The system for automatic remote reading of meter indications is suggested. Two existing alternative solutions of the system implementation are considered. The main drawback of both solutions is specified, which particularly consists in the necessity of using of the specialized meters. The proposed system, in contrast, is universal, does not require meter replacement, has high reliability due to the absence of moving parts, and permits manual remote diagnostics. The general structure of the proposed system is described. It consists of primary reading device, based on a low-cost videosensor, a microcontroller and a wireless data transmission module, as well as the receiver (server part) with optical recognition software. The software component is considered in more details. Specifically, it is suggested to be implemented using the two-stage procedure. In the first stage, the meter scale is detected and partitioned into separate digit positions. The search of meter scale is proposed to be implemented using the Hough transform, which performs the lines location. Since the Hough transform, which takes binary images as an input is employed, the Canny edge detection algorithm is supposed to be applied first. The meter scale detection is based on the location of lines that forms a rectangular region with specific proportions. When the detection of the scale is performed, the located region is segmented into a separate digit subregions using the binary morphological operations and connected components analysis. The described process is applied relatively seldom (usually when the primary reading device is (re)installed), thus the operator of the system may manually correct the results of the detection and segmentation (these results will be valid for the further use). It reduces the quality requirements for the results, which obtained in the first stage. In the second stage, the recognition of each digit is performed. For these purposes, the use of convolutional neural network is suggested. This network is based on the architecture similar to the LeNet: it operates on small grayscale images and has seven layers (three convolutional, two max pooling, one ReLU and one fully connected layer). The training of network was performed on a specially composed training set. The training itself was conducted using the standard stochastic gradient descent (SGD) with momentum. The “softmax” was used as the loss function. In the result of the training, the correct digit recognition rate in 99,2% is achieved.Ref. 16, fig. 6.