Research On Fast Automatic License Plate Detection Algorithm In The Wild

Automatic license plate recognition algorithm is widely used in the world.This technology is mainly used in parking lots,highways,residential and other places.Also,the technology can provide some help in unmanned driving,intelligent transportation and other aspects.The automatic license plate recognition(ALPR)algorithm mainly includes two steps: license plate detection(LPD)and optical character recognition(OCR).License plate detection is the first step in the automatic license plate recognition algorithm and is extremely important.Currently,there are many fast and precise optical character recognition algorithms,and most license plate detection research still has some limitations,which lead to narrow applications.After studying the latest typical automatic license plate detection algorithms in academia,we find that some works,such as the unconstrained license plate detection method proposed by Silva,focus more on the accuracy of the license plate detection,and the speed is slow.Some works like the end-to-end fast license plate detection method proposed by Zhenbo Xu,pursue too much on speed,resulting in low detection accuracy and poor robustness.These two kinds of methods have certain limitations when be applied.Based on these,we study algorithms of automatic license plate detection and propose a fast automatic license plate detection algorithm in the wild.This algorithm includes a lightweight license plate positioning network and a lightweight license plate detection network,which can achieve a balance between speed and accuracy,and ensure the real-time output of license plate characters on the basis of robust license plate detection.The main work of this article is as follows:1.Aiming at the problem that the current research of license plate detection still can’t balance the speed and accuracy well.we propose a fast automatic license plate detection algorithm in the wild.We design a YOLO-like lightweight network LPLO(License Plate Location)for positioning license plates,which can quickly locate the approximate location of the license plate by regressing to a rectangular box convering all parts of the license plate.If the rectangular block image regressed by LPLO network is directly fed to the OCR network,the recognition error will be large due to character distortion and other information interference except for license plate.Thus,the rectangular image located by the LPLO network is fed into another lightweight network named CDC(Corner Detection and Correction),which is used to accurately detect the four corner coordinates of license plate and reduce the redundant information interference of non-license plate area.Then we use perspective transformation to correct the license plate part to the front view,to reduce the recognition error caused by character distortion.After testing,the results show that our detection method can get license plates accurately from most images.The method has high accuracy and robustness,and the accuracy of our method reaches 98.6%,and the target IOU with real label reaches0.894.It can detect an image of 720 P at a real-time speed of 33.7FPS.Compared with the existing typical license plate recognition algorithms such as the End-To-End method proposed by Zhenbo Xu and the two-step large inclination method proposed by Silva,our algorithm is more advantageous in comprehensively robustness and speed,which proves its practicability.(Corresponding to Chapter 3).2.Although the automatic license plate detection method proposed in work 1 has been achieved at a real-time speed of 33.7FPS,there is still room for improvement in the analysis of its algorithm and model structure.Therefore,we propose three methods to accelerate the above-mentioned automatic license plate detection algorithm.Firstly,the LPLO network takes up more time in the process of license plate detection.Preprocessing operation is added before LPLO network.By extracting the color information of the image,the region that may be the license plate is pre-screened.To increase the proportion of license plate in the whole image,reduce the size of the feature image,and reduce the calculation time of the model.Because the redundant information of the preprocessed image is reduced except for the license plate region,it is easier for the network to locate LP,so the backbone of the license plate location network(LPLO)is reduced to 11 layers.In order to make up for the loss of the backbone network,a parallel feedback branch and an information supplement branch have been added to redetect some wrong detection images,or extract the whole license plate region from the original image before preprocessing to correct some detection errors.The speed of the entire license plate detection module is increased by nearly 1.5 times,reaching 46.9FPS,and the accuracy is kept at 94.9%,and the Io U is kept at 0.88.Secondly,we use depth separation convolution instead of traditional convolution for calculation in convolution neural networks,which can speed up the operation by reducing the number of convolution calculation.After testing,our method can be improved by nearly 49% on the equipment with insufficient computing power.Thirdly,using the model quantization tool,we can reduce the storage time and computing time of the network by reducing the number of model calculation bits from float32 to int8.Thirdly,We use model quantification tools.By reducing the number of calculation bits of the model from float32 bits to int8 bits,the network storage and calculation time can be shortened(Corresponding to Chapter 4).