Classify the Unknown or invalid image in image classification

[bara...@gmail.com]

Hi,

I am using the BOW and trained a set of images. I need to classify the invalid or unknown images. I think i need to manipulate the distance threashold to find out the same. For example, i have a medical image database.. if end user uploads the CAR image. this should fall under unknow category. that is KSVM should give value like -1. But now ,the inavlid image is coming under the valid category value. Can anyone suggest me to acheive the same in accord.net. I currently stuck up with the same.

[César]

Hi there,

If you are using multiclass support vector machines, you may be able to use the ksvm's raw distance output to detect when an image should be rejected. It is given as an out parameter in the Compute method. If you calibrate the machines first using ProbabilisticOutputLearning you will also be able to retrieve probabilities from the machines, which you can then use to define a rejection threshold.

Hope it helps!

Cesar