Using ME and PCA for Farsi Font Recognition

This paper presents a new method based on a combination of features and classifiers for font. Features are extracted and combined from textures of 128×128 pixels size. For feature extraction, Sobel-Roberts Features, Gabor Filter and different Wavelet transform same as Daubechies and Haar are employed. Extracted features combined together in a pairwise manner. Three MLP (Multi Layer Perceptron) used for classification, the MLPs run on different features. Then their output combined together by employing PSO (Position swarm optimization) algorithm for finding optimum weights. Experimental results have also verified this hypothesis. The proposed algorithm is examined on the Hoda dataset of 21000 samples prepared from 10 different Farsi fonts. The characteristics of the fonts are properly extracted in the method adopted by this research, achieved 97.45% accuracy in recognition rates so it has outperformed previous contributions and higher recognition rates.