Metaheuristic approach on feature extraction and classification algorithm for handwrittten character recognition

Abstract

Handwritten Character Recognition (HCR) is a process of converting handwritten text into machine readable form and it comprises three stages; preprocessing, feature extraction and classification. This study acknowledged the issues regarding HCR performances particularly at the feature extraction and classification stages. In relation to feature extraction stage, the problem identified is related to continuous and minimum chain code feature extraction at its starting and revisit points due to branches of handwritten character. As for the classification stage, the problems identified are related to the input feature for classification that results in low accuracy of classification and classification model particularly in Artificial Neural Network (ANN) learning problem. Thus, the aim of this study is to extract the continuous chain code feature for handwritten character along with minimising its length and then proceed to develop and enhance the ANN classification model based on the extracted chain code in order to identify the handwritten character better. Four phases were involved in accomplishing the aim of this study. First, thinning algorithm was applied to remove the redundancies of pixel in handwritten character binary image. Second, graph based-metaheuristic feature extraction algorithm was proposed to extract the continuous chain code feature of the handwritten character image while minimising the route length of the chain code. Graph theory was then utilised as a solution representation. Hence, two metaheuristic approaches were adopted; Harmony Search Algorithm (HSA) and Flower Pollination Algorithm (FPA). As a result, HSA graphbased metaheuristic feature extraction algorithm was proposed to extract the continuous chain code feature for handwritten character. Based on the experiment conducted, it was demonstrated that the HSA graph-based metaheuristic feature extraction algorithm showed better performance in generating the shortest route length of chain code with minimum computational time compared to FPA. Furthermore, based on the evaluation of previous works, the proposed algorithm showed notable performance in terms of shortest route length of chain code for extracting handwritten character. Third, a feature vector was derived to address the input feature issue. The derivation of feature vector based on proposed formation rule namely Local Value Formation Rule (LVFR) and Global Value Formation Rule (GVFR) was adopted to create the image features for classification purpose. ANN was applied to classify the handwritten character based on the derived feature vector. Fourth, a hybrid of Firefly Algorithm (FA) and ANN (FA-ANN) classification model was proposed to solve the ANN network learning issue. Confusion Matrix was generated to evaluate the performance of the model in terms of precision, sensitivity, specificity, F-score, accuracy and error rate. As a result, the proposed hybrid FA-ANN classification model is superior in classifying the handwritten characters compared to the proposed feature vector-based ANN with 1.59 percent incremental in terms of accuracy model. Furthermore, the proposed hybrid FA-ANN also exhibits better performances compared to previous related works on HCR.