【作者】 王世雄；

【导师】 刘晓宁；

【作者基本信息】 西北大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 随着数据采集技术和计算机技术的发展,以数字化三维颅骨为对象进行颅骨性别和种族的自动化鉴定成为法医学、人类学、考古学等领域内十分重要的研究课题。在颅骨性别和种族的鉴定研究中,按时间划分,主要分为传统鉴定方法和计算机新技术鉴定方法,其中传统鉴定方法主要包括形态学方法和测量方法。由于传统方法需专家参与、人工标定特征点和繁琐的测量操作,存在鉴定结果主观性强、误差大、耗时等问题。而基于计算机新技术能自动化实现颅骨的性别和种族鉴定,不仅缩短了鉴定周期、避免了对颅骨造成二次伤害,还提高了鉴定准确率。因此,针对上述问题,本文提出了一种基于颅骨侧面轮廓与卷积神经网络的颅骨性别鉴定算法,以及一种基于颅骨深度图与卷积神经网络的颅骨种族鉴定算法,并设计开发了颅骨性别和种族鉴定的原型系统。本文研究工作主要包括:(1)设计一种用于神经网络中特征提取的Wide模块。将上层的输出结果经过1×1的卷积核,再分别连接到混合卷积层中的3×3大小的正常卷积核和空洞率为k的空洞卷积核,最终将得到的特征与前一层1×1的特征层进行通道的合并。这种多通道合并的方式采用了稀疏矩阵转化为多个稠密矩阵计算的方式来使网络快速收敛,并且不易过拟合。(2)针对目前颅骨性别鉴定方法的主观性强、准确率不高等问题,提出一种颅骨侧面轮廓结合卷积神经网络的颅骨性别鉴定方法。首先,通过对颅骨数据的一系列预处理,得到训练需要的侧面轮廓图像;然后,使用Wide模块设计的神经网络模型SINet对完整颅骨和破损颅骨的侧面轮廓图进行性别鉴定。实验结果表明,颅骨的侧面轮廓可以很好的代表颅骨的性别特征,本方法对完整颅骨性别鉴定准确率达到96.5%,对四类破损颅骨鉴定准确率分别达到94.5%、87.5%、90.0%和75.0%。(3)针对直接采用颅骨模型进行种族鉴定需要手工标注特征点、准确率不高的问题,提出三维颅骨的两种深度投影表示方法。针对归一化的颅骨三维模型,提出两种深度投影方式:局部深度投影和全局深度投影。局部深度投影是对颅骨的正、侧、底方向与颅骨模型相切位置建立投影面,然后将点到面的距离作为灰度值进行投影。全局深度投影是对颅骨模型建立圆柱形投影面,并将圆柱的中心轴设置为法兰克福坐标系中的Z轴,然后根据点到面的距离作为灰度值进行投影。(4)针对目前颅骨种族鉴定方法的自动化程度低、准确率不高等问题,提出一种颅骨深度图像结合卷积神经网络的颅骨种族鉴定方法。首先,使用本文提出的两种深度投影方式得到颅骨的深度图像;然后,根据SE block和Wide模块搭建用于种族鉴定的网络ANINet;最后将两种深度图像进行结合输入到ANINet中进行模型训练,得到最终的鉴定模型。实验表明,本文提出的方法准确率达到了99.03%。(5)颅骨性别和种族鉴定系统的设计与实现。利用上述提出的方法,设计并实现颅骨性别和种族鉴定原型系统。实验表明,本文设计的系统可以有效地实现颅骨性别和种族的鉴定。更多还原

【Abstract】 With the development of data acquisition technology and computer technology,the skull automatic identification for sex and ancestry with digital three-dimensional skulls has become a very important research topic in the fields of forensic science,anthropology,archeology and so on.Skull sex and ancestry identification methods are mainly divided into traditional identification methods and computer technology identification methods,of which traditional identification methods mainly include morphological methods and measurement methods.Because traditional methods require experts participation,manual calibration of feature points,and cumbersome measurement operations,there are problems such as strong subjectivity,large errors,and time-consuming.Computer technology is used to automatically realize the sex and ancestry identification of the skull,which not only shortens the recognition cycle,but also avoids secondary damage to the skull and improves the recognition rate.It is proposed in this paper that a skull sex identification algorithm based on the skull side contour images and convolutional neural network,and a skull ancestry identification algorithm based on the skull depth map and convolutional neural network.A prototype system is developed for skull sex and ancestry recognition.The research work in this paper mainly includes:(1)A Wide module for feature extraction in neural networks is proposed.The output result of the upper layer is passed through a 1×1 convolution kernel,and then connected to the normal convolution kernel of 3×3 size and the hole convolution kernel with k hole ratio respectively in the hybrid convolution layer.Combine the final feature with the previous1×1 feature layer on the channel.This multi-channel merge method uses a sparse matrix to convert into multiple dense matrixes for calculation to make the network quickly converge.and it is not easy to overfit.(2)Aiming at the problems of highly subjectivity and low accuracy of current skull sex ide ntification methods,a method for skull sex identification based on the lateral profile of the skull combined with a convolutional neural network is proposed.First,through a series of preprocessing of the skull data,3D models of the skull are reconstructed to obtain the side images of the 3D models.Then extract the edges of the side images to obtain the side contour image required for training.Next,the neural network SINet with Wide module is used to sex-identify the lateral contours of the complete and broken skulls.The experimental results showed that the lateral profile of the skull can represent the sex characteristics of the skull well,the accuracy rate of gender identification of intact skull was 96.5%,and the accuracy rate of four types of broken skull was 94.5%,87.5%,90.0%and 75.0%,respectively.(3)Aiming at the problem that the ancestry identification requires manual labeling of feature points and the result had low accuracy,two depth projection representations for3 D skull are proposed.For the normalized 3D skull models,two depth projection methods are proposed: local depth projection and global depth projection.The principle of the local depth projection is to establish a projection plane for the front,side,and bottom of the skull and the tangent position of the skull model,and the distance from the point to the surface as the gray value of the skull deep images to project.The principle of global depth projection is to establish a cylindrical projection surface for the skull model,setting the center axis of the cylinder to the Z axis in the Frankfurt coordinate system,and the distance from the point to the surface as the gray value of the skull deep images to project.(4)Aiming at the problems of the low degree of automation and the low accuracy rate of the current skull ancestry identification methods,a method for skull ancestry identification using deep skull images combined with convolutional neural networks is proposed.First,the depth images of the skull are obtained by two methods of depth projection proposed in this paper.Then the network ANINet for ancestry recognition is established according to the SE block and Wide modules.Finally,the two depth images were combined and input into ANINet for model training to obtain the final identification model.Experiments showed that the accuracy of the method proposed in this paper is 99.03%.(5)A skull sex and ancestry identification system was designed and implemented.A prototype system for skull sex and ancestry identification was designed according to the methods mentioned before in the paper.Experients showed that the system designed in this paper can effectively estimate the sex and ancestry of the skull.更多还原