1 of 20

Dog Breed Classifier

Project Expo

2 of 20

Dogs vs Cats Summary

  • Kaggle Dogs v Cats Competition

​

  • 25000 differently sized training images (half dogs half cats)

​

  • Trained custom ‘traditional’ CNN

3 of 20

~87% Test Accuracy

4 of 20

Breed Classifier Project Goal

  • Predict breed of dog from an image

​

  • Dataset: Stanford Dogs
    • 120 Breeds
    • 20,580 Images (~170 images / breed)
    • Somewhat Unbalanced

​

5 of 20

Utilize Pre-Trained General Image Models

  • Trained on millions of images for general classification

​

  • Models Used
    • Inception V3
    • VGG-16
    • ResNet-50

​

6 of 20

Final Accuracies

Inception V3 ~ 43%

ResNet-50 ~ 60%

VGG-16 ~ 53%

7 of 20

Bella

Miniature Pinscher/Chihuahua Mix

Actual

Predicted

Inception V3

Chihuahua

VGG-16

Cardigan

​

ResNet-50

Cardigan

8 of 20

Luna

Golden Retriever

Actual

Predicted

Inception V3

Golden Retriever

VGG-16

Saluki

​

ResNet-50

Saluki

9 of 20

Morgan

English Cocker Spaniel

Actual

Predicted

Inception V3

Cocker Spaniel

VGG-16

Welsh Springer Spaniel

​

ResNet-50

Sussex Spaniel

10 of 20

Kiva

Shiba Inu

Actual

Predicted

Inception V3

Malamute

VGG-16

Cardigan

​

ResNet-50

Norwegian Elkhound

(Not in dataset)

11 of 20

Maggie

Yorkshire Terrier

Actual

Predicted

Inception V3

Lakeland Terrier

VGG-16

Irish Terrier

​

ResNet-50

Airedale Terrier

​

​

12 of 20

Snowy

Yellow Lab

Actual

Predicted

Inception V3

Weimaraner

VGG-16

English Foxhound

​

ResNet-50

Rhodesian Ridgeback

​

​

13 of 20

Buddy

Black Lab

Actual

Predicted

Inception V3

Flat-Coated

Retriever

VGG-16

Labrador

​

ResNet-50

Labrador

14 of 20

ResNet-50 Top 10 Best

Inception V3 Top 10 Best

No Overlap

15 of 20

ResNet-50 Top 10 Best

Inception V3 Top 10 WORST

16 of 20

ResNet-50 Top 10 Best

Inception V3 Top 10 WORST

4 Overlaps!

17 of 20

ResNet-50 Top 10 WORST

Inception V3 Top 10 Best

18 of 20

ResNet-50 Top 10 WORST

Inception V3 Top 10 Best

3 Overlaps!

19 of 20

Implement Ensemble Model If More Time

20 of 20

Areas For Improvement

  • Ensemble Model

​

  • Expand dataset to include missing breeds

​

  • Balancing dataset

​

  • Check breed pairs with highest confusion

​

​