基于自然场景的文本识别技术研究

发布时间：2019-06-25 21:25

【摘要】：随着智能设备的普及,智能设备拍摄的场景图像中包含的高级语义信息越来越受研究人员的关注。传统的光学字符识别技术往往无法很好地应用于场景图像文本的提取和识别,因为场景图像具有扫描图像所不具备的特性。造成识别困难的原因一方面是由于场景图像本身背景复杂,且拍摄场景图像时往往条件不可控,使得场景图像存在分辨率低、光照不均、模糊等问题;另一方面,由于场景图像中的字符往往在字体、大小和颜色等方面各有差异。因此,针对场景图像的文本识别问题需要新的解决方法。本文的工作是基于这样的背景开展的。本文主要工作包括:(1)深入研究和分析了当前自然场景文本识别领域的相关技术和现状;(2)基于深度学习中的卷积神经网络,实现了一个端到端的场景文本识别系统——MatE2E系统。该系统利用卷积神经网络学习字符特征,训练了两个分类器,分别用于字符的判定和识别。MatE2E系统的主要模块有两个。第一个主要模块是文本检测模块,该模块利用字符判断分类器和滑动窗口扫描场景图像来检测可能的文本区域,之后根据图像中不同位置的文本置信度过滤非文本区域。第二个主要模块是文本识别模块,该模块结合字符识别分类器和滑动窗口扫描文本区域图像并识别其中的文字,之后利用词典更正文本识别结果;(3)使用ICDAR2011数据集、ICDAR2015数据集和街景数据集验证了系统的准确率。实验结果表明,本文提出的系统具有较好的识别效果。本文在实验数据集上的验证结果表明MatE2E系统在实际应用中具有一定的参考价值。不过,MatE2E系统还有局限性,例如:只能识别英文字母和数字;另外还需在识别速度、倾斜文本检测等方面进行改进等。
[Abstract]:With the popularity of intelligent devices, more and more researchers pay attention to the advanced semantic information contained in the scene images taken by intelligent devices. The traditional optical character recognition technology can not be well applied to the extraction and recognition of scene image text, because the scene image has the characteristics that scanning image does not have. On the one hand, the difficulty of recognition is due to the complexity of the background of the scene image itself, and the conditions are often uncontrollable when shooting the scene image, which makes the scene image have some problems, such as low resolution, uneven light, blurring and so on; on the other hand, the characters in the scene image are often different in font, size and color. Therefore, new solutions are needed to solve the problem of text recognition of scene images. The work of this paper is based on this background. The main work of this paper is as follows: (1) the related technologies and present situation in the field of natural scene text recognition are deeply studied and analyzed; (2) based on the convolution neural network in deep learning, an end-to-end scene text recognition system, MatE2E system, is implemented. The system uses convolution neural network to learn character features and trains two classifiers for character determination and recognition respectively. There are two main modules of MatE2E system. The first main module is the text detection module, which uses character judgment classifier and sliding window to scan the scene image to detect the possible text area, and then filters the non-text region according to the text confidence of different positions in the image. The second main module is the text recognition module, which combines character recognition classifier and sliding window to scan the text area image and recognize the text, and then uses the dictionary to correct the text recognition results. (3) ICDAR2011 dataset, ICDAR2015 dataset and street view data set are used to verify the accuracy of the system. The experimental results show that the system proposed in this paper has a good recognition effect. The verification results on the experimental data set show that the MatE2E system has certain reference value in practical application. However, MatE2E system has limitations, such as: can only recognize English letters and numbers; in addition, it also needs to be improved in recognition speed, tilt text detection and so on.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.4

【相似文献】