A suitable shape descriptor

[matt von]

Dear Dr.Franz Graf,

 

First of all I would like to thank you for this excellent library!

 

I am currently doing my thesis where the aim is to create a texture retrieval system that can be used to recover textures from a collection that can stimulate particular high-level psychological feeling or emotions in the viewer. Such a texture retrieval system can be used for many applications such as in video games and film production industry as well as for traditional industries such as those of textiles and wallpapers.

Till know I made use of three feature descriptors from the JFeatureLib namely, ColorHistogram, Tamura and Gabor filter which produced great good results.

Now, I also need to implement a shape descriptor. The intention is that the user pick a shape from a provided number of different shapes and then the system recognize which images contain the chosen shape. Now, from the literature search that was performed I found that some of the most successful and frequently used shape descriptors are Fourier Descriptors, Curvature Scale-Space and Zernike moments. However, non of these descriptors are part of this library.

 

Can you please suggest the most ideal shape feature descriptor from the JFeatureLib for such application?

 

Kind regards,

Matthew

[Franz Graf]

Hi Matthew,

it is great to hear that you were able to use JFeatureLib for your research!

Hm yes you're right, there are not too much Shape descriptors in
JFeatureLib. I know Fourier Descriptors and Zernike Moments because I
once used them on my own research (unfortunately before I started
JFeatureLib). I can hardly give you an advice if or which of the present
descriptors can help you best.

If possible, I'd just to extract all of them and just try & compare.

Sorry that I can't help you in that point.
Kind regards
Franz

> --
> You received this message because you are subscribed to the Google
> Groups "JFeatureLib" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to jfeaturelib...@googlegroups.com
> <mailto:jfeaturelib...@googlegroups.com>.
> To post to this group, send email to jfeat...@googlegroups.com
> <mailto:jfeat...@googlegroups.com>.
> For more options, visit https://groups.google.com/d/optout.

--
Dr. Franz Graf
http://www.Lockede.de

[matt von]

Thanks Franz for your quick reply!

I will try all the shape descriptors and check which return the best result, as you suggested.

One last question... 

Unlike a CBIR that query by example (use an image to search for images), the distance measure between the feature vectors have to be performed between one particular shape (the user choice..the query) and the entire image. 

Since I am fairly new to image processing I would like to make sure if such shape descriptor of the JFeatureLib can be used for such purpose. All I want is to be sure that I not completely misguided! Does this involve some kind of masking on the images in the collection?

Thanks again for your help! 

--
You received this message because you are subscribed to a topic in the Google Groups "JFeatureLib" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/jfeaturelib/LSCjZRHW29A/unsubscribe.
To unsubscribe from this group and all its topics, send an email to jfeaturelib...@googlegroups.com.
To post to this group, send email to jfeat...@googlegroups.com.

[Franz Graf]

Hi Matt

Yes you definitely will have to apply masking. The descriptors only work on masked or binarized images.

Just for the sake of it : you might want to have a look at the work of Funkhouser from Princeton. He did a lot of shape matching. Maybe it helps a bit.

Good luck!
Franz

Dr. Franz Graf
http://www.Locked.de