Characters Type Recognition in Moroccan Documents Using CNN

EasyChair Preprint 8774

Abstract

In recent years, the classification and recognition became a hot topic in computer studies. Deep Learning algorithms present the most outstanding performance in classification and recognition issues. In this paper, we focus on applying these techniques to extract the characters types from the Moroccan official legal documents. Because of the variety of many pre-trained models. we designed a system able to Loop over each model available on TensorFlow Keras API, based on Transfer Learning technique, we trained the models on the dataset that we have built. And the outcome was that the DensNet and VGGNet models have achieved the best performance, with a validation accuracy of 98%. In addition to this, we proposed a modified model based on DenseNet201, the result achieved is 98.99% of overall accuracy.

Keyphrases: Keras, Transfer Learning, character recognition, image classification, pre-trained models

@booklet{EasyChair:8774,
  author    = {Ali Benaissa and Abdelkhalak Bahri and Ahmad El Allaoui},
  title     = {Characters Type Recognition in Moroccan Documents Using CNN},
  howpublished = {EasyChair Preprint 8774},
  year      = {EasyChair, 2022}}