Purpose. To ensure recognition of faults in conditions of changes in lighting, distance and camera position relative to the fault, as well as rapid adjustment to the recognition of new types and components of symbols of faults that can be detected in the process of implementation and practical use.
Specifications. The average processing time for an input image with dimensions of 800×600 and 1600×1200 is 20 ms and 60 ms, respectively. The horizontal size of the license plate in the image must be at least 65 cells.
Application area. Public and state security, municipal economy, motor transport sector, information technology.
Advantages. Recognition algorithms are used that are relatively resistant to changes in the symbol image, the presence of interference, as well as to changes in the symbol image due to changes in the position of the surveillance camera, the slope and thickness of the lines. Unlike specialized recognition systems based on deep learning, the technology allows you to quickly configure the system to recognize additional formats of information and symbols and does not require a large number of examples for training.
Technical and economic effect. To learn recognition, it is enough to have a relatively small number of examples of the types of NCs and the symbols used. This makes it possible to quickly learn to recognize new symbols and types of NCs within about ten minutes, which is significantly less time-consuming and labor-intensive compared to other technologies. When testing the software that implements the claimed method, the accuracy of fault recognition was 95% of correct recognition of nineteen types of faults on a sample of images in which cars were photographed at different distances and angles of shooting objects relative to the camera. Recognition was performed in real time; the number of character standards formed as a result of training varied from one to four for different characters.
Description. The method is based on the fact that, firstly, character recognition in NO is performed by searching for their key points, which helps to tune in to processing new types of NO using a small number of examples, and secondly, correction of the resulting text string is performed in the process of calculating its distance to sets of similar NPs, each of which is determined by the type of such NPs.