人类学学报 ›› 2019, Vol. 38 ›› Issue (02): 265-275.doi: 10.16359/j.cnki.cn11-1963/q.2018.0030cstr: 32091.14.j.cnki.cn11-1963/q.2018.0030

• • 上一篇    下一篇

结合改进卷积神经网络和最小二乘法的颅骨性别鉴定

杨稳(), 刘晓宁*(), 刘雄乐, 朱丽品   

  1. 西北大学信息科学与技术学院,西安 710127
  • 收稿日期:2017-08-01 修回日期:2018-02-22 出版日期:2019-05-15 发布日期:2020-09-10
  • 通讯作者: 刘晓宁
  • 作者简介:杨稳(1993-),男,陕西安康人,主要从事机器学习与模式识别研究,Email: 920759301@qq.com
  • 基金资助:
    国家自然科学基金项目(61673319);国家自然科学基金项目(61363065);陕西省自然科学基金项目(2018JM6061);陕西省自然科学基金项目(2014JM8358);研究生自主创新项目(YZZ17181)

Skull sex identification using improved convolution neural network and least squares method

YANG Wen(), LIU Xiaoning*(), LIU Xiongle, ZHU Lipin   

  1. College of Information Science and Technology, Northwest University, Xian, 710127
  • Received:2017-08-01 Revised:2018-02-22 Online:2019-05-15 Published:2020-09-10
  • Contact: LIU Xiaoning

摘要:

颅骨性别鉴定在法医学和颅骨面貌复原等领域具有重要研究意义和应用价值,针对传统颅骨性别鉴定需要专家参与且主观性强、计算机辅助方法需要人工标定特征点等问题,本文提出了结合改进卷积神经网络和最小二乘法的颅骨性别鉴定方法。首先,获取三维颅骨模型多角度颅骨图像,利用改进的卷积神经网络计算每个样本的每张图像属于男性和女性的概率;其次,基于概率均值采用最小二乘法计算每张图像对性别鉴定的权重;最后,利用上述步骤得到的最优参数构造决策函数,通过决策值完成颅骨性别鉴定。本文方法抛弃了繁琐的手动测量,对完整颅骨的性别鉴定正确率高达94.4%,对不完整颅骨的性别鉴定正确率高达87.5%,能够获得较好的颅骨性别鉴定性能。

关键词: 颅骨性别鉴定, 多角度图像, 改进卷积神经网络, 最小二乘法

Abstract:

Skull sex identification has significant research and applied value in forensic anthropology and skull reconstruction. The traditional skull sex determination methods need expert participation and to some extent, is not objective, because computer-aided methods require to marking landmarks the feature points manually. We present a novel sex determination method based on improved Convolution Neural Network and Least Square. Firstly, obtain multi-angle skull images of three-dimensional skull model, and calculate the probability of each image belongs to male or female. Secondly, the weight of each image is calculated using the Least Squares method based on the probability mean. Lastly, the sex determination function is constructed by using the optimal parameters obtained through the above steps. This method does not need to mark the feature points or do the measurement. Experiments show that the proposed method can get quite a reliable performance with an accuracy of 94.4% for the the complete skull and 87.5% for the incomplete skull.

Key words: Sex determination, Imagery, improved convolution neural network, Least squares method

中图分类号: