Does the following image classification problem be termed as image retrieval ?

%------------ taken from matlab help ---------- %

%-----------------------------------------------%

%-------------- svmclassify code --------------%

load fisheriris

xdata = meas(51:end,3:4);

group = species(51:end);

svmStruct = svmtrain(xdata,group,'showplot',true);

species = svmclassify(svmStruct,[5 2],'showplot',true)

hold on;plot(5,2,'ro','MarkerSize',12);hold off

%--------------------% code ends %------------------------%

The above code gives the result ---> species = 'virginica'

The species is classified as 'virginica'. This is just one image. Can this process of classification be termed as "image retrieval" ???

Or do we have to retrieve many images, in order for it to be termed as image retrieval ?

Answer 1

[Walter Roberson]

No, there are no images involved in the above. The fisheriris data does not use images.

[Mohan]

Thank you for the reply Walter. I have changed the code according to my classification problem. I can retrieve an image from the result.

The problem is, I am retrieving only one image. Can this be termed as image retrieval ?

Why I am asking this question is, usually many images are retrieved in an image retrieval system when it is provided with a query image. In my system, only one image is retrieved.

[Walter Roberson]

Retrieving one image at a time can be okay for an image retrieval system. It depends on the purpose of the system.

If the purpose is to display exact matches for the retrieval criteria, then it could be that only one image matches the criteria exactly.

If the purpose is to display close matches for the retrieval criteria, then it could be that only one image is "close" to within the cut-off. For example, if there is only one red image in the database and the retrieval request requires strong red, then there might not be any point in retrieving the images of the blue sky and clouds.

Now consider if you had asked to classify [5 1.55]. If I read the graphs correctly, that would be just a hair over the dividing line into virginica, but it would also be very close to one of the versicolor samples. There would be reason to doubt the classification, but svmtrain is going to tell you virginica without qualification. That close to the boundary you should "morally" be retrieving both as possibilities. You should be figuring out a certainty measure for the classification and displaying multiple possibilities when the certainty is low enough.

Also, if there are only two images in the database, it would not usually be called image retrieval.