In this thesis, a digit recognition system using Radial Basis Function Neural Networks classifier based on HOG algorithm for feature extraction is designed in order to recognize the handwritten digits. In the pre-processing part of the digit recognition system, Histogram of Oriented Gradient is applied to extract the features of digit image. The extracted features in the digit image are reduced from high-dimension to low-dimension by using pre-processing methods such as Principal Component Analysis, 2-dimensional 2-directional PCA, and Incremental-PCA. After performing the dimensional reduction, the features of digit images are learned and then the digit recognition is performed by using polynomial Radial Basis Function Neural Networks. The architecture of RBFNNs consists of three functional modules such as condition, conclusion, and inference part. In the condition phase, the input space is partitioned with the use of fuzzy clustering realized by means of Fuzzy C-Means algorithm. Also, this algorithm is applied instead of gaussian function to consider the characteristic of input data. In the conclusion part, connection weights are used as the extended diverse types in polynomial expression such as constant, linear, quadratic and modified quadratic.
In this study, three main experiments of the digit recognition systems are carried out based on pre-processing algorithms for comparative studies and analysis.
The first experiment is related to the comparative studies of PCA and (2D)2PCA based digit recognition systems. Thus, the digit recognition system is configured in order to compare the recognition rates by using PCA and (2D)2PCA for dimensional reduction.
The second experiment is focused on HOG-PCA and HOG-(2D)2PCA for constructing the digit recognition systems. In order to extract the features of digit images, HOG algorithm is applied before performing dimensional reduction algorithm. And then this second experiment is compared with the first experiment using PCA and (2D)2PCA.
In the final experiment, IPCA is considered to carry out the incremental feature extraction instead of PCA.
The advantage of the proposed pattern classifier is quantified in detail by running a number of experiments using the benchmarking MNIST handwritten digit database.