Research On Vehicle License Plate Recognition Algorithm Based On Deep Learning

In recent years,with the improvement of computer computing power,the introduction of convolutional neural networks,image segmentation and recognition have also developed rapidly.In terms of image segmentation,problems such as image distortion,missing and multiple noise caused by bad weather will cause problems such as missing characters and misalignment when the object detection model pre-selection frame outlines the license plate area.In addition,in terms of license plate recognition,traditional filtering and morphological denoising,in the case of too high image noise,there will be no deeper denoising of image details,resulting in reduced image quality after denoising and reduced recognition rate.In view of the proposed problem,in terms of segmentation algorithm,this paper uses the U-net segmentation model as a backbone network instead of the object detection model to segment the edge of the license plate in detail to reduce the segmentation problem caused by pre-selection box selection.Moreover,on the basis of U-net network,combined with adaptive weight feature pyramid and attention mechanism,the neck layer and head layer of the network are improved to improve the model performance.In terms of text recognition,because bad weather is mostly uniform noise,and the noise generation process of the diffusion model is in line with the uniform distribution,according to this characteristic of the model,the diffusion model is used to add noise processing to the image,simulate the noise generation process,train different levels of noise conditions,and rely on the trained "generation noise" to reverse denoising the image,so as to restore more image details.At the same time,considering that the traditional character segmentation method text recognition will cause certain errors in the results,this paper selects the SAR model to implement the text recognition module of the article,and treats the text information as a whole to reduce the error.After a large number of experiments,combined with the model used in this paper for license plate recognition,the license plate information can be efficiently and accurately identified,and the accuracy rate of recognition under normal circumstances is as high as 97.2%,and in bad weather,the license plate recognition rate also reaches 92.8%,and the accuracy of this model is 1.8% higher than that of mainstream single-stage detection algorithms,and the accuracy of traditional character segmentation methods is about 3%.