Yasmine ElGlaly

In this project we build a system to classify isolated handwritten Arabic characters. The system we propose consists of three phases; preprocessing, feature extraction, and classification. The preprocessing phase concerns with denoising and normalizing the images. Then we extract 14 features from the processed images. Finally, we train the extracted data using back propagation neural network.

 

We succeed to build a character recognition system for the Arabic characters using two classification algorithms; the BPNN and the kNN. Although we expected that neural network will be the best solution for our problem, we find that kNN performs better. kNN has very low error rate in classifying new datasets, and its accuracy is 100% for the training datasets. Also, kNN is very robust to high noise in the input images. On the contrary, BPNN has high error rate in the training data, and very low accuracy in the testing data. Also, BPNN is sensitive to noise. According to the training time, BPNN needs to train for 5000 epochs before it responds with the character label. On the other hand, kNN is very fast in training the datasets. Using our 2.00 G.Hz processor and 3.0 G.B. RAM computer, kNN responds with the right class 150 times faster than the BPNN for the same input data. The proposed system can be enhanced by adding further operations in the preprocessing phase so that the letters will be the same size and in the same position, such as scaling, rotation, and translation normalization. We believe that the neural network can behave better and its accuracy increased with these normalized data. For more details, read this.