Abstract:
Many languages have made great progress in the field of character recognition, including:
Latin, Chinese, Japanese ... where high rates of recognition reach 100%, while the recognition
of Arabic letters is witnessing low rates, due to some characteristics in the Arabic language that
hinder It is difficult to identify. Therefore, the work carried out in the framework of this
memorandum was related to the development of the system for the recognition of printed Arabic
characters. In this regard, a study on the structure of the Arabic language was presented,
followed by a presentation of the convolutional neural network technique, which proved its
efficiency in rapid recognition and obtaining more reliable results. Where two models CNN1
and CNN2 are proposed to measure the recognition accuracy and monitor the effect of the depth
of convolution layers on the recognition performance. CNN1 consists of two convolutional
layers (3x3) and CNN2 consists of three convolutional layers. Where the weights are adjusted
after entering all the examples (Bath training), in contrast to the weights are adjusted after each
example is entered (online training). The results showed that the CNN2 model was
distinguished to get better accuracy and loss Less.
