基于CUDA的快速车牌字符识别

FAST LICENSE PLATE CHARACTERS RECOGNITION BASED ON CUDA

传统的车牌识别研究主要目的是提高识别准确率。利用CUDA技术在准确率不降低的情况下实现识别速度的提高。为此,对常用的SVM分类方法进行改进,使其能够在GPU上实现并行计算,再利用改进后的SVM训练和预测车牌字符数据。实验结果表明,相对于运行在CPU上的LIBSVM方法,经过改进的在GPU上运行的SVM方法能够带来1-30倍训练速度和50-72倍预测速度的提高,且随着样本数量的增加,加速效果会更加显著。

Traditional research of license plate recognition, mainly focuses on the improvement of identification accuracy; in this paper we will do some research on improving the recognition speed while guaranteeing the same accuracy. For this purpose, the common SVM classification method is improved to realise parallel computing in GPU. Then, the improved SVM is applied to license plate characters for training and predicting. Experimental results show that the improved SVM running on GPU is able to bring enhancement in 1 - 30 times faster training speed and in 50 -72 times faster prediction speed respectively than the popular solver LIBSVM on CPU, and the acceleration effect will be significantly enhanced with the increase of sample size as well.