【作者】 陈泉；

【导师】 骆斌； 冯桂焕；

【作者基本信息】 南京大学 ， 软件工程， 2016， 硕士

【摘要】 随着计算机技术的不断发展,人机交互的研究也开始由传统人机交互向智能人机交互转化,以人为中心的智能人机交互研究成为了当前的一个研究热点,手绘草图作为智能人机交互的一个分支,符合人们长期书写习惯,易于表达个人意图,易于表达个人设计思路。研究使得计算机理解人们手绘草图意图的系统和技术也就变得尤为关键。在线手绘草图识别是由多门学科交叉的模式识别技术,囊括了计算机图形学、模式识别、CAD技术和人机交互等多门学科。本文研究在线手绘流程图的符号识别技术,将在线手绘流程图的符号识别划分为三个阶段：笔划成组、符号识别和流程图的语法解析。笔划成组即将采集到的时序笔划序列组合成可能构成流程图基础符号的笔划组合。本文主要根据笔划之间的时间和空间关系,运用动态规划算法来实现笔划成组工作,这阶段工作得到的结果,本文称之为假设。符号识别就是对手绘流程图的基础符号进行识别,最后将训练好的分类器用于在线手绘流程图的整体识别。本文分别使用了Random Forest、神经网络和卷积神经网络三种分类器算法对流程图基础符号进行识别,对分类器的准确率、灵敏度、特异度和ROC曲线进行评估,最后选择卷积神经网络作为本文对流程图整体识别分类器算法,该阶段分类器的输出结果叫做候选符号集。卷积神经网络较于Random Forest和神经网络能够更加充分的利用符号的笔划之间的空间关系。流程图语法解析即从候选符号集中通过流程图的上下文结构语法来获得最终的识别结果。本文在在流程图语法解析阶段,根据流程图的结构约束,递归的定义了流程图结构语法规则,最后根据语法解析,获取流程图的最终识别结果。流程图语法解析主要是因为得到的候选符号集大约是流程图真实符号集数目的20倍左右,因而需要进一步利用流程图上下文信息获取最终的识别结果。本文流程图整体识别方法结果验证了基于语法的流程图识别方法,在对流程图库FC Database中的流程图的整体识别的识别结果为80.2%,同时本文对流程图结构定义的语法可以运用到其他的具有结构特性的手绘草图的识别当中,如UML,数学公式和化学符号等。更多还原

【Abstract】 With the rapid development of computer technology, human-computer interaction research began by the traditional human-computer interaction to intelligent human-computer interaction transformation. At the same time, intelligent human-computer interaction has become a research hotspot recently. Sketches as a branch of intelligent human-computer interaction, conform to the people for a long time writing habit, easy to express personal intention, and easy to express personal design idea. In order to let computer usderstand the intention of the sketches, sketches recognition systems and technology become extremely important. Online hand-drawn sketch recognition is an interdisciplinary technology, including computer graphics, pattern recognition, CAD technology and human-computer interaction and other subjects. In this paper, we study online hand-drawn flowchart symbols recognition technology. The online hand-drawn flowchart symbols recognition is divided into three stages:stroke group, symbol classification and syntax parsing of flowchart.In the first stage, stroke group is to gernate stroke combinations which are time and spatial adjacent. We make full use of the temporal and spatial relations between strokes and the dynamic programming algorithm to achieve the stroke group working. We call the stroke combination as hypothesis in this paper.In the second stage of symbol classification, we first train the flowchart symbol as our classifier, and then the classifier is used to classify the stroke combinations for the overall flowchart recognition. In this paper, we use three classifier algorithms, which are the random forest, neural network and the convolutional neural network, to train flowchart symbols recognition, assessed on the classifier’s accuracy, sensitivity, specificity and ROC curve. Finally, we choose the convolutional neural network as our flowchart overall recognition classifier algorithm. We call the output results of this stage as candidate symbol sets.In the third stage, the aim of syntax parsing of flowchart is to obtain the final flowchart recognition result from the candidate symbol sets based on the flowchart of the context structure grammar. Firstly, we recursively definite the flowchart structure grammar according to the structural constraints of the flowchart. Then, we introduce a description language to reliaze the flowchart structure grammar. Finally, we obtain the final recognition result of flowchart recognition from the syntax parsing. We have to do the syntax parsing mainly because of the candidate symbol sets we get from the second stages more than 20 tiems of the flowchart real symbols. Therefore, it is necessary to use of flowchart of context information to obtain the final recognition result.Through the proposed method of this paper, we obtain 80.2% accuracy on the flowchart symbol recognition from the FC Database, a handwritten-flowchart database. At the same time, the recognition method based on the structure grammar can be extended to the other hand-drawn sketch recognition, such as UML diagram, mathematical formula, chemical symbol and so on.更多还原