联机数学公式手写体识别的研究与实现

发布时间：2018-05-13 16:23

本文选题：联机手写体识别 + 手写体数学公式　；参考：《电子科技大学》2017年硕士论文

【摘要】：在教育行业,为了实时追踪学生的学习轨迹和知识薄弱环节,机器自动识别学生答题的手写笔迹成为必要的技术需求。因此,本文的研究重点为联机数学公式手写体识别,旨在提出一个稳健可行的解决方案来识别学生的手写数学公式笔迹,主要研究内容包括如下几点:1、提出了一种融合CNN和DBN的单字符识别模型应用卷积神经网络(Convolutional Neural Network,CNN)搭建和训练了一个单字符分类器模型,针对神经网络对对抗样本的脆弱性表现,同时创造性的采用了深度信念网络(DeepBeliefNetwork, DBN)的解码重构损失作为识别置信度评价模型,最后融合了 CNN和DBN的置信度评价,增强了对对抗样本的拒识能力。2、提出了一种基于组合排序的手写体识别算法手写数学公式中存在着大量的二义性,二维结构准确判定存在着相当的难度,同时还有许多容易混淆的相似字符,这些情况都增大了机器自动识别的难度。本文提出了一种基于组合排序的手写体识别算法,先把这些不确定的情况都保存下来作为候选,产生候选组合路径,再基于词组频率表,语义模型,识别置信度等进行路径排序就能保证识别正确率,可以大大简化系统的复杂度,同时增大识别系统的鲁棒性。3、提出了一种基于错误识别案例的快速学习方法识别出错案例的调试工作非常繁重,本文提出了一个基于错误案例的学习方法,可以快速的从我们提供的错误案例标记数据中学习到新的组合映射知识,并将这些知识作为系统的组合补充分支,避免了再次出现同样的识别错误。基于上述的方法,实现了一个识别系统,实验结果表明本文提出的联机数学公式手写体识别方法具有较高的识别率及较好的鲁棒性。
[Abstract]:In the education industry, in order to track the students' learning track and knowledge weakness in real time, it is necessary to recognize the handwritten handwriting of students' answer questions automatically by machine. Therefore, this paper focuses on on-line mathematical formula handwritten recognition, in order to propose a robust and feasible solution to recognize the handwriting of students' handwritten mathematical formula. The main research contents are as follows: 1. A single character recognition model combining CNN and DBN is proposed. A single character classifier model is built and trained by convolution neural network Convolutional Neural Network. At the same time, it creatively adopts the decoding reconstruction loss of deep belief network (DBN) as the evaluation model of recognition confidence. Finally, it combines the confidence evaluation of CNN and DBN. In this paper, we enhance the ability of rejecting the antagonistic samples. 2. A handwritten recognition algorithm based on combinatorial sorting is proposed. There is a lot of ambiguity in the handwritten mathematical formula, and it is very difficult to determine the 2D structure accurately. At the same time, there are many confusing similar characters, which make automatic recognition more difficult. In this paper, a handwritten recognition algorithm based on combinatorial sorting is proposed. Firstly, these uncertainties are saved as candidates to produce candidate combination paths, and then based on phrase frequency table, semantic model. The correct rate of recognition can be guaranteed by the path sorting of recognition confidence, and the complexity of the system can be greatly simplified. At the same time, the robustness of the recognition system is increased. A fast learning method based on the error recognition case is proposed. The debugging work of identifying the error case is very heavy. In this paper, a learning method based on the error case is proposed. We can quickly learn new combinatorial mapping knowledge from the error case tag data provided by us and use this knowledge as a complementary branch of the system to avoid the same recognition errors. Based on the above method, a recognition system is implemented. The experimental results show that the on-line mathematical formula handwritten recognition method proposed in this paper has a higher recognition rate and better robustness.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】