Real-time and efficient license plate recognition method and device for embedded terminal and medium

Abstract

The invention provides an embedded terminal real-time efficient license plate recognition method and device and a medium, and the method comprises the following steps: reading a to-be-recognized license plate image, transmitting the to-be-recognized license plate image to a license plate recognition network model, carrying out the layer-by-layer calculation, and outputting a license plate recognition result; wherein the license plate recognition network model sequentially comprises a first residual block, a second residual block, a third residual block, a fourth residual block, a fifth residual block, a sixth residual block, a seventh residual block, an eighth residual block and an output layer; the first residual block is used for performing primary convolution calculation on the to-be-recognized license plate image to obtain a first shallow license plate feature; the second residual block is used for performing convolution calculation on the first shallow license plate feature to obtain a second shallow license plate feature and the like. The real-time and efficient license plate recognition method for the embedded terminal is high in real-time performance and not prone to being influenced by illumination, license plates with various character numbers can be recognized, and the license plates can be seamlessly adapted to the embedded terminal through verification.

Description

technical field [0001] The invention relates to the technical field of license plate recognition, in particular, to a real-time and efficient license plate recognition method, equipment and medium at an embedded terminal. Background technique [0002] The license plate is one of the external significant identity information of the car, and the driving route, type, driver and other information of the vehicle can be obtained through the license plate. License plate recognition has a wide range of applications, and is widely used in smart parking, highway vehicle monitoring, and city number restrictions. [0003] At present, the traditional license plate recognition algorithm used has high requirements on the quality of the license plate picture, and there are many factors that affect the quality of the license plate picture, such as light, the distance and angle between the license plate and the image acquisition equipment, etc., so the traditional license plate recognition al...

AI Technical Summary

Problems solved by technology

[0003] At present, the traditional license plate recognition algorithm used has high requirements on the quality of the license plate picture, and there are many factors that affect the quality of the license plate picture, such as light, the distance and angle between the license plate and the image acquisition equipment, etc., so the traditional license plate recognition algorithm cannot meet the complex light requirements. , multi-color license plate, high recognition rate, fast recognition requirements

[0004] It can be understood that the traditional license plate recognition algorithm needs to separate the characters in the license plate area first. Segmentation algorithms include template-based character segmentation algorithms, clustering algorithm character segmentation, etc. Such methods are seriously affected by light, and inaccurate segmentation directly leads to Subsequent recognition is wrong, and the accuracy of license plate recognition is low; machine learning algorithms such as svm and knn are used to recognize the segmented characters. Character recognition is the most important link in license plate recognition, and its accuracy directly affects the accurate recognition of license plates ; However, using machine learning algorithms to recognize characters, there is still a lot of room for improvement in the recognition rate

[0005] With the emergence of deep learning, end-to-end recognition methods have appeared, such as crnn (full name: An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition), lprnet (License Plate Recognition via Deep Neural Networks ), etc. Although the accuracy of license plate recognition based on deep learning has been greatly improved, the computational complexity of the model is relatively large, and some layers of operators do not support the phenomenon when transplanting on the embedded end, such as the processor rk3399. The deep learning reasoning framework rknn does not support the lstm (Long short-term memory) layer in crnn and the 3D pooling layer in lprnet. It is difficult to implement these layers with code and the real-time performance is difficult to guarantee, resulting in high accuracy. The deep learning algorithm cannot be adapted to the embedded side, resulting in low accuracy and real-time performance of the embedded algorithm when applied to complex light and multi-color license plate recognition

[0006] Although the Chinese patent with the application number CN201810736742.9 discloses a license plate recognition method based on a simplified ResNet residual network, the patent does not change the structure of resnet18, and the output of the fully connected layer of the last layer is 7 ( Blue card seven characters), 7 neurons correspond to each character in the blue license plate, once the model is trained, it can only recognize blue or yellow cards (7 characters), even if it is used to recognize green cards (8 characters ) The result obtained is still 7 characters, so this license plate recognition method can only recognize license plates with fixed characters, and cannot recognize blue cards, yellow cards and green cards at the same time, let alone recognize green cards with 8 characters

Images