License Plate Recognition System For Multi-frame Images In Surveillance Videos

Intelligent transportation system has come into people's lives.It is widely used in parking management,high-speed toll management,etc.Vehicle plate recognition becomes a research hotspot as the most important part of the intelligent transportation system.Many researchers have conducted in-depth study on it.The vehicle plate recognition has been very mature and have very excellent effect on clear license plate.However,when the image quality decline,the recognition rate is too bad.There are three parts in the vehicle plate recognition: vehicle detection,vehicle license plate acquisition and the vehicle plate recognition.Here are the specific introduction:In vehicle detection,we research the high efficient vehicle detection technology to reduce the cost when we make the training vehicle plate samples.Vehicle plate recognition needs thousands of vehicle plate samples.It consumes too much time if we only cut them by hands.This paper uses convolutional neural network to realize the vehicle detection and it greatly reduce labor costs.In vehicle license plate,vehicle plate characters should be split precisely.So we use manual punctuation to get the vehicle plate from the vehicle figure.Then the divisible vehicle plate characters are pretreated by graying,histogram equalization,zero-mean and inclination correction according to the points users marked.It can eliminate the interference factors and highlight useful information.In this way,we can get thousands of the high-quality vehicle license plate samples.In test phase,the recognition results may be sensitive to the points we marked.When users mark the incorrect points,it effects the vehicle plate division and reduces the recognition rates.We make two points-optimization systems.These systems can find the best points according to the vehicle plate pictures and the points users marked.It improves the division quality and increases the recognition rates.Vehicle character recognition is the key of the vehicle plate recognition.In this phase,we study the multi-frame vehicle character recognition.About the number and letter vehicle plate characters,we use sparse auto encoder to get the sparse features of the characters.Then we combine the features using the support vector machine.About the Chinese characters,we can get the features using the dictionary learning models and then combine these features by softmax.Based on the single frame vehicle character recognition.This paper we combine the different vehicle pictures from different frames on the same vehicle to recognition the vehicle characters and use the information between them fully.We design two multi-frame vehicle recognition namely result fusion recognition and character fusion recognition.Experiment results show that multi-frame recognition can earn higher recognition rates than single-frame recognition.