Analysis Of Failure In Offline English Alphabet Recognition With Data Mining Approach

• Main Author: Munnian, Ruthrakumar
• Format: Monograph
• Language: English
• Published: Universiti Sains Malaysia 2019
• Subjects: T Technology; T351-385 Mechanical drawing. Engineering graphics
• Online Access: http://eprints.usm.my/58271/1/Analysis%20Of%20Failure%20In%20Offline%20English%20Alphabet%20Recognition%20With%20Data%20Mining%20Approach.pdf; http://eprints.usm.my/58271/

Offline handwriting recognition is a long existing approach to identify the handwritten phrase, letters or digits. Earlier studies in the handwriting recognition field were mostly focused on recognizing characters using Neural Network Language Model (NNLM) classifier, Hidden Markov Model (HMM), and Support Vector Machine (SVM) with segmentation technique, Hough Transform method, and structural features. However, these approaches involve complex algorithms and require voluminous dataset as the training model. Therefore, this study attempts a data mining approach to the analysis of failure in offline English alphabet recognition. The objectives of the study are to improve the pattern recognition approach for classifying English alphabets and to determine the root of classification failure in handwritten English alphabets. Handwritten data of capital letters of the English alphabet by 50 Universiti Sains Malaysia student experimented. The data was pre-processed to remove the outliers prior to classification analysis with the aid of the Waikato Environment for Knowledge Analysis (WEKA) tool. Classification analysis was initially performed on all seven classifier’s algorithms at 10-fold dross validation mode. At phase one, Stroke and Curve are added into the dataset and classified respectively. At phase two, Sharp Vertex, Closed Region, and Points are added in the dataset. The top three classification algorithms were selected: IBk, LMT and Random Committee for further classification. The classified result was further analyzed to identify the root of classification errors. At the raw dataset classification, the classification accuracy is low with 25%. As the attributes are added to raw dataset respectively, the accuracy of classification was successfully increased to 89%. Conclusively, the accuracy of the classification depends on the added attributes to distinguish characteristics of the alphabets.