Research Articles
BAI Siyue, CAI Kanchen, ZHOU Lan, CHEN Ying, WAN Daliang, ZHOU Shengbin
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.
