Abstarct: Categorizing and recognizing Persian handwritten numbers is a key process and as the first and most important step, it is used in many applications in the field of machine vision and image processing, such as the postal system (sorting letters baxsed on postal code), the banking system (paper analysis. check) and authentication. Due to the uneven and uncoordinated structure in writing numbers and the use of different imaging devices, the classification of handwritten numbers has always been a challenging issue. Some of the existing classification methods are semi-automatic and depend on the human factor. In recent years, deep learning-baxsed networks have achieved advanced performance in text image classification. Recognizing handwritten numbers using deep neural networks is an important task in the field of machine vision. To recognize handwritten numbers, convolutional neural networks are usually used. In the proposed method of this thesis, first, several different methods are put to test and by combining all the experiences, an innovative deep network is proposed and invented, in which the best results are obtained with the least possible laxyer. The existing network with 50 iterations was able to obtain an accuracy of 99.44%, which improved the highest available result by 44%.