This paper concerns the issue of Arabic letters recognition by means neural networks and concentrate on some intricacy’s aspects in Arabic script. In order to achieve superior identification accuracy on a variety of fonts and sizes, the paper have proposed a novel solution which fuses some traditional image processing methods with deep learning based technology. Second approach: This is a proposed way in which scanned photos are translated into binary format, quality of the images gets better and then it undergoes letter identification using CNNs (Convolutional Neural Networks). The accuracy rate in the Tahoma font was around 92% for both types of experiments with a recognition range between 60 and 95%. The topic of this paper was the method, results and potential applications for STP in use cases such as (automatic translation services), text-to-speech systems or digital archiving.