Розпізнавання символів автомобільних номерів різних форматів

Abstract

Urgency of the research. Typically the recognition process includes the following steps: license plate detection, license plate normalization, segmentation of the license plate image into separate symbols and symbols recognition. The effectiveness of license plate recognition depends on each of the indicated recognition stages, but for recognition of the license plates of different formats the key stages are segmentation and recognition stages. Therefore the development of the recognition method of the license plates symbols of different formats is an actual task. Target setting. Different formats of car numbers have different fonts and different arrangement of characters, which complicates the process of recognizing car numbers. Actual scientific researches and issues analysis. General trends that have been identified by the analysis of publications indicate that for character recognition of car numbers used convolutional neural network, fully connected neural networks, correlation analysis, binarization images and histograms of brightness. Uninvestigated parts of general matters defining. All analyzed methods are well suited for recognition of the symbols of wellvisible license plates. This makes difficult to apply such methods in real conditions as the license plates can be dirty or poorly visible. The research objective. The purpose of the article is to describe the method of recognizing car numbers of different formats, which has a high percentage of correct recognition and can be used to recognize car numbers on the video stream from cameras located above the tracks. The statement of basic materials. For recognition of the symbols of license plates it is suggested to use the brightness histogram of the binarized image, for symbols recognition - a specially created neural network with the ability of recognition the alternative parts of the original image of the license plate and for removing the incorrectly recognized symbols - the list of license plates formats. Conclusions. The proposed method successfully copes with the task of license plate recognition with confidence 95-99 %. But as the test results show the method has several drawbacks. First, this method does not easily recognize the “trash” in the image and often confuses it with the symbol ‘I’. Second, on the dirty license plates or on false detection this method repeatedly uses alternative recognition which leads to a significant load on the processor.