基于R-ELM的实时车牌字符识别技术

Real-time license character recognition technology based on R-ELM

提出新的实时车牌字符识别技术.该技术对分割后的字符图像进行灰度化,归一到固定像素大小,采用R-ELM算法进行训练.该技术的优点在于能够采用较小的样本集,快速达到理想的识别率.实验数据显示,与传统的BP算法相比,效率能够提高2～3个数量级.为了有效地挖掘GPU的运算能力,系统采用弹性队列与动态符合调整方法,将字符数据组合成数据包,保证在使用图形处理器（GPU）进行识别的过程中,运算效率最大化.实验显示,与CPU相比,能够得到近2个数量级的速度提升.通过对大量实际样本图像的测试,采用该方法获得了良好的识别及加速效果.

A new real-time recognition technology was provided for license plate character.Character image segmentation was normalized to a fixed pixel size after gray processing,and then R-ELM algorithm for training was applied.The desired recognition accuracy was rapidly achieved by using comparatively smaller sample sets.Experimental results show that recognition efficiency can be improved from 100to 1000times compared with the traditional BP algorithm.To apply the GPU technology more effectively,the system assembles the character data into packets by using elastic queue and dynamic compliance adjustment method which can acquire maximum efficiency in the recognition process.Experiments indicated that recognition accuracy was raised nearly 100times compared with CPU.By testing of a large number of practical sample images,the method can obtain a good recognition and accelerated result.