A License Plate Recognition SoC System Based On Software And Hardware Co-design

Since the new century,artificial intelligence technology is developing rapidly,which has changing the social industrial production and human life constantly.The typical example is license plate recognition technology.License plate recognition first obtains the recognition result by collecting the license plate image,processing it through a special image processing system,and extracting the target features.This technology is currently widely used in traffic management,police management,intelligent community management,automated charging and other fields.At the same time,license plate recognition technology faces new challenges.New energy vehicles are gradually becoming popular,but new energy vehicle license plates are significantly different from traditional license plates in terms of license plate size,license plate background color,and number of characters.The license plate recognition system needs to be able to recognize hybrid license plates uniformly;at the same time,with the increasing complexity of the license plate recognition algorithm and the improvement of image resolution,it is more and more difficult for the pure ARM processor to complete the recognition task.Therefore,it is necessary to deploy the license plate recognition algorithm on the processor system and programmable logic resources based on the software and hardware co-design,so as to achieve the hardware acceleration of the algorithm and build a more powerful license plate recognition system.This article first introduces the overall workflow and principle of the license plate recognition system,and divides the license plate recognition system into five modules:image＿capture,image＿preprocess,license＿plate＿location,character＿segmentation and character＿recognition.Also introduced the hardware and software co-design,highlighting the principle of classification of hardware and software modules.Based on this,by analyzing the characteristics of different modules in the license plate recognition system,hardware design is adopted for image＿preprocess module and character＿recognition module,and software design is adopted for license＿plate＿location module and character＿segmentation module.Based on the results of software and hardware division,the AXI bus is used to realize the interconnection between PS and PL,and a complete license plate recognition SoC system is constructed.The hardware design module in the license plate recognition SoC system is based on Verilog for RTL-level circuit design.The image＿preprocess module is based on the deep pipeline,which realizes the circuit design of gray process,median filter,image enhancement and binarization process;The character＿recognition module is implemented based on CNN hardware accelerator,Chinese characters and numbers&letters are recognized by the improved Le Net-5 neural network with different network parameters and the same network structure,and the circuit realization of convolutional layer,pooling layer,fully connected layer and classifier is completed.The software design module in the license plate recognition SoC system is implemented based on the embedded Linux system running Open CV function library.The license＿plate＿location module first divides the license plate image into different connected domains,and then realizes the license plate region extraction through prior knowledge,and finally obtains the license plate image containing only character information through the license plate tilt correction and boundary positioning;The character＿segmentation module first distinguishes the license plate type,and then,different templates are used for character segmentation of the mixed license plate,and the character boundary is adjusted by evaluating the segmentation effect,and finally the single character image is standardized to a size of 32x32.Finally,a license plate recognition SoC system was built based on the Zedboard hardware development platform,and 600 different types of license plate test samples were used for functional testing and performance testing.The functional test results proved that the license plate recognition SoC system can complete the unified recognition task of mixed license plates.The performance test results show the license plate recognition SoC system has better picture resolution,and the accuracy of the license plate recognition with a small angle is higher.The system running time is reduced by more than 50% compared with the pure software implementation.The final recognition accuracy of the license plate recognition SoC system is 95.1% and the time is 0.821 s,which meets the requirements of the project.