An Efficient Arabic HMM System Based on Convolutional Features Learning

Abstract

Published works recently indicate that the generic features extracted from the convolutional neural networks are very powerful. This paper shows that CNN feature can be used with a HMM system [1] for Arabic handwritten word recognition, to yield classification results that outperform the handcrafted features. These features are usually based on heuristic approaches that describe either basic geometric properties or statistical distributions of raw pixel values. The CNN features based HMM is shown satisfactory recognition accuracy on the well-known IFN/ENIT database and outperformed some other prominent existing methods.