Design And Implementation Of QR Code Recognition System Based On ZYNQ

QR code automatic recognition technology is a crucial technology in various domains including industrial Internet of Things,manufacturing,logistics management,and consumer industries.Traditional QR code recognition products employ image processing algorithms like pattern recognition and wavelet transform.However,these products often face challenges such as interference from other image elements,as well as issues related to blur and noise,leading to low accuracy rates.In recent years,the advancement of deep learning in image processing has significantly enhanced the accuracy of QR code recognition in complex scenarios.The main research work in this thesis encompasses the following three aspects:1、Firstly,the system principle is analyzed to clarify the overall system design scheme,which includes the target detection algorithm,image deblurring algorithm,and system implementation and testing.In order to achieve a cost-effective and highly customizable solution,the ZYNQ platform is selected for hardware design,along with the integration of the OV5640 camera.Moving on to the software design phase,a three-level detection mechanism is devised to ensure fast recognition speed.To enhance recognition performance,Vitis AI is employed for quantization and compilation of deep learning models.Lastly,the Linux system is ported,the algorithm is deployed,and the complete system is implemented on the ZYNQ platform.2、Design the algorithms in the system.Firstly,the QR code target detection algorithm is designed.After introducing the YOLOv3 algorithm principle,two improvement methods are incorporated: dataset Masic data augmentation and Kmeans prior bounding box aggregation.In the experimental section,the original algorithm and the improved algorithm are tested for performance using a custom QR code dataset.The results from the test set demonstrate that the improved method achieves mAP values of 0.99 and 0.89 for the two types of target objects,whereas the original method achieves mAP values of 0.93 and 0.84.These results indicate that the improved YOLOv3 algorithm exhibits superior performance,thus it is selected as the pre-processing algorithm for QR code recognition.Next,the QR code image deblurring algorithm is designed,based on the principles of image deblurring,to eliminate Gaussian blur and motion blur from the image.This method is trained using theGAN(Generative Adversarial Network)model framework,with Mobile Netv1 from DeblurGANv2 as the generator,and PatchGAN as the discriminator.To effectively remove Gaussian blur in QR code images,the relative entropy loss function is incorporated.Experiments are conducted using a custom QR code blurred dataset for training and testing,comparing the proposed method with DeblurGANv2 original method,SRN,and CycleGAN.The results indicate that the proposed method achieves a PSNR of 22.68 dB and an SSIM of 0.866 dB between the deblurred images and the original high-quality images in the test set,outperforming the other three comparison algorithms.3、The first step involves completing the hardware circuit design and porting the Linux system onto the ZYNQ platform.Subsequently,the deep learning algorithms are quantized and compiled using Vitis AI.Finally,the model is deployed to the ZYNQ platform,and the system implementation and testing are carried out.The results demonstrate that the system achieves a recognition rate of 85.9% on the complex environment test set,with a single image recognition speed of 92 ms.These outcomes meet the system’s requirements for both recognition rate and speed.