Handwritten Character Recognition Using Multiclass SVM Classification with Hybrid Feature Extraction

In this paper, we describe hybrid feature extraction for offline handwritten character recognition. The proposed technique is a hybrid of structural, statistical and correlation features. In the first step, the proposed technique identifies the type and location of some elementary strokes in the character. The strokes to be looked for comprise horizontal, vertical, positive slant and negative slant lines–as we observe that the structure of any character can be approximated with the help of a combination of simple straight line strokes. The strokes are identified by correlating different segments of the character with the chosen elementary shapes. These normalized correlation values at different segments of the character give correlation features. For making feature extraction more robust, we add in the second step certain structural/statistical features to the correlation features. The added structural/statistical features are based on projections, profiles, invariant moments, endpoints and junction points. This enhanced, powerful combination of features results in a 157-variable feature vector for each character, which we find adequate enough to uniquely represent and identify each character. Prior, handwritten character recognition problem has not been addressed the way our proposed hybrid feature extraction technique deals with it. The extracted feature vector is used during the training phase for building a support vector machine (SVM) classifier. The trained SVM classifier is subsequently used during the testing phase for classifying unknown characters. Experiments were performed on handwritten digit characters and uppercase alphabets taken from different writers, without any constraint on writing style. The obtained results were compared with some related existing approaches. Owing to the proposed technique, the results obtained show higher efficiency regarding classifier accuracy, memory size and training time as compared to these other existing approaches.