Search

Offline handwritten character recognition is a part of the arduous area of research in the domain of document analysis and recognition. In order to enhance the recognition results of offline handwritten Gurumukhi characters, the authors have applied hybrid features and adaptive boosting approach in this paper. On feature extraction stage, zoning, diagonal, centroid, and peak extent-based features have been taken into account for extracting the meaningful information about each character. On the classification stage, three classifiers, namely decision tree, random forest, and convolution neural network classifier, are used. For experimental work, the authors have collected 14,000 pre-segmented samples of Gurumukhi characters (35-class problem) written by 400 writers where they have used 70% data as training set and remaining 30% data as testing set. The authors have also explored fivefold cross-validation technique for experimental work. The AdaBoost approach along with the fivefold cross-validation strategy outstands the existing techniques in the relevant field with the recognition accuracy of 96.3%. [ABSTRACT FROM AUTHOR]

Massive dataset can be analyzed by establishing physical distributed environment or by hiring cloud-based distributed environment. The advantage of cloud-based environment over physical environment is that, it provides scalable virtual resources on demand and thus makes it suitable for handling increase in volume of the data. The various hidden patterns in data can provide knowledge bases for decision making. The statistical or data mining based methods can be used for finding knowledge patterns. Among the decision tree based classification algorithms, implementable in distributed environment, an efficient algorithm can be selected based on few parameters such as execution time, accuracy of prediction and complexity of the tree structure. In this study, Apache Hadoop-based distributed environment is established on virtual machine. Apache Spark is installed to execute machine learning algorithms. The comparative study of binary classification methods such as decision tree, gradient boosted tree and random forest tree is performed to judge their performances on the basis of defined parameters. It is found that Random forest tree performs best among all three algorithms for the considered dataset. [ABSTRACT FROM AUTHOR]