基于拍照的银行卡卡号检测

发布时间：2018-08-11 17:36

【摘要】：随着成像设备的广泛使用,只需要嵌入一个模块,移动设备就可以通过拍照获取的银行卡图像自动输入银行卡账号而不用人工输入。因此,基于拍照的银行卡卡号检测和识别技术具有重要的研究价值。和自然场景中的文本检测一样,基于拍照的银行卡卡号检测面临着同样的难题,卡号文本同样存在字体、大小、排列方向上的多样性,也受光照条件、透视变换和对比度的影响,另外,卡号的复杂背景也加重了卡号检测和识别的难度。本文在自然场景中文本检测的基础上对基于拍照的银行卡卡号检测做了系统的研究,提出了基于特征提取、机器学习的卡号检测方法。本文的主要工作如下:首先,本文算法是用来检测水平卡号行的,需要水平校正银行卡图像。本文提出了两种预处理算法来改进已经提出了的Radon变换倾斜校正算法,第一种是对输入图像做边缘检测,第二种是对输入图像做直线段检测,然后对边缘或直线段图像做Radon变换,检测银行卡的倾斜角度。实验结果显示两种预处理改进能够提高银行卡图像的倾斜校正效果。其次,根据卡号和它的相邻背景间存在瞬态颜色,有一定的对比度,本文采用形态学算法来提取卡号的这种对比度特征;接下来,本文巧妙地将水平投影和k-means结合,得到了比较好的候选卡号行定位效果。最后,在卡号验证过程中,本文对传统的LBP算法进行了改进,提出了改进的LRBP(Region Local Binary Pattern)特征,该特征对卡号的纹理特征的描述能力更好,提高了银行卡卡号行的检测效果。接下来,算法分别提取了滑动窗的HOG和改进的LRBP特征通过训练好的SVM分类器来验证卡号域,在这一过程中,算法使用了分类器集成来提高分类器的检测精度。最后通过实验数据集检测,本文算法能很好地检测出银行卡号。
[Abstract]:With the wide use of imaging equipment, only one module needs to be embedded, and the mobile device can automatically input the bank card account without manual input by taking pictures of the bank card image. Therefore, the bank card number detection and recognition technology based on photograph has important research value. Like text detection in natural scenes, the bank card number detection based on taking pictures faces the same problem. The card number text also has the diversity of font, size, arrangement direction, and is also subject to illumination conditions. The influence of perspective transformation and contrast, in addition, the complex background of card number also increases the difficulty of card number detection and recognition. Based on the Chinese text detection of natural scene, this paper makes a systematic research on the bank card number detection based on taking pictures, and puts forward a method of card number detection based on feature extraction and machine learning. The main work of this paper is as follows: firstly, the algorithm is used to detect the horizontal card number line, and the horizontal correction of bank card image is needed. In this paper, two preprocessing algorithms are proposed to improve the proposed Radon transform skew correction algorithm. The first is to detect the edge of the input image, and the second is to detect the line segment of the input image. Then the edge or straight line image is transformed by Radon to detect the tilt angle of bank card. The experimental results show that the two preprocessing improvements can improve the skew correction effect of bank card image. Secondly, according to the transient color between the card number and its adjacent background, there is a certain contrast. In this paper, morphological algorithm is used to extract the contrast feature of the card number. Then, the horizontal projection and k-means are skillfully combined in this paper. A good candidate card number line location effect is obtained. Finally, in the process of card number verification, the traditional LBP algorithm is improved, and an improved LRBP (Region Local Binary Pattern) feature is proposed, which can describe the texture feature of the card number better and improve the detection effect of the bank card number line. Then, the HOG of sliding window and the improved LRBP feature are extracted respectively to verify the card number domain through the trained SVM classifier. In this process, the classifier integration is used to improve the detection accuracy of the classifier. Finally, through the experimental data set detection, the algorithm can detect the bank card number well.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】