基于笔迹的性别识别方法研究

发布时间：2019-01-18 21:06

【摘要】：笔迹既包含着书写者先天的生理特征,又受后天学习的影响,能在一定程度上反映书写者的书写习惯和生物特征。从笔迹中提取的信息可以用来判断书写者的性别、年龄和使用右手或者左手的习惯。其中,性别在书写者笔迹风格形成过程中的作用是不容忽视的。在取证分析和人口调查统计中,以性别将人群进行划分是非常有用的。确定书写者的性别能够缩小调查研究的范围,并提高笔迹识别和笔迹验证的效果。同时,结合性别与其他生物特征对案件分析有一定的启发作用。本文着眼于基于笔迹的性别识别的研究。从轮廓特征、纹理特征和深度神经网络自动提取特征三个方面入手,实现了根据笔迹判断书写者性别的目标。本文设计了链码和边界方向提取笔迹图像的轮廓信息,利用SVM算法进行分类,在IAM On-line数据库上得到了71.2%的准确率。本文研究了局部二值模式(LBP)的多种形式,以多尺度LBP提取笔迹图像的纹理信息,通过实验构建了多尺度LBP特征并确定了合适的K值,同时使用KD树分类,在IAM On-line数据库上得到了73.25%的准确率。在深度神经网络方面,本文分析了成熟网络结构的设计思路,在扩充数据的基础上,利用深度学习工具caffe搭建了包含七个卷积层和相应功能层的卷积神经网络,并使用多种技巧提高网络性能,通过合理地设置参数和微调,在IAM On-line数据库上得到了76.17%的准确率,这是该数据库上笔迹性别识别的最高准确率。
[Abstract]:Handwriting contains not only the inherent physiological characteristics of the writer, but also the influence of acquired learning, which can reflect the writing habits and biological characteristics of the writer to a certain extent. Information extracted from handwriting can be used to determine the sex, age, and habit of using the right or left hand of the writer. Gender plays an important role in the process of writing style formation. In forensic analysis and demographic statistics, it is very useful to divide the population by sex. Ascertaining the sex of the writer reduces the scope of research and improves the effectiveness of handwriting recognition and handwriting verification. At the same time, the combination of gender and other biological characteristics has a certain enlightening effect on case analysis. This paper focuses on the study of gender recognition based on handwriting. From three aspects of contour feature texture feature and depth neural network automatic extraction of features the goal of judging the sex of the writer according to handwriting is achieved. In this paper, we design chain code and boundary direction to extract the contour information of handwriting image and classify it with SVM algorithm. The accuracy rate is 71.2% in IAM On-line database. In this paper, the various forms of local binary mode (LBP) are studied. The texture information of handwriting image is extracted by multi-scale LBP. The multi-scale LBP feature is constructed through experiments and the appropriate K value is determined. At the same time, the KD tree is used to classify the texture information. The accuracy rate is 73.25% on IAM On-line database. On the aspect of depth neural network, this paper analyzes the design idea of mature network structure. Based on the extended data, a convolutional neural network including seven convolution layers and corresponding functional layers is constructed by using the depth learning tool caffe. By setting parameters and fine-tuning reasonably, the accuracy rate of 76.17% is obtained in IAM On-line database, which is the highest accuracy rate of handwriting gender recognition in this database.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】