Efficient Way of Feature Extraction for the Recognition of Handwritten Arabic Characters

The fact that Arabic script is semi-cursive and the direction is from right to left has an effect on the shape of the characters, also the baseline helps to differentiate between characters. Unfortunately, this information is implicit. To take advantage of this information, character recognition is considered according to two visions: global and local. Depending on a global vision, the character must be considered as a whole. Once a word is correctly segmented, each character could be recognized separately, but the information of the baseline gets lost. Furthermore, the existing datasets of handwritten Arabic letters do not integrate this information. However, when the feature extraction is done after decomposing the image into four regions from the centre of gravity, a score of 96.22% is obtained. This process circumvents the lack of this information. And according to a local vision, most of the information is located in the right part of the character, extracting features from the useful part of the character gives a recognition rate of 95.67%. Extracting features using both visions boost the recognition rate to 96.58%. In this work we have used an Artificial Neural Network (ANN) classifier and a Local Binary Pattern (LBP) as a descriptor on the IFHCDB dataset.