基于指数矩的车牌识别研究

发布时间：2018-03-31 01:33

本文选题：计算机视觉　切入点：指数矩　出处：《北京邮电大学》2017年博士论文

【摘要】：计算机视觉,正在从根本上改变我们的世界,以及我们每个人的生活方式。让机器之眼代替人眼,看懂我们的世界,实现智能化控制,解放人类的双手,是无数科学家梦寐以求的理想。视觉智能的潜在应用是无限的,人工智能几乎触及人类生活的各个方面,本文主要针对计算机视觉在智慧交通领域的应用展开研究,将指数矩的多畸变不变性运用到车辆追踪、车牌定位、以及车牌字符识别中,形成一套基于指数矩的车辆识别算法。指数矩的相关理论,是计算机视觉领域一个新的研究方向,不仅可以用于车牌识别,还可应用于常规的物体识别、场景识别等,因为指数矩的平移、缩放、和旋转不变性,对于目标物体的倾斜,远近变化,光照不足,天气恶劣等情况有很强的抗干扰能力,在不佳环境下依然具有很高的识别效率和准确度。基于指数矩的相关研究,对于未来智慧交通、智慧城市的建设有一定的价值。本文将指数矩作为图像的特征参数,对车辆追踪、车牌定位、及车牌字符识别展开了一系列研究,主要的研究工作和创新点有下列几个方面:(1)提出了基于指数矩的车辆跟踪算法。在反复的实验中,作者发现,在收费站和交通关卡,车辆相对严格的直线行驶,从某一固定点观察,车辆向远向近行驶可以视为连续的缩放变化,利用这一特征,作者将指数矩的缩放不变性运用到车辆跟踪中,提出了一种新的车辆跟踪算法。首先利用帧间差分法确定目标车辆,然后提取目标车辆的指数矩作为跟踪参数,通过不断调整搜索窗口的位置,实现多车辆的自动跟踪。较传统车辆跟踪算法,本文算法利用了指数矩的缩放不变性,降低了光照和天气对识别的影响,提高了跟踪的鲁棒性。车辆自动跟踪算法在收费站、交通关卡等有着广泛的应用前景。(2)提出了基于指数矩的车牌定位算法。本文提出了一种全新的车牌定位方法:基于指数矩特征的车牌定位方法。车牌定位是后续车牌字符识别的前提和基础,在车牌识别过程中具有至关重要的作用。本文将指数矩运用到车牌定位中,利用指数矩的平移、缩放、和旋转不变性,在车牌倾斜、车辆远/近变化、天气变化、光照不足等环境信息变化的情况下,依然具有良好的识别效果。本方法在不必进行倾斜校正、不必进行中心点调整以及比例调整的情况下,即可定位车牌,缩短了定位时间,具有良好的实际应用价值。(3)提出了基于指数矩和网格计算的车牌字符识别算法。作者根据车牌的字符形态学特征,对将车牌的所有字符分为12组:第1组是汉字组;第2-11组为形近的数字和字符组;第12组为模值无关组。将12个分组对应12个神经网络分类器,进行指数矩和网格特征训练,待处理的字符依次进入相应分类器,用指数矩特征进行初级分类,再利用网格特征进行第二次判定,确定最终的识别结果。该方法有效利用了指数矩的识别优势,同时,利用网格特征,弥补了指数矩由于旋转不变性在形似字符的判定中产生的误差。
[Abstract]:Computer vision, is changing our world radically, and the life of each of us. Let the eye machine to replace eyes, we understand the world, the realization of intelligent control, liberation of human hands, is a dream of many scientists. The potential application of visual intelligence is infinite, almost all aspects of artificial intelligence touch of human life, this paper focuses on the application of computer vision in the field of intelligent transportation research, the index of moment distortion invariance applied to vehicle tracking, vehicle positioning, and license plate character recognition, a vehicle recognition algorithm based on moment theory index. The index of moment, is a new research direction the field of computer vision, not only can be used for license plate recognition, can also be used in conventional object recognition, scene recognition, because the moment index translation, scaling, and rotation invariance, To tilt, the object distance changes, insufficient light, bad weather conditions and have strong anti-interference capability, in the poor environment still has high recognition efficiency and accuracy. The related research index based on moment, for the future of intelligent transportation, smart city construction has a certain value. This paper will index moments as the feature parameters of the image, the vehicle tracking, vehicle license plate location, license plate character recognition and launched a series of research, the main research works and innovations are as follows: (1) tracking algorithm is proposed based on the exponential moment. The author found the vehicle in repeated experiments, and at the toll station and crossing traffic the vehicle, relatively tight straight, from a fixed point observation, the vehicle to travel away from the past can be regarded as the continuous change of the zoom, using this feature, the author will use the zoom invariant moment index to vehicle tracking, Put forward a new vehicle tracking algorithm. Firstly using frame difference method to determine the target vehicle, then the exponential moment is extracted as the target vehicle tracking parameters, by adjusting the position of the search window, automatic tracking of multiple vehicles. Compared with the traditional vehicle tracking algorithm, this algorithm uses zoom invariance moment index, reduce the effects of light and weather on the recognition, improve the robustness of tracking algorithm. In the toll station, automatic vehicle tracking, traffic levels and has a wide application prospect. (2) proposed a license plate location algorithm based on the exponential moment. This paper proposes a new method of license plate location: license plate location method of exponential moments based on the characteristics. License plate location is the premise and basis of the license plate character recognition, has a crucial role in the process of license plate recognition. In this paper, the exponential moment applied to license plate location, using the exponential moment Pan, zoom, tilt and rotation invariance in the vehicle, far / near changes, changes in the weather, illumination changes such as lack of environmental information, it still has a good recognition effect. This method needn't tilt correction, without center adjustment and ratio adjustment, can be license plate positioning, shorten the positioning time and has good practical value. (3) proposed character recognition algorithm based on Grid Computing and the exponential moment. According to the morphological features of license plate characters, all the characters are divided into 12 groups: the first group is the 2-11 group is the group Chinese characters; and digital form the character group of nearly twelfth groups; independent group value model. The 12 groups corresponding to the 12 neural network classifier, exponential moments and mesh character training, the character in order to enter the corresponding classifier, the primary classification index of moment features, and then The second decision is made by grid feature to determine the final recognition result. This method takes advantage of the identification of exponential moment effectively, and makes use of grid characteristics to compensate for the error caused by the rotation invariance of the exponential moment in the determination of the similar character.

【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】