in Python to implement the CNN, and Shapley Additive Explanations (SHAP) to interpret the model.

Data Sources

• I scraped Google Images for the dataset of 35 breed images
  • about 400 images per breed
• I added the Stanford dog breed dataset images
  • 100-200 images per breed, but some breeds not included

Optimizing the algorithm

• Tweak different parameters, including:
  • How large the image is that the CNN is finding features from? 244x244 pixels, 350x350, etc.
  • What do you do to the image? Crop, zoom, skew...
  • The structure of the “fully connected layers” at the end
  • How many times (“epochs”) it should run
• Improved accuracy from 70% to 80% for 35 breeds

Results

• 3 breed model: 95.7% accuracy
• 35 breed model: 80% accuracy

How the neural net “sees” the image

Future work

• Implement a Densely Connected Convolutional Network instead of a ResNet-50, which has been shown to have a higher degree of accuracy.
• Further refine model to account for errors (for example, along border of image).
• Look into an ensemble model, which can result in higher accuracy.

For further information

• Email: [email protected]
• LinkedIn: http://www.linkedin.com/in/michaelgsuttles
• Website: http://www.doggydatazone.com