The Relevant CNNs Features Based HMM for Arabic Handwriting Recognition
Mustapha Amrouch, Mouhcine Rabi · International journal of imaging and robotics · 2017
This paper presents a work that aims to compare the learning features with Convolutional Neural Networks (CNN) and the handcrafted features. In order to determine the relevant features among of those kinds. We consider our previous baseline HMM system Rabi et al. 2016, for Arabic handwriting recognition. Experiments have been performed on the benchmark IFN/ENIT database. Obtained results with CNN features surpass those achieved using the hand-crafted features. This demonstrates the high efficiency of CNN results from the strong capability for hierarchical feature learning given a large amount of data. However, Hand-crafted features are not extracted from an optimization process to be adequate with the specific problem, and insufficient to be encoded with supervision.