image multi classification with keras

Question

so if I have two labels "dogs" and "cats" and I want to create multi classification neural network.

now if I provided a new random image which is not a dog or a cat, is there a way I can teach the classifier to tell me that this image is not a dog or a cat instead of saying how much percent it maybe cat or dog?

Anshul Rai · Accepted Answer · 2018-09-22T22:48:52.633

1

The best way to accomplish this is to create a new class in addition to dog and cat to handle images you have no interest in. So now, your labels would be ["dogs", "cats", "other"].

In your current architecture, your model is forced to predict a random image as either a dog or cat as those are the only two options it has. Adding a new class to deal with other images is generally the easiest way to make your classifier more robust to incorrect predictions.

edited Sep 22 '18 at 22:48

answered Sep 22 '18 at 22:43

Anshul Rai

772
7
21

but should I add some random examples in this "other" class or not ? – osama Sep 23 '18 at 21:23
Yes, you would need to add some random images. – Anshul Rai Sep 23 '18 at 22:36

image multi classification with keras

1 Answers1