Research On High Efficiency License Plate Recognition In Complex Scenes

In recent years,with the development of the economy,China's car ownership has grown rapidly.How to quickly obtain license-plate information has become a difficult problem for the traffic management department.In this regard,an algorithm with strong adaptive ability,fast detection speed and high precision is the research and development direction of the license plate detection algorithm.Although the traditional license-plate detection algorithm has been put into practical use,the detection of the traditional license-plate detection algorithm is not effective,and the traditional detection algorithm is difficult to adapt to a variety of scenarios.The deep learning algorithm has good image feature representation ability and can be applied to complex scenes,but the detection process is time consuming.Based on the above reasons,we have done a lot of research on license-plate recognition technology.The license-plate recognition algorithm in this paper has the following improvements:1)We design a license-plate location system based on Fully Convolutional Networks for Semantic Segmentation(FCN)and Faster R-CNN modules:For license-plate recognition system,the segmentation module FCN and detection module Faster R-CNN are effectively combined for license-plate location.These two CNN stems are designed for feature extraction in FCN and Faster R-CNN modules,respectively.The FCN module is employed to detect the license-plate candidate region while the Faster R-CNN module is used for the license-plate detection via obtaining the specific location.Due to the proposed network is small,so the trained model is also small and the speed is greatly improved under ensuring the accuracy,which achieves the real-time requirements.2)We optimize training procedure:Because of the problems in license-plate recognition task,such as low recognition rate caused by illumination,false recognition of fuzzy images.In the paper,a classifier is employed to classify the license-plate samples,and to put the negative samples into the negative sample set again to obtain more identifiable feature representation,which improved that the recognition ability of the model to complex scenes and complex images effectively.3)We reasonably improved Non Maximal Suppression(NMS)module:During the process of license-plate detection,when several license-plates are found to be close to each others,the NMS module can directly delete all the candidate boxes,whose coincidence degree is highest,To solve the problem of partial license-plate omission,we reduce the license-plates candidate box's confidence instead of deleting the candidate box directly.This algorithm retains the original deleted candidate box,reduces the fall-out ratio under the condition that multiple license-plates are close to each other.It improves the result of license-plate detection effectively.At the same time,the license plate recognition algorithm was successfully applied to the commercial project of Chongqing Kaiser Company...