Since the dawn of time, humanity has always been obsessed with two things: First up is their appearances. With the increased boom in social media, photo editing tools and tweaking images have been at an all time high. The other is dogs. So we decided why not combine the two.
To start with we chose two dog breeds which are very similar in appearance to see if the classifier can differentiate the two. As you can see the classifier does a very good job in differentiating the Italian Greyhound and the Whippet.
We used three major corruption techniques on our images: editing, doodling and superimposing the images.
Editing:
The first method we used is editing, where we changed the saturation, exposure, contrast and sharpness of the image.
After altering the image’s exposure, contrast, saturation, and sharpness, the accuracy of the image classifier in identifying the dog breed decreased.
Overlapping:
Since we noticed that even though the classifier was a bit confused with the edited images, we decided to combine the two images as well as overlap them to see how well it could identify them. And the results are as below:
Doodling:
We then proceeded to doodle on the image by creating a “santa dog” and outlining certain features, the classifier had a hard time to identify the images especially when the different colors were taken into consideration.
We think the classifier had a hard time classifying the two images mainly due to the fact that we did choose very similar dog breeds that it was initially hard for us to differentiate between the two and then proceeded to make the situation worse for the classifier with the noise. But overall the classifier did a better job than most humans would have.
Pooja Srinivasan and Parithosh Dharmapalan authored this blog