基于级联卷积神经网络的车牌定位  

License Plate Location Based on Cascaded Convolution Neural Network

作　　者：傅鹏 谢世朋[1] 

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《计算机技术与发展》2018年第1期134-137,共4页Computer Technology and Development

基　　金：江苏省科技重点研发计划-产业前瞻与共性关键技术(BE2016001-4);教育部-中国移动科研基金(MCM20150504)

摘　　要：针对多车辆、低分辨率等复杂环境下的车牌定位问题,提出了一种基于人眼视觉特性的车牌识别方法。通过模仿人眼视觉原理,利用级联卷积神经网络分层提取目标区域特征,逐步缩小搜索区域的方法,实现车牌的精准定位。首先通过运动目标检测算法定位出目标运动热点区域;然后使用卷积神经网络识别热点区域中的车辆;最后使用卷积神经网络从定位的车辆图片中识别车牌。数据集采集于多个交通路口的天网摄像头,然后对5 000幅图像,约15 000个目标进行人工标注,同时对训练图片进行随机变换,从而提高训练的有效性。实验结果表明,通过提取运动区域可提升卷积神经网络运行的速度和识别的精度。相比于传统车牌识别算法,提出的方法极大地提高了复杂场景下的车牌识别率,同时在处理高分辨率的图片时具有更高的车牌定位率。Aiming at the problem of license plate positioning in complex environments such as multi-vehicle and low resolution, we present a license plate recognition method based on human vision. By imitating the visual principle of human eyes, the precise positioning of license plate is realized by the approach of extraction of target region characteristics through cascade convolution neural network and gradually narro- wing the search area. Firstly ,the target motion region which we are interested in is located by the motion detection. Then ,the vehicle identifi- cation is performed on the hot spot region by convolutional neural network. Finally,license plates are located in vehicle picture. Training pic- tures are collected in 20 different traffic junctions of the skynet camera images,as well as nearly 5 000 images and about 15 000 targets la- beled by manual. At the same time, the labeled images are transformed randomly to improve the effectiveness of the training. According to the experiments,the extraction of motion region enhances the speed and recognition precision of convolutional neural network, and greatly improves the license plate recognition rate in complex scenes compared to the traditional license plate recognition algorithm. Moreover, it per- forms better in dealing with high-resolution pictures.

关 键 词：车牌定位 运动目标检测 视觉特性 卷积神经网络 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]