基于易混淆字符集神经网络的车牌识别算法研究

发布时间：2018-12-14 16:20

【摘要】：随着21世纪智能交通和信息技术的发展,计算机辅助人们处理交通问题成为了科学家不断研究的方向。我国交通流量随着汽车销量增长而日益增大,通过交警使用传统方法处理道路事故和违章已经变得不切实际。智能交通技术通过引入先信息技术、控制技术和计算机技术,形成一套先进的智能交通系统。目前,诸如公交GPS控制系统、车辆追踪系统、车辆信息管理系统、ETC不停车电子收费系统等等,都属于当前智能交通子系统的应用。而作为车辆信息管理系统中的关键一环,为了有效、快速的判别车辆身份,车牌识别系统成为了研究者不断改革创新的一个部分。车牌识别系统主要依靠计算机图形处理技术、模式识别技术、智能计算技术,将车牌图片从视频流得到的图片中提取出来,并依次进行字符的分割与识别。车牌识别技术至今依然存在许多困难,如车牌抓取、车牌去噪、字符识别、系统性能要求等等困难。在一个标准4米高卡口上抓取的1080P图片中,车牌部分仅仅占有约120*35像素大小。并且道路卡口照片中存在着大量自然背景和多车辆等等干扰因素。车牌图片噪声很多,包括过曝光电、柳丁、车牌边框、车牌污损等等噪声影响。字符识别成功率更是由上面两步成功率和预处理效果以及识别方法所决定,并且系统对实时性和准确性都有明确的要求。本文主要通过对国内外最近方法研究,通过选取合适的车牌提取、字符分割、字符识别算法,在原始方法上加以改进,使得识别时间平均在300ms以内,数字及英文字符正确率在95%以上,车牌识别率在80%左右。论文主要进行了以下方面的研究与改进： (1)车牌提取方面：在Sobel算子垂直方向边缘检测后运用纵向噪声与横向噪声消除方法去除大部分车体噪声和环境噪声,运用改进型的二值图像快速矩化算法将候选车牌位置标识出,运用车牌矩形特征和颜色特征得到正确的车牌位置。 (2)字符分割方面：在灰度拉伸、二值化得到较为清晰的车牌图片后,运用字符高度逼近方法去除大部分车牌的上下边框和柳丁,在运用简化投影特征和车牌模板特征,以及采用一定容错算法将字符正确分割。 (3)字符识别方面：通过对比两种字符图片处理和特征提取方法,最终选择粗网格特征与投影特征提取方法进行字符特征提取。通过采用两个结构简单的神经网络,分别用来识别汉字,英文与数字,并且通过易混淆字符判别神经网络区分易混淆字符。这种方法可以提高字符识别率与字符识别速度。
[Abstract]:With the development of intelligent transportation and information technology in the 21st century, computer aided people to deal with traffic problems has become the research direction of scientists. The traffic flow in China is increasing with the increase of car sales. It has become impractical to deal with road accidents and violations by traffic police. Intelligent transportation technology forms an advanced intelligent transportation system by introducing information technology, control technology and computer technology. At present, such as bus GPS control system, vehicle tracking system, vehicle information management system, ETC non-stop electronic toll system and so on, all belong to the current application of intelligent transportation subsystem. As a key part of the vehicle information management system, in order to identify the vehicle effectively and quickly, the license plate recognition system has become a part of the innovation and innovation of the researchers. The license plate recognition system mainly relies on computer graphics processing technology pattern recognition technology and intelligent computing technology to extract license plate images from images obtained from video stream and to segment and recognize characters in turn. There are still many difficulties in license plate recognition technology, such as license plate capture, license plate denoising, character recognition, system performance requirements and so on. In a 1080P image captured on a standard 4 m high bayonet, the license plate is only about 120 pixels in size. And there are a lot of interference factors such as natural background and multiple vehicles in the road bayonet photos. License plate images have a lot of noise, including overexposure, Liuding, license plate frame, license plate fouling and other noise effects. The success rate of character recognition is determined by the success rate of the two steps above, the effect of preprocessing and the method of recognition, and the system has clear requirements for real-time and accuracy. This paper mainly through the domestic and foreign recent method research, through the selection suitable license plate extraction, the character segmentation, the character recognition algorithm, in the original method carries on the improvement, causes the recognition time to average within 300ms, The correct rate of digital and English characters is more than 95%, and the recognition rate of license plate is about 80%. The main research and improvements in this paper are as follows: (1) license plate extraction: after the vertical edge detection of Sobel operator, most of the car body noise and environment noise are removed by using longitudinal noise and transverse noise elimination method. The candidate license plate position is identified by the improved binary image fast moment algorithm, and the correct license plate position is obtained by using the rectangular and color features of the license plate. (2) character segmentation: after grayscale stretching and binarization to get clear license plate images, the method of character height approximation is used to remove the upper and lower frames and Liuding of most license plates, and the simplified projection features and license plate template features are used. And a certain fault-tolerant algorithm is used to segment the characters correctly. (3) character recognition: by comparing two methods of character image processing and feature extraction, the coarse mesh feature and projection feature extraction method are selected for character feature extraction. Two neural networks with simple structure are used to recognize Chinese characters, English characters and numbers, and to distinguish easily confused characters by neural networks. This method can improve character recognition rate and character recognition speed.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【参考文献】