You can use many svm to classify .

In this paper introduce this method.

title :

One-Against-All Multi-Class SVM Classification Using ...

http://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=10&cad=rja&uact=8&ved=0CIkBEBYwCQ&url=http%3A%2F%2Fwww.stat.umn.edu%2F~xshen%2Fpaper%2FOne-Against-All-Zheng.pdf&ei=dbtTU5uYGYXEsgbp2oGQBA&usg=AFQjCNGZvvO_BtE4smlmrG_44uoBkAe0XQ&sig2=t7sWuL7cByGwscclBhW2Sw&bvm=bv.65058239,d.Yms

LIBSVM can perform multi-class classification in MATLAB. The only difference from two-class classification on your side will be to make sure that all classes are represented in the training set and that the label vector has numerical values representing each class.

http://www.csie.ntu.edu.tw/~cjlin/libsvm/

the following paper reported some methods of multiclass svm. http://eprints.soton.ac.uk/271333/1/hasdam.pdf

In case of one aginst rest method you need to be aware of balanceing the number of classes samples because the number of samples of the rest classes might be sgnificantly more tha the number of one class.

My first choice for multiclass svm in matlab was libsvm but after reading its description I found that it is originally developed for binary classification and multiclass classification using libsvm is an ongoing research topic.And the multiclass work done using libsvm is using binary classification technique which I don't know how so I am not sure is libsvm the right choice for multi class classification in matlab for image dataset

Thank you for all links but these are all research papers giving theoretical background I want practical implementation guideline for multiclass svm in matlab for image dataset. Can some one please help?

Sounds like you are looking for something similar to bag of words (or features) type approach. It is unclear if you need to classify single images or image contents, but in either case good starting point might be work done by A.Vedaldi et al. They offer a load of information (and also some code) to study on the topic [http://www.vlfeat.org/applications/apps.html].

That might be an option to start - but,... Please notice that you are threading on an area that is not trivial. You cannot expect just to "implement" something without knowing what you are doing and expect that the things work magically. So I would spend an considerable amount of time to understand the problem and study the theory too [Vedaldi et al have also a load of tutorials], otherwise it may be difficult to go beyond the state of art...

Yes I have already seen some of work by A.Vedaldi et al but I would look more into it.Lets see if I get any clue,thank you all.

HI

In order to improve the predictive accuracy of the SVM , you need to focus on the optimization of its parameters.

Next and the best option is to create HYBRID SVM that will surely give you better performance,.

Also if you follow the above option, then most probably you will end up in publishing your work in good reputed HIGH IMPACT FACTOR journals.

Hope this will help you to some extent.

Thank you.

Best,

INDRAJIT MANDAL

http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=5439692&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D5439692

LIBSVM toolbox, will a good option. These LIBSVM implementation of SVM algorithm is available online. Otherwise, testing with WEKA software can be a good alternative to Matlab.