应用于监控视频中的多帧图像车牌识别系统

发布时间：2018-08-26 07:44

【摘要】：智能交通系统已经走进人们的生活,它广泛应用于收费站、停车场等诸多场景之中。车牌识别作为其中最为重要的部分成为一个研究热点,许多专家学者提出了优秀的识别算法。目前车牌识别技术已经相当成熟,对清晰车牌有着较高的识别率,但是一旦图像质量有所下降,识别率将会大大降低。车牌识别系统主要分为三大部分:车辆检测,车牌获取和字符识别。本文将对此展开深入研究。车辆检测部分,研究高效率的车辆检测算法。本文采用基于卷积神经网络的车辆检测算法,实现了从原始视频图片中自动截取并保存车辆图片,极大的降低了训练样本的获取成本。车牌获取部分,研究了图像灰度化,直方图均衡化,去均值以及车牌倾斜校正等预处理操作,通过预处理可以降低干扰因素,突出车牌有用信息,便于后续的识别。该车牌获取器可以方便快捷的从车辆图片中截取高质量的车牌图片。车牌算法的识别结果对于手动标点情况过于敏感,标点位置偏差极大的降低了车牌分割和识别效果。本文研究了两套标点优化算法,根据用户标点和图像信息,算法自动矫正车牌标点,进一步提高车牌的分割效果,最终提高车牌识别率和识别结果稳定性。车牌字符识别部分,研究了多帧字符识别算法。对于数字字母车牌字符,先通过稀疏自编码器提取字符的稀疏特征,再由支持向量机完成识别工作。对于中文字符,则通过费希尔判别准则的字典学习提取字符的残差信息,再利用softmax完成中文字符的识别。不同于常见的单帧车牌识别算法,本文利用车牌在监控视频不同帧中的多张图片共同参与识别,充分利用多帧图像间的相对信息和自身的图像信息。在单帧车牌识别的基础上设计两套多帧识别算法,分别为结果融合型多帧识别算法和特征融合型多帧识别算法。试验结果表明多帧识别对于较模糊车牌有着更高的识别率。
[Abstract]:Intelligent Transportation system (its) has come into people's life, it is widely used in many scenes such as toll station, parking lot and so on. As the most important part of license plate recognition, many experts and scholars put forward excellent recognition algorithms. At present, the license plate recognition technology is quite mature, and has a high recognition rate for clear license plate, but once the image quality is reduced, the recognition rate will be greatly reduced. License plate recognition system is mainly divided into three parts: vehicle detection, license plate acquisition and character recognition. This article will carry on the thorough research to this. In the part of vehicle detection, the efficient vehicle detection algorithm is studied. In this paper, the vehicle detection algorithm based on convolution neural network is used to automatically capture and save the vehicle images from the original video images, which greatly reduces the cost of obtaining training samples. In the part of license plate acquisition, the preprocessing operations such as image grayscale, histogram equalization, de-mean and license plate tilt correction are studied. Through the preprocessing, the interference factors can be reduced, the useful information of license plate can be highlighted, and the subsequent recognition can be facilitated. The license plate acquirer can conveniently and quickly intercept the high quality license plate image from the vehicle image. The recognition result of license plate algorithm is too sensitive to manual punctuation, and the deviation of punctuation position greatly reduces the effect of license plate segmentation and recognition. In this paper, two sets of punctuation optimization algorithms are studied. According to the user punctuation and image information, the algorithm automatically corrects the license plate punctuation, further improves the segmentation effect of the license plate, and finally improves the recognition rate and the stability of the recognition result. In the part of character recognition of license plate, multi-frame character recognition algorithm is studied. For the characters of alphabetical license plate, the sparse features are extracted by sparse self-encoder, and then the recognition is completed by support vector machine (SVM). For Chinese characters, the residual information of characters is extracted from the dictionary of Fisher criterion, and the recognition of Chinese characters is completed by softmax. Different from the common single-frame license plate recognition algorithm, this paper uses multiple images in different frames of the surveillance video to participate in the recognition, and makes full use of the relative information between the multi-frame images and their own image information. On the basis of single frame license plate recognition, two sets of multi-frame recognition algorithms are designed, one is result fusion multi-frame recognition algorithm and the other is feature fusion multi-frame recognition algorithm. Experimental results show that multi-frame recognition has a higher recognition rate than fuzzy license plate.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】