Online Recognition of Outlines in Teeline Shorthand Language

Abstract

To summarize the speech delivered by the VIPs in the short time the short hand newlinelanguage is employed. Mainly there are two most popularly used shorthand languages namely newlinePitman and Teeline. An automatic shorthand language recognition system is essential in order newlineto make use of the handheld devices for speedy conversion to the native English language. To newlineaddress the above issue we had attempted the system in this thesis using the three methods, newlinefirstly using the Finite Automata, using the machine learning approaches and lastly using the newlineDeep learning (CNN). The method involves collecting the input image, preprocessing, newlinesubtracting the context to the grayscale image, extracting the features and applying them to newlinethe classifier or directly giving the Deep Learning images for English letter prediction. newlineAny of the pattern recognition system starts with the data collection step, an android newlineapplication along with the digital pen iball8060U is interfaced to extract the data drawn from newlinethe user. Considered the 45 to 50 different writers to generate the data. The study reviles that newlinethe data generation becomes considerably well as user became familiar with the application newlineand digital pen. The android application allows the user to write the alphabets in the provided newlinetemplate space only. The 10,600 dataset has been generated in it 60% is reserved for training, newline20% of dataset is reserved for testing and validation. newlineThe handwriting process is documented by the acquisition system as a stream of (X, newlineY) coordinates with the correct pen position sensor and pen-up / pen-down switching. No newlinepressure level was reported. The dataset distribution consists of 53 files (one for each author) newlineand a file for data collection. This file contains descriptions of the Id (ID), the name of the newlinecharacter (Label) and the specific type (Char). The stroke information is collected in order to newlinerecognize the behavior of the Finite State Automata. newlineThe generated dataset is used to extract the feret features from each image reserved newlinefor training process. The ten shape-based features are extracted from the images to form the newlinefeature vector space with dimension 1 x 13. The feature mapping is fed to the Support vectors newlineand K-Nearest neighbourhood. The same process is extended to the Teeline character newlinerecognition. newlineTo train the deep learning model have employed a 5000-image dataset is provided to newlinethe fifteen layered network including input and output layer. During the training process newlineevery image are passed through every layers and respective weights and bias are calculated. newlineOnce a single image reaches the final layer, before taking a new image weights and bias of newlineevery layers are tuned to increase the accuracy. The same procedure is repeated on the rest of newlinethe imagein the training process. newlineThe finite automata method to determine Teeline shorthand language is implemented in the newlineJAVA of version 1.8 and achieved the considerable accuracy. The machine learning and deep newlinelearning approaches are implemented in MATLAB2018a. The accuracy of the machine newlinelearning approaches is considerably less compared to the deep learning method. The machine newlinelearning approaches have achieved the accuracy in the range of 74 to 82 percentage. The deep newlinelearning approach had achieved the accuracy in the range of 90 to 96 percentage. By newlineobserving the accuracy rate the deep learning can be considerably applied to the real time newlineapplication for speedy conversion from Teeline shorthand language to English letter. newline