Comparative Analysis of Models for Real-Time Pattern Recognition

• Main Author: Nordlöf, Jonas
• Format: Others
• Language: English
• Published: Umeå universitet, Institutionen för matematik och matematisk statistik 2014
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:umu:diva-96781

The different applications for optical character recognition in real-time applications will most likely increase in the future as innovations as autonomous vehicles or elder care robots become a reality. This analysis therefore aims to evaluate different "off-the-shelf" models that can be used in these applications. Four different classifiers have been combined with three different feature extraction procedures, giving a total of twelve models, have been used in the analysis. The evaluated classifiers are Artificial neural networks (ANN), -nearest neighbour ( NN), Support vector machines (SVM) and Random forest, while the raw image, wavelets and principle component analysis (PCA) were used as feature extraction procedure. The analysis used handwritten numerals from the MNIST-library as training and test data. Four different properties have been studied; these are dependencies of training data, accuracy of prediction, time for prediction and robustness against noise. The ANN classifier was the fastest, SVM had the highest accuracy, NN was the most robust against noise while the Random forest model had the highest accuracy when smaller training sets were used. Using principal-components as features to the classifiers increased the model robustness against noise.