目的 尸体角膜随死后时间延长发生的形态学变化是规律性较好的指标,常用来判断死亡时间（postmortem interval, PMI）。本文尝试用机器视觉代替人的肉眼主观判断,收集尸体样本以建立通过人体角膜图像推断PMI的模型。方法 收集实际案例建立包含505例人体死后角膜图像的数据库,PMI范围为0.24h（约死后14min）至492h（约死后20.5d）,大致分为三类（依次为：0~<6h、6~<20h、20h及以上）或二类（0~<15h、15h及以上）;使用由华盛顿大学陈天奇博士提出的Xgboost模型分别进行二分类与三分类分析;使用多种卷积神经网络模型分别进行分类和回归学习,并通过比较最终选择了由微软研究院提出的ResNet模型进行分析。结果 Xgboost在三分类时预测准确率依次为71.8%、40.7%、65.7%,二分类时为90%、48.5%。ResNet分类模型中,精准率、召回率在三分类时分别依次为：81%、75%,30%、50%,61%、71%,二分类时为：70%、92%,76%、38%。ResNet回归模型中,比较整个模型的预测结果,0~6h内的预测值与真实值较为接近,均值误差为0.5616,均方误差为0.5873,6h之后开始出现较大误差。结论 分类和回归模型都在0~6h之内得到了很好的结果,说明在此时间段内,角膜图像噪声较低,可预测性强。

The morphological changes of cornea are an important indicator for postmortem interval (PMI) estimation, thus having frequently been used in forensic practice when available. In this paper, an attempt was carried out to estimate PMI from human corneal images through machine vision instead of human visual subjective judgment. Based on routine forensic examination, a PMI database, enclosing 505 corneal images of their respective PMI labeled from 0.24-492h, was established, consequently being roughly divided by PMI into three categories: 0-6h, 6-20h and more than 20h, or two categories: 0-15h and more than 15h. Xgboost, proposed by Dr. CHEN Tianqi of the University of Washington, was used to perform two- and three-category classifications. The convolutional neural network model was also selected to perform both the classification and regression learning. However, ResNet, developed by Microsoft Research Institute, was the final chosen model for analysis because of its outperformance. For Xgboost, its accuracy showed with three-category classification at 71.8%, 40.7% and 65.7%, and two-category classification at 90% and 48.5% in their respective designated PMI categories. For ResNet, the three-category classification contributed its precision rate 81% and recall rate 75% with the first category 0-6h, plus the corresponding 30% and 50% about the second category 6-20h or 61% and 71% for the category 20h and more, respectively. When ResNet was put under the two-category classification, its precision rate was 70% and recall rate 92% for the first category 0-15h, together with the second category more than 15h demonstrating the respective 76% and 38%. For ResNet to play role into regression learning, its predicted numeral was closer to the true value for the 0-6h PMI, with the mean error value 0.5616 and mean squared error value 0.5873, contrasting to large errors appearing after 6h. Therefore, the selected models proved their performance in both classification and regression learning, showing better for the 0-6h PMI estimation because the corneal images in the interval were of low noise and high predictability.