Experimental Study of Convolutional Neural Network Architecture for Pattern Recognition in Images

The choice of optimal tools and instruments of synthesis and modelling of neural network for pattern recognition in images is carried out. A comparative analysis of the results of training a neural network using the open library of machine learning TensorFlow, libraries Keras and NumPy, data sets from the open database MNIST in the recognition of handwritten input has been carried out. The software based on the neural network of the convolutional type by topology is developed, which solves the problems of handwritten numerical symbols recognition. The main design, technological and technical-operational characteristics: accuracy 99.2%, mini-packages - 200 pieces, the ratio of training and training sets - 0.2, the number of epochs - 10. The created software can be used in areas of visual analysis of data of paper documentation of enterprises, where it is necessary to transfer data from paper to electronic form. And also, to serve as a starting point for the development of the core of more powerful software for handwriting recognition, namely, a bunch of digits, symbols (car numbers, postal codes, etc.). #CSOC1120.