基于深度卷积神经网络的表单中手写签名位置定位方法

发布时间：2018-10-09 09:09

【摘要】：在人们的日常生活中经常会使用到各种表单或者票据,他们是很重要的法律凭证。为了自动对表单或者票据中的签名验明真伪,比如银行存取款凭证签名真伪识别、保险公司保单签名真伪识别、快递单据识别等等,需要首先确定签名在表单中的位置。现如今大部分表单的签名真伪识别工作都是由人工来完成的,费时费力且受人为主观因素影响较大。因此,开发一套表单手写签名自动检测系统具有十分重要的研究意义。本文主要研究了手写签名自动检测系统中的目标定位功能。目标定位一直是一个热点的问题,目前常用的方法有基于颜色、纹理、形状、空间、模板匹配等方法。本文采用卷积神经网络获取候选区域的方法,实现了表单中手写签名的定位功能。首先,使用卷积神经网络获取候选区域,其实质是使用滑动窗口对图像进行穷举搜索;其次,利用分类层对每个特征图上的候选区域做分类任务,判断该候选框是前景还是背景;最后,利用回归层对每一个候选框的位置做回归。本文主要研究内容有:(1)收集了大量、多样的表单图像,自定义表单图像数据库。收集的图像包含银行表单、保险公司表单、快递公司表单等,对收集好的表单图像进行旋转、平移、缩放、加噪等处理,这样设计出的表单图像数据库会使得训练结果更好。(2)为了提高区域候选框质量,根据设计的表单图像数据集中目标的长宽比,优化滑动窗口中获取候选框的窗口比例。(3)为了减少高层卷积细节特征信息损失,尽可能多的获取到更多图像特征,对RPN网络模型连接结构进行改进。文中提出了 RPN-X的网络结构,并且对该网络模型中的训练参数进行了优化。(4)基于caffe框架,搭建训练神经网络的平台,并训练出RPN-X神经网络模型。(5)在matlab编程环境下,开发了一个手写签名位置定位系统,实现了实时拍照上传并且进行手写签名位置的定位。
[Abstract]:Forms or notes are often used in people's daily lives, and they are important legal documents. In order to automatically verify the authenticity of the signature in the form or bill, such as bank deposit and withdrawal certificate signature authenticity identification, insurance company policy signature authenticity identification, express document identification and so on, it is necessary to first determine the position of the signature in the form. Nowadays, the identification of signature authenticity of most forms is done manually, which is time consuming and influenced by human subjective factors. Therefore, the development of a form handwritten signature automatic detection system has a very important significance. In this paper, the function of target location in automatic detection system of handwritten signature is studied. Target location has always been a hot issue. At present, the commonly used methods are based on color, texture, shape, space, template matching and so on. In this paper, we use convolutional neural network to obtain candidate regions and realize the localization function of handwritten signature in the form. Firstly, using convolutional neural network to obtain candidate regions, the essence of which is to search the images by sliding window. Secondly, the candidate regions on each feature map are classified by classification layer. Finally, the location of each candidate is regressed by regression layer. The main contents of this paper are as follows: (1) collect a large number of various form images and customize the form image database. The collected images include bank forms, insurance company forms, courier company forms, etc. The collected images are rotated, translated, scaled, noised, etc. The designed form image database will make the training result better. (2) in order to improve the quality of the region candidate, according to the aspect ratio of the target in the designed form image dataset, (3) in order to reduce the loss of high-level convolution detail feature information and obtain as many image features as possible, the RPN network model connection structure is improved. In this paper, the network structure of RPN-X is proposed, and the training parameters in the network model are optimized. (4) based on the caffe framework, the platform of training neural network is built, and the RPN-X neural network model is trained. (5) in the matlab programming environment, A position location system for handwritten signature is developed, which can locate the location of handwritten signature by taking pictures and uploading them in real time.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】