基于神经网络的机动车号牌字符识别

以定位、 分割后的机动车号牌字符为研究对象. 首先, 对机动车号牌图像进行大小、 灰度方差、 灰度均值的标准化处理. 其次, 根据机动车号 牌字符的特点, 抽取字符3种不同的特征, 构造3个BP神经网络对机动车号牌字符进行识别. 并根据字符在机动车号牌中所处位置的差异, 在每个BP神经网络中构造4种不同的子网络分 别进行训练和识别. 最后, 每个BP神经网络的输出通过加权求和的组合方式得到最终识别结 果. 在组合各网络输出前, 采用对字符图像求取局部二阶差分值的方法, 将字形相近的字符 进行再分类, 从而有效地降低误识率. 通过分析实验结果, 表明本算法在机动车号牌识 别应用中达到了理想的识别效果.

Character Recognition of Vehicles' License Plates Based on Neural Network

Aiming at located and partitioned characters on motor vehicle plates. First, we standardized the size, variance of gray scale and mean value of gray scale of the motor vehicle plate images. Second, according to the characteristic of characters on the plate, we selected three different characteristics, and constructed three BP nerve nets (NN) to identify the characters on the motor vehicle plate. At the same time, for the position differences of characters on the plate we constructed and trained four different subnets to recognize characters in every BP NN. Finally, we used weighted summation of out-up of every BP NN to get the final result. Furthermore, before combining the result of each BP NN, we calculated part two-rank difference in the (characters') image to secondly classify the similar characters in the shape, and so that we can decrease the (wrong) identification rate. According to the analysis based on the results of the experiments, it is demonstrated that the algorithm has a perfect identification effect.