Recognizing Arabic handwritten words using multiple features and classifier selection

In this paper, we present a new robust offline Arabic handwritten words recognition based on the holistic approach. We propose to use three descriptors namely Multi-Level Local Phase Quantization (ML-LPQ), Histogram of Gradient (HOG) and Gabor features to describe images. Each feature is fed to three well-known supervised classifiers K-Nearest Neighbors, Naïve Bayes and Support Vector Machine (SVM), and then for each feature, we retain and select the classifier that yielded the best results. Moreover, in order to improve recognition performance, we opt for using majority vote technic as a combination scheme. Experimental results are carried out using our own created database. The obtained results, given at the end of the paper, have demonstrated the efficiency of the proposed method where an average recognition of 97.84% has been achieved.