Vehicle license plate recognition method based on deep neural network

Abstract

The vehicle license plate recognition technology is a specific application of the text mode recognition technology in the field of transportation. A problem of unstable all-weather recognition rate still exists in an existing vehicle license plate recognition system in domestic at present. Acquisition of vehicle images is affected by variations in ambient light, and stains, color fading, inclination and movement of vehicle license plates also result in image blurring and distortion, so that a vehicle license plate positioning algorithm and a vehicle license plate recognition algorithm are difficult to be implemented under such conditions. The accuracy is improved by applying the deep neural network to image recognition, and manual characteristic extraction is avoided. A convolutional neural network model is a system structure of the deep neural network, and has a close relationship with image processing. According to the invention, a convolutional neural network is applied to vehicle license plate recognition and enabled to have a high recognition rate for abnormal characters such as broken characters, inclined characters and the like. It is indicated by experiments that the algorithm has a strong ability for distinguishing similar characters and has high robustness.

Description

technical field [0001] The invention relates to a license plate recognition technology based on a deep neural network. Background technique [0002] License plate recognition technology is a specific application of text pattern recognition technology in the transportation field. It mainly includes a series of steps such as image capture, image processing, license plate positioning, character segmentation, and character recognition. At present, the existing license plate recognition systems in China generally have problems such as unstable recognition rate around the clock. When collecting vehicle images, they will be affected by changes in ambient light, and the license plate itself will also appear stains, fading, tilting, and blurred images caused by motion. , so it is difficult to implement the license plate location algorithm and recognition algorithm in these cases, and the most widely used recognition method is the improved BP neural network method. BP algorithm is an...

AI Technical Summary

Problems solved by technology

At present, the existing license plate recognition systems in China generally have problems such as unstable recognition rate around the clock. When collecting vehicle images, they will be affected by changes in ambient light, and the license plate itself will also appear stains, fading, tilting, and blurred images caused by motion. , so it is difficult to implement the license plate location algorithm and recognition algorithm in these cases, and the most widely used recognition method is the improved BP neural network method

BP algorithm is an optimization algorithm for local search, but it needs to solve the global extremum of complex nonlinear functions, so the algorithm may fall into local extremum; the approximation and generalization ability of the network are closely related to the representativeness of learning samples, and from It is very difficult to select typical sample instances to form the training set in the problem; there is no unified and complete theoretical guidance for the selection of the network structure, and generally it can only be selected by experience; the contradiction between the generalization ability of the network and the training ability, In general, when the training ability is poor, the prediction ability is also poor, and to a certain extent, with the improvement of the training ability, the prediction ability also improves, but when it reaches a certain limit, with the improvement of the training ability, the prediction ability decreases instead. , that is, the so-called "overfitting" phenomenon occurs, which is because the network has learned too many sample details and cannot reflect the laws contained in the sample.

Images