How to put a good threshold in comparing images

Hi sir,

i already have a feature vector containing color and texture features , and i am measuring the similarity between the query image and the database images. i want to set a threshold that gives me an accurate precision and recall

Would you help me with this please?

Answer 1

[Junaid]

It totally depends on type of images and applications. But you can keep various values of Threshold and on each value you calculate the F-score ( also knows as F-1 score), click here for detail. Best threshold value is that value which gives you maximum F-score ( near to one). Once you get values which gives you maximum F-score, then you can use this value on your application and see the performance of Threshold (sensitivity of Threshold).

I hope it may be helpful to you.

Answer 2

[Image Analyst]

What you want to do is called CBIR, Content-Based Image Retrieval. It's getting better and better all the time. It's way too complicated to give you any code here so all I can do is to refer you to some links where you can do some more research into several methods that are used:

http://en.wikipedia.org/wiki/Content-based_image_retrieval

http://en.wikipedia.org/wiki/List_of_CBIR_engines

Section 19.4 here:

http://iris.usc.edu/Vision-Notes/bibliography/contentsapplicat.html#Implementations%20and%20Applications,%20Databases,%20QBIC,%20Video%20Analysis,%20Hardware%20and%20Software,%20Inspection