In this study, we present two offline hybrid classification approaches for Indian (Arabic) handwritten numerals in effort to obtain higher reliability and classification rates than those obtained by single classifiers. Both methods work at the pixel level. No feature extraction methods are used as the purpose of this study is to focus on classifiers. The first hybrid classifier introduced is the serial hybrid classifier. It consist of three consecutive single classifiers. The first level is Fuzzy C-Means classifier followed by Support Vector Machine for as second level when more details are required and finally confirmation of classification will be through unique pixels method which forms the third classification level. The second hybrid classifier is the parallel hybrid classifier. It fuses the decisions simultaneously obtained from a Fuzzy C-Means classifier with the decision of a Neural Network to obtain the final decision. Both algorithms are tested on the CENPARNI Indian (Arabic) handwritten numerals dataset. The overall testing accuracy reported is 88%, 89% for the serial hybrid classifier and the parallel hybrid classifier, respectively. The paper also reports the results obtained using different types of single classifiers and compares with the above mentioned hybrid classifiers results. It shows the superiority of the hybrid classifiers over single classifiers.
|Number of pages||14|
|Journal||International Journal of Applied Engineering Research|
|Publication status||Published - 2015|
- Bayesian Fusion
- Fuzzy- C Mean
- Handwritten Indian (Arabic) numbers
ASJC Scopus subject areas