Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

Fast and Accurate Digit Classification

Subhransu Maji and Jitendra Malik

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2009-159
November 25, 2009

http://www.eecs.berkeley.edu/Pubs/TechRpts/2009/EECS-2009-159.pdf

We explore the use of certain image features, blockwise histograms of local orientations, used in many current object recognition algorithms, for the task of handwritten digit recognition. Existing approaches find that polynomial kernel SVMs trained on raw pixels achieve state of the art performance. However such kernel SVM approaches are impractical as they have a huge complexity at runtime. We demonstrate that with improved features a low complexity classifier, in particular an additive-kernel SVM, can achieve state of the art performance. Our approach achieves an error of $0.79%$ on the MNIST dataset and $3.4%$ error on the USPS dataset, while running at speeds comparable to the fastest algorithms on these datasets which are based on multilayer neural networks and are significantly faster and easier to train.


BibTeX citation:

@techreport{Maji:EECS-2009-159,
    Author = {Maji, Subhransu and Malik, Jitendra},
    Title = {Fast and Accurate Digit Classification},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {2009},
    Month = {Nov},
    URL = {http://www.eecs.berkeley.edu/Pubs/TechRpts/2009/EECS-2009-159.html},
    Number = {UCB/EECS-2009-159},
    Abstract = {We explore the use of certain image features, blockwise histograms of local orientations, 
used in many current object recognition algorithms, for the task of handwritten 
digit recognition. Existing approaches find that polynomial kernel SVMs trained on 
raw pixels achieve state of the art performance.  However such kernel SVM approaches 
are impractical as they have a huge complexity at runtime. We demonstrate that with improved 
features a low complexity classifier, in particular an additive-kernel SVM, can 
achieve state of the art performance. Our approach achieves an error of $0.79%$ on 
the MNIST dataset and $3.4%$ error on the USPS dataset, while running at speeds 
comparable to the fastest algorithms on these datasets which are based on multilayer neural 
networks and are significantly faster and easier to train.}
}

EndNote citation:

%0 Report
%A Maji, Subhransu
%A Malik, Jitendra
%T Fast and Accurate Digit Classification
%I EECS Department, University of California, Berkeley
%D 2009
%8 November 25
%@ UCB/EECS-2009-159
%U http://www.eecs.berkeley.edu/Pubs/TechRpts/2009/EECS-2009-159.html
%F Maji:EECS-2009-159