Traffic Sign Character Recognition Based On The Deep Learning

Traffic signs provide the information to drivers as natural scene in the daily life. As an important reference in navigation, traffic signs are becoming more and more essential when cars are widespread. Traffic signs usually set on the top and right of the road. Drivers need to search and identify traffic signs' information, and will easily be absentminded. For drivers who are not familiar with the road system in city, they may slow down their vehicle to get the information from traffic signs. Suddenly slowing down vehicle may cause accident. Detection and Recognition information from traffic signs may help achieve the hub displaying traffic signs' information in car.Great progress have been achieved on machine learning, such as machine vision, natural language processing, voice recognition and image recognition, therefore machine learning becomes an important branch of artificial intelligent and new research direction. As a new method in machine learning, deep learning achieves great progress in research and commercial aspect. Deep Boltzmann Machine as a method in deep learning has advantage in ambitious data processing. It may be good at recognition in natural image recognition.From above, the model constructed by Restricted Boltzmann Machines is introduced in this thesis, and is based on the application in traffic signs character recognition. A deep learning method based on Mixing optimization Deep Boltzmann Machine is proposed.Firstly, for the traffic signs' images, the color space, edge detection, projection, connected component methods are analyzed. Traffic signs' character extraction method based on color space and panel space information is proposed.Secondly, this dissertation introduces the history of deep learning, the thought of deep learning, and the fundamental unit, Restricted Boltzmann Machine. And then, this dissertation introduces the deriving and training method of restricted Boltzmann machines. In addition, the deep Boltzmann machine method is introduced.Thirdly, the training method is essential in deep learning. The training method of Deep Boltzmann Machine is introduced in this thesis, which includes the pre-training method and the fine-tuning method. The pre-training method is an unsupervised method, and influences the parameter adaption and the fine-tuning algorithm. As a supervised method in Deep Boltzmann Machine, the fine-tuning method, conjugate gradient method, is introduced in dissertation.Fourthly, in order to improve the recognition accuracy of the Deep Boltzmann Machine, and based on the thought of sampling accuracy and fine-tuning improvement, a method called the mixed optimization pre-training is proposed.From the experiments, the Mixed optimization Deep Boltzmann Machine method in character recognition is better than the original Deep Boltzmann Machine.