复杂自然环境下车牌识别算法研究

发布时间：2018-09-06 14:24

【摘要】：车牌识别技术是智能交通系统中的重要组成部分,是计算机视觉、图像处理与模式识别在智能交通领域的重要研究课题之一。但在实际环境下采集到的车牌图像,容易受到光照变化、尺度变化、目标干扰等诸多不利因素影响,因此在复杂多变的自然下识别车牌仍然是一个十分具有挑战的课题。车牌识别技术主要解决车牌的定位、分割、识别三个问题。本文分别对这三个部分进行了研究,并提出了相应算法。本文提出了一种基于目标区域的车牌定位算法,采用逐步求精的定位策略。该算法适用于光照变化、尺度变化和目标干扰等复杂的自然环境。本文引入了Selective Search算法对输入图像进行目标区域提取,根据车牌特征筛选出车牌候选区域,并通过一个预训练的支持向量机对候选区域进行判别,保留车牌区域。对获得车牌区域进行非极大值(NMS)抑制剔除重合区域。最后精确定位到车牌位置。本文提出了一种基于连通区域的字符分割算法。该算法首先对输入车牌进行预处理和倾斜校正,结合连通区域标记法和数学形态学处理法获得字符区域。同时,本文对传统的字符归一化方法进行了改进,有效的解决了由字符归一化造成的字符形变的问题。本文提出了一种基于卷积神经网络的车牌字符识别算法,设计了两个卷积网络NET1和NET2,其中NET1用做识别汉字,NET2用做识别字母和数字。本文引入了 rectifier作为神经元的激活函数,并使用mini-batch随机梯度下降法训练网络,可以加速目标函数的收敛。采用卷积神经网络可以从输入的字符图像中自动提取出图像特征,并进行分类,从而获得识别结果。在整个过程中,不需要手动选定图像特征或对图像作局部处理。实验表明,本文算法可以有效的在复杂自然环境中定位车牌,分割字符和识别字符。并将本文算法与同类型算法做了比较,均有显著的提升。
[Abstract]:License plate recognition technology is an important part of intelligent transportation system. It is one of the important research topics of computer vision, image processing and pattern recognition in the field of intelligent transportation. However, the license plate images collected in the actual environment are easily affected by many unfavorable factors, such as light change, scale change, target interference and so on, so it is still a challenge to recognize the license plate in the complex and changeable nature. License plate recognition technology mainly solves three problems of license plate location, segmentation and recognition. In this paper, the three parts are studied, and the corresponding algorithms are proposed. A license plate location algorithm based on target region is proposed in this paper. The algorithm is suitable for complex natural environment, such as illumination variation, scale change and target interference. In this paper, the Selective Search algorithm is introduced to extract the target region of the input image, and the candidate region is selected according to the license plate feature, and the candidate region is identified by a pre-trained support vector machine to preserve the license plate area. Non-maximum (NMS) suppression is used to eliminate the coincidence area of the obtained license plate. Finally, the location of the license plate is accurately located. In this paper, a character segmentation algorithm based on connected region is proposed. The algorithm firstly preprocesses and corrects the input license plate, and combines the connected region marking method and the mathematical morphology processing method to obtain the character region. At the same time, the traditional method of character normalization is improved, which effectively solves the problem of character deformation caused by character normalization. A license plate character recognition algorithm based on convolution neural network is presented in this paper. Two convolution networks NET1 and NET2, are designed in which NET1 is used to recognize Chinese characters and NET2 is used to recognize letters and numbers. In this paper, rectifier is introduced as the activation function of neurons, and the mini-batch stochastic gradient descent method is used to train the network, which can accelerate the convergence of the objective function. By using convolution neural network, the image features can be automatically extracted from the input character images and classified, and the recognition results can be obtained. In the whole process, there is no need to manually select image features or make partial image processing. Experimental results show that the proposed algorithm can effectively locate license plates, segment characters and recognize characters in complex environments. This algorithm is compared with the same type algorithm, and has a significant improvement.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】