Research On License Plate Recognition Based On Deep Learning

Recently,more mature research results have been obtained for license plate recognition technology for specific environments.The license plate positioning,character segmentation and character recognition are the basic steps of the traditional license plate recognition and recognition algorithm.The dependence among these three steps is strong,and the error accumulation and even the error amplification are prone to occur.In a complex environment such as insufficient lighting or blurred license plates,license plate positioning errors and character segmentation errors can seriously affect the character recognition rate.The accuracy of license plate positioning plays a decisive role in the recognition rate of the entire license plate recognition algorithm.This thesis optimizes the existing target detection network,realizes the license plate location algorithm based on deep learning,and uses Opencv to implement the traditional license plate location algorithm.In this thesis,through the comparison of two license plate location algorithms,it is found that the license plate location algorithm based on deep learning has strong robustness.In license plate location based on deep learning,since the license plate can be used as a detection target,each license plate character can also be used as a detection target.But plate character region is smaller and more difficult to detect than license plate region.Therefore,based on this idea,this thesis combines deep learning technology to further research on license plate recognition and proposes an end-to-end license plate recognition method.This method firstly uses the deep neural network to extract the basic features,then uses a scale transformation strategy to extract the multi-scale features on the basic features,and finally inputs the multi-scale features to the position regression layer and the category classification layer to complete the location and recognition of the license plate and characters.The network model obtained after training as a whole can realize the end-to-end license plate recognition method.This method does not require step-by-step license plate location,character segmentation,and character recognition.Instead,it directly recognizes a license plate image,avoiding the effects of error accumulation and character segmentation.In this thesis,more than 2,000 license plate images in a complex environment have been collected,and training sets and test sets in four different environments have been produced by machine assisted and manually annotated methods.Finally,an end-to-end license plate recognition network has been trained and tested.Compares various traditional license plate recognition algorithms.Experimental results show that the end-to-end license plate recognition method proposed in this thesis has high recognition rate and strong robustness in a complex environment.