【作者】 李杰；

【导师】 方木云；

【作者基本信息】 安徽工业大学 ， 计算机科学与技术， 2016， 硕士

【摘要】 文字是人类文明的重要标志,作为传承人类历史与文明的重要媒介和工具,它对人类的进步起到促进的作用。在信息化程度非常发达的今天,文字几乎无处不在,报纸、书籍、文档、报表、海报、名片、广告牌等等,都包含大量的文字信息。而人们对于这种信息获取的方式和要求促使文字检测和识别技术在办公自动化、人机交互、机器人导航、无人驾驶汽车等诸多领域的广泛发展。如果能让计算机自动的去识别图像中的文字,那么将会为人类提供一种非常便利的与计算机之间进行沟通的交流方式。经实验验证,在大样本测试集下国内现有成熟的OCR识别软件的首位识别准确率为95%~97%之间,在准确率上仍有提升的空间。因为识别算法的核心是文字图像提取的特征和相似度度量方法,所以文字识别的两个很重要的内容就是图像的特征工程和相似度度量。文字识别的主要步骤是先提取文字图像中相关特征,然后依据某种度量准则对提取的特征进行衡量。由于提取的特征和采用的度量方法不一样,所以各种文字识别的准确率和效率也不一样。针对当前文字识别的准确性不足,在基于统计模式识别和结构模式识别方法基础上,提出了一种基于概率特征和结构特征融合的自适应文字识别算法。该算法模拟人类学习的模式,通过对训练样本的学习去构建汉字在测量空间的概率分布矩阵,比对原始图像和标准汉字库中汉字的概率分布矩阵的相似度来达到汉字分类的效果。其中相似度度量准则是从矩阵空间的结构和概率两个角度出发去构建的,充分考虑了结构模式识别和统计模式识别方法的优缺点。实验结果显示该算法在训练样本下的首位识别正确率可以达到99.66%,在1623张非训练样本文字图像下的首位识别正确率可以达到99.13%,在5515张非训练样本文字图像下的首位识别正确率可以达到98.57%。由此可以证明文中提出的相似度度量方法在文字识别的有效性。由于算法还有较为充足的改进空间,并且随着学习样本的增加,当前建立在相似度度量标准上的算法的准确率还有进一步提升的空间。更多还原

【Abstract】 Character is an important symbol of human civilization, as a heritage of human history and civilization of the important tool and media, It is to promote the role of human progress. Especially in the modern urban environment, the characters are almost everywhere. Newspapers, books, documents, reports, posters, cards and billboards, contains a large amount of characters, so automatic character detection and recognition technique in image search, office automation, human-computer interaction, robot navigation, unmanned aerial vehicles and other fields have broad application prospects. If the computer can automatically identify the characters in the image, it will provide a very natural communication between the computer and the communication mode.Experimental results show that the first recognition accuracy of the existing mature OCR recognition software in the large sample test set is 95%~97%,there is still room for improvement in the accuracy rate. Because the core of the recognition algorithm is the feature of text image extraction and similarity measure, the two important contents of text recognition are the feature engineering and similarity measure of image. The main step of character recognition is to extract the text image of the relevant features, and then based on a measure of the extracted features to measure. Since the extracted features and the measurement methods are not the same.so the accuracy of a variety of text recognition and efficiency is not the same.In view of the lack of the accuracy of the current character recognition. Based on statistical pattern recognition and structural pattern recognition, we propose an adaptive character recognition algorithm based on probability characteristics and structural features, through the different number of training samples to build the probability distribution matrix of Chinese characters in the measurement space. By comparing the similarity of the probability distribution matrix of the Chinese characters in the original image and the standard Chinese character library, the results of the classification of Chinese characters are achieved. The similarity measurement criteria is from point of view of the matrix of spatial structure and the probability of building. Experimental results show the algorithm in the training samples of the first recognition accuracy can reach 99.66%, in 1623 non training sample text image the first recognition correct rate reached 99.13%, 5515 non training sample text image in the first recognition correct rate can reach 98.57%. It can be concluded that the similarity measure method proposed in this paper is effective in character recognition.with the increasing of learning samples, this paper establishes the accuracy of the algorithm in the similarity measure standard and the space of progress could be have more adequate space for improvement.更多还原