End-to-end license plate identification method

Abstract

The invention discloses an end-to-end license plate identification method. The method comprises the steps of firstly segmenting a license plate region to serve as an input according to a license plate locating result; secondly predicting a license plate character thermodynamic diagram by adopting a designed three-layer full convolutional neural network, and obtaining candidate character regions according to the thermodynamic diagram; thirdly performing classification and position correction on the candidate character regions by adopting a designed seven-layer deep convolutional neural network model to obtain character sequences and position sequences; and finally selecting out the character sequences meeting Chinese mainland license plate standards by adopting a template-based optimal path algorithm, wherein the selected character sequences are identification results. The whole network adopts a multi-task combined training mode; the input is a license plate color image, and an output is a license plate number; and the method avoids the defect that a license plate needs to be accurately segmented in a conventional license plate identification method, and can effectively increase the license plate identification rate in a complex scene.

Description

technical field [0001] The invention relates to the technical field of license plate recognition, in particular to an end-to-end license plate recognition method. Background technique [0002] With the development of computer technology and information processing technology, the information processing ability of computer has been continuously improved, and computer vision technology has been widely used in intelligent transportation and electronic police systems based on multimedia, pattern recognition and artificial intelligence technology around the world. Among these applications, 96% of the automated systems use the automatic license plate recognition technology, and more than 75% of the systems use license plate recognition as the core application. [0003] Traditional license plate recognition methods usually include license plate detection, license plate correction, license plate segmentation, and character recognition. For license plate pixels greater than 120 pixels...

AI Technical Summary

Problems solved by technology

[0003] Traditional license plate recognition methods usually include license plate detection, license plate correction, license plate segmentation, and character recognition. For license plate pixels greater than 120 pixels, and the license plate is not stained, the license plate deflection angle is less than 30 degrees, and the license plate has no backlight. , the license plate recognition rate is greater than 95%, but for small license plates (less than 100 pixels) and the license plate is backlit or defaced or the deviation angle is too large, the license plate recognition rate often drops sharply. License plate segmentation methods such as vertical projection cannot effectively deal with such dirty or large-off-angle license plates, because in this case, the license plate characters and the background often blend with each other, and it is very difficult to accurately segment, and the segmented Accuracy often affects the accuracy of character recognition

Images