Classification of Dogs

My implementation of the Convolutional Neural Networks (CNN) algorithm for identifying a canine’s breed from an image. Additionally, it supply the resembled dog breed if provided an image of a human.

Classification

You can see an example classification for the German Shepherd picture below:

And an example of a misclassification for a rotweiler.

Notes

You can see my full analysis of the classifier in the notebook, but a snippet is included below

The output result is about where I expected it to be. ResNet50 is very good with large image data, and I provided minimal layers to the algorithm, I expected a good base performance. However the model has only achieved 81% score, which would not work very well in a production environment (app or SaaS).

Potential improvements that are available for this model are:

• Reduce overfitting with usages of dropout and batch_normalization layers
• Add batch_normalization to reduce covariate shift in the calculation process
• Change the optimizer to another type, to find a better optimizer fit for my problem (adagrad or adam)

None of the above are guaranteed to produce a better result, just potential directions that I could pursue to improve the performance of the machine.

Project

This project was submitted by Jonathan Beverly as part of the Nanodegree At Udacity. The source was originally pulled from https://github.com/udacity/dog-project.git.