自然场景中文本检测识别系统的设计与实现

发布时间：2019-05-17 00:51

【摘要】：随着智能手机等移动互联网设备以及4G网络等无线高速上网方式的普及,使得人们获得信息的途径越发的丰富、广泛和便捷。图像在传播信息媒介中所占的比例越来越大,已经远远超过传统的文本形式。图像中的文本包含着很多有价值的信息。在图像中可被获取的信息中,人眼的注意力尤其会被文本所吸引[17]。传统的文本检测识别技术专注于扫描文档、票据以及格式统一的文本检测,很少涉及自然场景图片中的文本检测。准确检测和识别自然场景中的文本对于图像的语义理解具有重要意义。本论文的工作内容是基于该作者的实习单位的研究项目进行的。研究重点为移动通信中违规图像的过滤。作者独立完成了以下工作内容:自然场景中文本检测识别系统的设计与实现。作者利用计算机视觉领域的相关知识,采用模式识别和深度学习中的一些算法,从自然场景图片中检测出标识文本,并且通过实验验证了其有效性。系统包括图像的预处理、文本位置检测、文本识别等功能模块。其中,图片的预处理模块对输入数据进行灰度化、增强文本对比度。文本检测模块利用最大稳定极值区域(Maximally Stable Extremal Regions,MSER)算法提取候选区域,并通过深度神经网络提取深度特征。文本识别模块利用识别神经网络进行文本识别,并通过语言模型修正识别结果。自然场景中文本检测识别系统被应用于检测和识别网络图像中的文本信息,提供文本信息线索给违规图像过滤系统。测试结果表明,融合了自然场景文本检测识别系统的违规图像过滤取得了更好的性能,从而达到了设计目标。
[Abstract]:With the popularity of mobile Internet devices such as smartphones and 4G networks, people's access to information is becoming more and more abundant, extensive and convenient. The proportion of images in the communication information media is more and more large, which has far exceeded the traditional text form. The text in the image contains a lot of valuable information. In the information that can be obtained in the image, the attention of the human eye is especially attracted to the text [17]. The traditional text detection and recognition technology focuses on scanning documents, bills and uniform format text detection, and rarely involves text detection in natural scene pictures. Accurate detection and recognition of text in natural scenes is of great significance for semantic understanding of images. The content of this paper is based on the research project of the author's internship unit. The research focuses on the filtering of illegal images in mobile communication. The author completes the following work independently: the design and implementation of the natural scene Chinese text detection and recognition system. Based on the relevant knowledge in the field of computer vision, the author uses some algorithms in pattern recognition and deep learning to detect the identification text from the natural scene picture, and verifies its effectiveness through experiments. The system includes image preprocessing, text position detection, text recognition and other functional modules. Among them, the preprocessing module of the picture grays the input data to enhance the text contrast. The text detection module uses the maximum stable extremum region (Maximally Stable Extremal Regions,MSER algorithm to extract the candidate region, and the depth feature is extracted by the depth neural network. Text recognition module uses recognition neural network for text recognition, and modifies the recognition results through language model. The natural scene Chinese text detection and recognition system is used to detect and recognize the text information in the network image, and to provide the text information clue to the illegal image filtering system. The test results show that the illegal image filtering combined with the natural scene text detection and recognition system achieves better performance, thus achieving the design goal.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】