[An LSI Infomercial.. ;-)] Fwd: Some initial results of application of information retrieval techniques to image retrieval

[Subbarao Kambhampati <rao@asu.edu>]

> Date: Sat, 21 Dec 2002 11:58:05 -0800 (PST)
> From: Kanav Kahol <kkahol@yahoo.com>
> Subject: Some initial results of application of information retrieval
>  techniques to image retrieval
> To: rao@asu.edu
>
>
> Dear Sir
>
> As i had said in class i was very inspired by the information retrieval 
> techniques and wanted to apply it to content based images retrieval.
>
> IN our lab we had collected data on defining images in terms of keywords. 
> SO for e.g. an image of mountains could be marked by words snowy, flaky 
> forest etc.  so we had 94 images of natural scenes and we collected data 
> on what keywords represent an image.
>
> simultaneously through some unfortunate souls we collected data on which 
> imaghes is most similar to which. so a person came in took one of the 94 
> images and told us the top 20 similar images. (parallel i guess with 
> relevance feedback where we ask user what is relevant).
>
> So now we had a document for each image + subjective measure of images 
> most similar to it.
>
> Before this class all we used to do is take dot products of keywords for 
> every image to find dot product measure. What we found was that the words 
> assigned to image lined up pretty well with subjective similarity.(Say for 
> image 1 we had words forest snow and for document 2 the words were forest 
> river there dot product will be [1 1 0]*[1 0 1] which is 1. )
>
> For me the blessing was this course. now i took this data of 94 
> images(each having 98 keywords) and put in  a term-document matrix. ONly 
> this is know term-image matrix.
>
> As a check up I performed LSI on this. And here is the exciting part. Only 
> 20 dimensions compared to 98 were now needed. While this is good here is 
> more exciting part for me. Actually the 20 dimensions are lining up 20 
> concepts in language. For e.g.  one of the dimension was actually a 
> weighted combination of all the words related to forests. while another 
> one was similar to river water related words.
>
> What this reduced dimensionality lining up with definable concepts or 
> dimensions does is instead of tell people to define 98 words is just 
> define vague concepts and still be able to retrieve images.
>
> While going from images to words sutomatically is still an big image 
> proceesing issue what this analysis gives us is a good measure of how much 
> we can do with this theory.
>
> I must say i am surprised with the results and will continue to probe in 
> this . []
>
> Regards
>
> Kanav