基于深度学习的车牌识别技术研究

发布时间：2017-12-30 23:51

本文关键词：基于深度学习的车牌识别技术研究　出处：《青岛科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：车牌识别是智能交通系统中极为重要的一部分,在智慧城市的理念当中也不可或缺,具有较高的研究与应用价值。虽然经过长时间的研究和努力,我国的车牌识别技术也取得了优秀的研究成果,能够解决一般场景下的车牌识别问题,如交通固定卡口、车库出入口和小区门禁等场景。但在自然场景下出现的车牌倾斜、车牌扭曲、光照条件较差、像素分辨率较低等情况,不能够准确的进行车牌定位与识别。针对这些问题,本文使用基于深度学习的目标检测方法进行以下几个方面的研究。首先,本文提出使用基于深度学习的目标检测方法对车牌进行定位,并对目标检测的卷积神经网络结构进行改进。在模型的训练过程中,使用图像标注软件对现场采集到的30521张图片进行手工标注。为了增加训练样本的数量,随机对图像进行镜像和缩放等操作。使用训练好的模型在多个不同交通卡口采集的数据进行测试。将测试结果与使用灰度图像进行车牌定位的方法进行比较,验证了基于深度学习的目标检测方法在车牌定位方面的鲁棒性。其次,同样使用基于深度学习的目标检测方法对车牌字符进行检测,并针对内地车牌和港澳台车牌对检测结果分别进行相对应的排序处理,最终得到车牌识别结果。在模型的训练过程中,本文使用图像标注软件对21269张图片进行手工标注。由于车牌中的字符具有不对称性,则不对图像做增强处理。使用不同的网络结构对该数据集进行训练、测试,并使用现场采集到的图像数据进行对比验证,选取最优网络模型。最后将车牌定位和车牌识别过程相结合,统一测试整个系统的识别速度。测试结果显示识别速度为每张图片0.12秒,比传统的车牌识别系统有很大的优势。另外该系统可以与车辆检测、车型识别完美结合。待检测图片先进行车辆检测和车型识别,再进行车牌的检测与识别。这些结果可以通过搭建大数据平台,实时上传检测到车辆的车型、车牌和位置等信息,对建设智慧城市、实现万物互联具有重大意义。
[Abstract]:License plate recognition is one of the most important part of the intelligent transportation system, is indispensable in the concept of smart city, has a high value of research and application. After long time research and efforts, the license plate recognition technology in China has also made outstanding research results, can solve the problem of general license plate recognition scenarios. Such as traffic fixed bayonet, garage entrance and residential access scenarios. But the license plate appears in natural scene tilt, plate distortion, poor light condition, low resolution pixel, can not carry out license plate location and recognition accuracy. Aiming at these problems, this paper uses the following research methods to detect deep learning based on the target. First of all, the paper proposes using the detection method of deep learning targets based on the license plate location, network structure and convolution neural network on target detection of The improvement in the training process. In the model, 30521 images using image annotation software to the collection of field of manual annotation. In order to increase the number of training samples, random image and zoom the image. Using the trained model was tested in a number of different traffic stations collected data. Compare the method of license plate the positioning will test results with the use of gray image, to verify the detection method of deep learning goals based on the license plate location robustness. Secondly, using the same method to detect deep learning targets based on the license plate characters were detected, and the mainland and Hong Kong and Macao on license plate license plate detection results respectively corresponding to the sorting, finally get the license plate recognition results. In the training process of the model, this paper uses the image of 21269 images were manually labeling software. Due to the license plate The asymmetry of the character has no image enhancement processing. Using different network structure of the data set for training, testing, and comparing the results using the image data collected, selecting the optimal network model. Finally, the license plate location and license plate recognition process combining unified testing of the whole system test the recognition speed. The results show that the recognition speed of 0.12 seconds for each picture, it has more advantages than the traditional license plate recognition system. This system also can be combined with vehicle detection, vehicle recognition. Perfect detecting image to vehicle detection and vehicle recognition, then the license plate detection and recognition. These results can build a big data platform, real-time upload to the vehicle license plate detection models, and location information, for the construction of smart city, to achieve interconnection of all things is of great significance.

【学位授予单位】：青岛科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】