基于图像处理的自动阅卷系统相关技术研究

发布时间：2018-02-03 01:34

本文关键词： 自动阅卷图像处理条形码识别手写字母识别 LVQ神经网络　出处：《太原理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：自动阅卷系统由于其高效的批阅处理、更为客观公正的评分机制以及更加方便的管理功能等优点正逐步替代着传统的人工阅卷方式。现在流行的自动阅卷系统多采用光标阅读机实现答题卡的自动识别,这种方式对答题卡纸质要求较高,且需要购置昂贵的专用识别设备,比较适用于大型考试中。对于近年来出现的基于图像处理的自动阅卷系统也是针对填涂模式的客观题进行识别,这种模式下的识别率对考生的填涂质量依赖太强,容易造成系统误判,而且也不符合考生的答题习惯,还会占用考生较多的填涂时间。针对填涂模式存在的问题,本文对基于手写字母识别模式的自动阅卷系统进行研究。同时只针对试题与答案分离的答题纸进行处理,以减少扫描工作量,提高图像处理的速度,节省系统运行时间与存储空间的开销。本文的主要研究内容如下:(1)结合答题纸图像的特征简化了倾斜校正的过程。对于Hough变换检测直线的过程中计算量较大的问题,先对答题纸图像的特征区域进行边缘检测,再对边缘图像中的横线点进行筛选,最后进行Hough变换得到图像的倾斜角度。(2)提出基于垂直投影的条形码识别方法。将条形码图像识别技术引入到考生的信息识别过程中,简化系统识别的过程,提高识别准确率。基于垂直投影的条形码识别方法,可以实现对受到严重污染和残缺不全的条形码图像的快速准确地识别。(3)提出了一种手写字母特征提取的新方法。针对传统手写字母特征提取方案获得的特征点数较多,造成识别系统结构较为复杂的问题,结合手写字母的特点,提出了八点特征提取方法。经过实验测试证明,对基于八点特征提取法提取的特征点进行识别可以准确地辨识出其代表的字母,同时识别准确率也比较高。(4)基于八点特征提取方法,通过改进遗传算法优化的LVQ神经网络,实现了手写字母的自动识别。对于神经网络因为初始权值设置不合理可能会出现“死”神经元的问题,加入了遗传算法对其进行优化。并对遗传算法进行了改进,加快收敛速度,避免陷入局部最优解。通过实验测试证明,经过改进遗传算法优化的LVQ网络的收敛性和分类性能都有明显的改善和提升。同时基于八点特征提取法的LVQ神经网络的网络结构也比较简单,对手写字母的识别正确率也比较高,满足了自动阅卷系统的性能要求。本文研究的技术和方法对解决基于图像处理的自动阅卷系统的关键问题有很大的借鉴意义,适合应用在中小型考试的阅卷工作中。
[Abstract]:Automatic marking system due to its efficient marking processing. The advantages of more objective and fair scoring mechanism and more convenient management function are gradually replacing the traditional manual marking method. Now the popular automatic marking system uses the cursor reader to realize the automatic recognition of the answer card. . In this way, the paper requirement of the answer card is high, and expensive special identification equipment is needed. The automatic marking system based on image processing in recent years is also used to identify the objective problems of the filling pattern. The recognition rate in this mode is too dependent on the quality of the examinee's filling, and it is easy to cause system misjudgment, and it does not accord with the examinee's habit of answering questions. Also will occupy the examinee more filling time. In this paper, the automatic marking system based on the pattern of handwritten letter recognition is studied. At the same time, only the answer paper which is separated from the answer is processed, in order to reduce the scanning workload and improve the speed of image processing. The main contents of this paper are as follows: 1). The process of skew correction is simplified by combining the feature of the answer paper image. The problem of large computation in the process of detecting straight line by Hough transform is discussed. First, the feature region of the answer paper image is detected, and then the transverse points in the edge image are screened. Finally, Hough transform is carried out to get the tilt angle of the image.) A bar code recognition method based on vertical projection is proposed, and the bar code image recognition technology is introduced into the information recognition process of the examinee. The process of system recognition is simplified and the recognition accuracy is improved. The barcode recognition method based on vertical projection is presented. Can realize fast and accurate recognition of seriously polluted and incomplete barcode images. A new method for feature extraction of handwritten letters is proposed. Because of the complex structure of the recognition system, combined with the characteristics of the handwritten letters, an eight-point feature extraction method is proposed, which is proved by the experiment. Recognition of feature points based on eight-point feature extraction method can accurately identify its representative letters, and the recognition accuracy is also relatively high. 4) based on eight-point feature extraction method. By improving the LVQ neural network optimized by genetic algorithm, the automatic recognition of handwritten letters is realized. For the neural network, the problem of "dead" neurons may occur because the initial weights are not set properly. The genetic algorithm is added to optimize it, and the genetic algorithm is improved to speed up the convergence speed and avoid falling into the local optimal solution. The convergence and classification performance of the improved genetic algorithm (GA) optimized LVQ neural network are improved and improved obviously. At the same time, the network structure of the LVQ neural network based on the eight-point feature extraction method is also relatively simple. The recognition accuracy of handwritten letters is also high. Meet the performance requirements of automatic marking system. The techniques and methods studied in this paper have great reference significance to solve the key problems of automatic marking system based on image processing. Suitable for small and medium-sized examination paper marking work.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】