EVOLUTIONARY-BASED APPROACH FOR SOLVING DIGITAL SIGNATURE RECOGNITION TASK

The work reported in the paper aims to develop an image registration methodology for digital signature recognition task. In our research we propose a solution to the problem of a particular component of banking security systems involving the client's signature. Any document reflecting legal transactions executed by a certain bank that includes client’s signature should be pre-processed by the security system in order to authorize it. One of the first steps in the authorization process involves the recognition of the client’s digital signature. Basically, the digital signature to be recognized is often different from the stored one from the geometrical point of view, in the sense that it could be a distorted variant of it. In our developments the working assumption is that the degradation is modeled by rigid transform, where only translation, rotation, and scaling are considered. The registration problem is solved based on the following general procedure. First, the binary variants of both the acquired image and the stored one are computed to make the entire recognition process tractable. Next an evolutionary-based technique is developed to align the input image to the target one. The proposed fitness function is defined in terms of mutual information computed between the transformed image and the target image. We use various mixtures of standard recombination schemes, involving local/global convex and discrete crossover. The mutation procedure comprises uncorrelated multiple sigma-type parameters. The registration quality is evaluated in the final phase using quantitative and qualitative measures. The experimental results together with some concluding remarks regarding the quality of the proposed methodology are reported in the final part of the paper.

本论文所阐述的研究工作，旨在面向数字签名（digital signature）识别任务，构建一套图像配准（image registration）方法。在本研究中，我们针对银行安全系统中涉及客户签名的特定模块难题，提出了一种解决方案。所有包含客户签名、反映某银行办理合法交易的文档，均需经安全系统预处理以完成授权核验。授权流程中的首要步骤之一，即对客户的数字签名进行识别。本质而言，待识别的数字签名与存储的签名在几何层面往往存在差异，前者可视为后者的畸变变体。在本研究的推导过程中，我们的工作假设为：此类图像畸变可通过刚体变换（rigid transform）建模，仅考虑平移、旋转与缩放三类变换。本研究基于下述通用流程解决配准问题：首先，对采集图像与存储图像分别生成其二值化版本，以使整个识别流程具备可操作性；随后，我们提出一种基于进化算法的配准技术，以实现输入图像与目标图像的对齐。所提出的适应度函数（fitness function），基于变换后图像与目标图像间的互信息（mutual information）进行定义。我们采用多种标准重组策略的组合方案，涵盖局部/全局凸交叉与离散交叉。变异过程包含不相关的多σ型参数。在最终阶段，我们通过定量与定性两类评估指标，对配准质量进行评估。本论文的最后部分将展示实验结果，并针对所提方法的性能给出若干总结性评述。