Research On Automatic Slant License Plate Recognition In The Wild

With the rapid development of computer and communication technology,the func-tion of automatic information processing is widely used in all aspects of people’s life.The use of Automatic License Plate Recognition system liberates manual labour and improves the efficiency of vehicle management.Nowadays,with the demand of vari-ous scenarios such as driverless technology and intelligent transportation,and the rapid development of intelligent technology also need to improve the speed and accuracy of Automatic License Plate Recognition technology.The technology of Automatic license plate recognition has important application value and research significance.Today,both the commercial and academic fields are investigating various methods that can be used for automatic license plate recognition systems.However,most of the data set images extracted by the current methods are concentrated in some specific li-cense plate areas,that is,most of the license plate images are almost frontal to detect the license plate.In this case,if the license plate position in the image that needs to be tested is tilted,distorted,etc.,the detection and recognition results are not ideal.In WPOD-NET^[1],it is proposed to correct and recognize images with oblique shooting angle in some unconstrained scenes.By testing this method,it cannot cover all the situations of capturing license plate image,and even the distorted license plate images collected in extreme scenes cannot be recognized.In this paper,according to the characteris-tics of the above unconstrained scheme,we have added the function of detecting the license plate image（the image is extremely tilted or the distortion is not parallelogram）taken under extreme conditions.A layer of feedback mechanism has been added to the network to make the license plate images that are wrongly detected in certain occasions automatically judged as errors and re-entered the network to exclude wrong answers for re-detection,thereby effectively improving the robustness and accuracy of the network.According to the baseline flow characteristics of the automatic license plate recognition system,multi-threaded GPU parallel acceleration is adopted to increase the license plate recognition rate.The main research work and innovation of this article are as follows:1.According to the characteristics of the unconstrained license plate recognition scheme WPOD-NET^[1],modify the loss function part of the full connection layer of the“DETECTION”model bounding box regression.Take the coordinates of the three corner points of the license plate and the L₁norm output by the network as the losses during the training.According to the three corners of the parallelogram,the fourth corner of the license plate can be obtained.The test results show that using the same network structure to learn fewer features can learn more accurate features.In order to enhance the network’s ability to detect images in extreme scenes,the shape of the license plate is not constrained to be parallelograms.All four corners of the license plate are involved in network training part.The test results show that the network has a stronger ability to learn the license plate images with more distortion.The method in this paper can detect the license plate more accurately.（see Section 1 and 2,Chapter 3in this thesis）.2.Due to the activation of some features in process of license plate detection,the network will detect a small number of vehicle images such as lights or wheels as license plates,and obtained wrong results.Combined with the output of character recognition after error detection,a layer of feedback mechanism was developed.Zero-setting the pixels in the license plate area that has been detected incorrectly to make it impossible to activate the network.Re-detection of some incorrectly detected license plates can significantly modify the detection results and improve the accuracy rate by about 4%.（see Section 3,Chapter 3 in this thesis）3.In actual traffic applications,in addition to the fixed-point shooting images,more license plate recognition is needed for the images in the video or large batches of images.Therefore,on the basis of ensuring the accuracy of the license plate,the improvement of speed has become necessary.Combined with the characteristics of the baseline process of the automatic license plate recognition system,the video,especially the batch images,is accelerated for license plate recognition.Accelerate according to the characteristics of video frame-to-frame translation and schedule GPU to execute parallel processing of segmentation tasks.According to the translation properties of video frames,the acceleration algorithm can be increased to approximately 1.30-2.82times the initial calculation speed.According to different scheduling acceleration algo-rithms,it can increase to 1.76-2.29 times of the initial calculation speed.（see Chapter4 in this thesis）...