| License Plate Recognition is one of the most important fundamental technologies of Intelligent Transport System, it has been widely used in public video surveillance systems. In the last 30 years, great improvement has been made in license plate recognition technology. Both the accuracy and time cost of recognition algorithm have become acceptable in most of the normal application. However, the accuracy would become very low as in some extreme environment such as foggy or rainy day, strong light etc. The recognition algorithm still need to be improved.This thesis focuses on solving the problem of recognition accuracy in extreme environment condition. Bringing the DBN(Deep Belief Nets) algorithm into license plate recognition field, this thesis builds a DBN based method for license plate character recognition. This thesis designs the parameters of the DBN classifier by studying the property of license plate data, discusses the problems in training the classifier, including the training parameters for RBMs, error function and stop criteria for global training etc. We also describe the way to generate samples.Some experiment was made to compare the proposed method in this thesis with the SVM based method and the ANN based method. The experiment result shows that, Testing in blurred set, noise set, flaw set and inclined set separately, the DBN based recognition method get accuracy of 94.9%,95.4%,96.3% and 85.6%,which is better than other traditional method. It will greatly improve the accuracy of recognition system.This thesis also studies the optimization method for the training and recognition process. In RBM training process, we improve the momentum factor, speeding up the training process by 16.6%. As to the recognition process, we adjust the recognition task to running in parallel. After the parallel optimization with 4 cores, the recognition progress speed up by 2.5 times, leading time cost drop to 19 ms, which is sufficient for engineering application. |
PDF Full Text Request