The paper discusses the technology of creating character recognition (using convolutional neural networks) systems on the image. These days, there are many approaches to solving this problem, and most of them are ineffective for images whose symbols are located on a complex background and are vulnerable to noise, affine and projection distortions. The proposed technique consists of the following stages: image pre-processing, text segmentation, and recognition by convolutional neural networks. During research was conducted a series of experiments, namely: experiment to select the most suitable method of binarization of digital images, experiment to select the most efficient convolutional neural network topology form text recognition problem. As a result of the experiments performed, this technique as applied to the recognition of car numbers demonstrates high reliability and accuracy, including in low light conditions, therefore, the developed recognition method can be recommended for commercial use. As an additional field of experiments was suggested a bunch of approaches of how to improve this technique.