摘要
车牌定位及车辆识别是智能交通管理的主要研究问题。车牌定位识别,通过对图像进行预处理并结合形态学能粗略获取候选车牌位置,对符合特征的候选车牌进行筛选,精确获取车牌位置,最后采用神经网络完成字符识别过程。车辆识别采用迁移学习,采用AlexNet卷积神经网络构造出深度特征向量。形态学能够应对灰度底质量差的情形,为字符识别提供保障。车辆识别时对比直接分类图片特征,迁移学习构造的深度特征分类精度为85.13%,提高了38%,验证了迁移学习的有效性,通过KNN算法表明深度特征能够表征图片属性。针对新数据集重新提取特征、训练样本将消耗大量时间,对比迁移学习和AlexNet框架发现分类精度持平,表明了迁移学习的鲁棒性。
Location of license plate and recognition of vehicles arc main issues in intel ligent traffic management. Through image preprocessing and morphology, the locations of license plates arc roughly recognized） which then be fil-tered to obtain the accurate locations. A recognition process for characters on the plates is completed by utilizing a neural network. Transfer Learning is used for recognition of vehicles？ and a deep feature vector is developed by using an AlcxNct Convolutional neural network. Morphology can be used to process poor quality of grey background and guarantee accuracy of recognition. Compared with a direct classification of image features, the classification accuracy of the deep feature vec-tors constructed by Transfer Learning is 85.13% , which increases by 38%. It verifies the effectiveness of Transfer Learn-ing and image attributes that characterized by deep features based on the KNN algorithm. However,it takes time when rc-cxtracting features and training samples for new data sets. This method is found to have the equal accuracy of classi fi-cation when it is compared with Transfer Learning and AlcxNct framework） which proves the robustness of Transfer Learning.
出处
《交通信息与安全》
CSCD
2017年第4期63-68,共6页
Journal of Transport Information and Safety
基金
国家自然科学基金项目(51178157)
江苏省普通高校专业学位研究生创新计划项目(SJLX16_0154)资助
关键词
智能交通
形态学
车牌定位
车辆识别
机器学习
深度学习
迁移学习
深度特征
intelligent transportation
morphology
location of the license plate
vehicle recognition
machine learning
deep learning
transfer learning
deep features
