精确的血迹的年龄估计在刑侦法医鉴定中有着重要的意义。 构建了以8个LED为照明光源、 以黑白CCD相机为成像单元的可见-近红外多光谱成像系统, 利用以k最近邻算法、 支持向量机算法和随机森林算法为基模型的融合模型分析预测血迹年龄, 研究了利用可见-近红外反射多光谱精确估计人体血液年龄的可行性, 并与前人利用高光谱进行血迹年龄预测的研究结果进行了对照, 还检验了血液特异性对模型的影响。 实验记录了11个人体血液样本在1～20 d的在400～940 nm之间的8个波长通道的图像, 使用标准正态变换校正(SNV)对光谱进行预处理, 以消除基线平移和散射作用带来的样本光谱差异。 随机选择经过预处理后的7个样本用作训练集以建立模型, 剩余的4个样本用作测试集以验证模型, 建立了基于k最近邻算法、 支持向量机算法和随机森林算法的融合模型, 并与k最近邻算法模型, 支持向量机算法模型, 随机森林算法模型进行比较, 结果显示融合模型的实验结果更好。 基于该融合模型得到的实验结果为： 0~2 d之间预测样本的正确分类率(CCR)为80％, 平均误差为0.053 d, 2~20 d之间预测样本的CCR为69%, 平均误差为0.442 d, 与利用高光谱获得的结果相当甚至更好。 为测试该方法的适用性, 检验了血液特异性对本模型的影响, 实验样本取自8个不同捐献者的20个血迹, 将其中4个捐献者的10个样本加入原有模型, 另外4个捐献者的10个样本作为测试集以检验血液特异性对模型影响。 实验结果为： 在0~2 d之间的CCR为75.6%, 平均误差为0.063 d, 2~20 d之间预测样本的CCR为65.6%, 平均误差为0.467 d。 CCR无显著降低, 表明该模型对不同来源的血迹样本, 仍然有着较好的适应性。 与采用高光谱的研究结果相比, 多光谱结合模型融合算法同样可以获得较好的血迹年龄估计结果, 并具有结构简单, 成本低廉, 稳定性好的优点, 有望成为一种高精度的血迹年龄预测手段, 在法医学领域中有重要应用前景。

The accurate estimation of blood age is of great significance in the forensic identification of criminal investigation. In this paper, a visible-near-infrared multispectral imaging system with 8 LEDs as illumination source and monochrome CCD camera as image input unit is constructed. The ensembling model based on k nearest neighbor method, support vector machine and random forest method is used to analyze and estimate the age of bloodstains. The feasibility of using the visible-near-infrared reflectance multispectral to accurately estimate the age of human blood was investigated, and the results were compared with the previous studies using hyperspectral techniques for blood age estimation. The influence of blood specificity was also tested. The experiment recorded images of 8 channels from 400 to 940 nm on days 1 to days 20 of 11 human blood samples, and the spectra were preprocessed using standard normal variate transformation (SNV) to eliminate spectral differences due to the baseline shift and scattering. Seven preprocessed samples were randomly selected as training set to build the model, and the remaining four samples were used as test sets to test the model, a model ensembling model based on k nearest neighbor method, support vector machine and random forest method was built. Compared with the results by k-NN model, SVM model and RF model the result is better. The correct classification rate (CCR) of the samples between 0 and 2 d is 80%, the average error is 0.053 d, and the CCR between 2 and 20 d is 69%. The average error is 0.442 d, which is comparable or better than that obtained by using hyperspectral techniques. In order to test the practical applicability of the method, this paper tested the effect of blood specificity on the model. The test sample was 20 blood samples taken from 8 different donors, 10 of which from 4 donors were used to refine the original model, and 10 samples from another 4 donors were used as test sets to test the effect of blood specificity. The estimated age of blood from different sources is: CCR is 75.6% between 0 and 2 d, and the average error is 0.063 1 d. After adding blood samples from different donors, there was no significant decrease in CCR, indicating that the model still has good adaptability to blood samples from different sources. The results show that compared with the previous research results, multispectral technology combined with model ensembling algorithm could obtain more accurate age estimation results, and has the advantages of simple set-ups, low-cost and good stability, which might be a high-precision blood age estimation method and have important application value in the field of forensic science.