Research On License Plate Recognition Algorithm Based On Deep Learning

License plate intelligent recognition is the core part to realize intelligent traffic management.It requires comprehensive application of related technologies such as artificial intelligence,image processing,computer vision,pattern recognition and deep learning.This thesis makes full use of the powerful computing and graphics processing functions of the computer to establish a license plate recognition algorithm model based on deep learning target detection.It has great theoretical significance for the intelligent detection and recognition of license plate images.This thesis analyzes the technical difficulties of license plate recognition at present and the diversity and uniqueness of Chinese license plates.The license plate recognition algorithm based on deep learning is proposed.The license plate recognition is divided into two parts: license plate detection and recognition.Owing to the characteristics of vehicle image and license plate image,different recognition algorithm models are designed to realize the intelligent recognition of license plates.The license plate detection and recognition algorithm use Inception-Res convolutional neural network for image feature extraction,and then use the SSD target detection algorithm to locate and identify based on the extracted feature map.In order to improve the robustness of the algorithm,combined with the characteristics of the vehicle image and the license plate image,the image input sizes of the two algorithms are respectively 300*300 and 224*88,and pre-selected boxes of different sizes and proportions are set.In the algorithm model training process,the data set was in the rule of VOC format,and the data set is manually labeled;in order to improve the quality of the data set,manual cleaning work is performed;color transformation,scaling transformation,translation transformation,random clipping,Gaussian are adopted.Data enhancement methods such as noise enrich the diversity of data,and use category reorganization algorithms to solve the problem of license plate data imbalance.The data set was trained by different network structures,what's more,a new image data set was used for comparison experiments.Finally,a license plate recognition system was developed based on the best-performing network model.Compared with the traditional license plate recognition algorithm,the license plate recognition algorithm based on deep learning not only has a high convergence speed of the network model convergence speed,but also has the high recognition accuracy and faster recognition speed.It is important for the license plate image and the character is blurred.The algorithm also has a good recognition effect in complex natural environments such as unclear and weak illumination.After the recognition algorithm is engineered,the license plate recognition can reach 98.9% accurately,and the license plate recognition speed of the average single vehicle image is 300 ms.