【作者】 周兰；

【导师】 刘良；

【作者基本信息】 华中科技大学 ， 法医病理学， 2010， 博士

【摘要】 [背景]死亡时间(Postmortem interval, PMI)推断一直刑事科学技术和法庭科学技术的重点和难点。如何实现准确、快速的推断,对迅速侦破案件,处理司法鉴定中的疑难问题等均具有重要意义。死亡时间推断实验成果颇多,但均因检材制备较为复杂,实验方法难以统一,离实际应用存在一定的距离。本实验以易于观察、结构单一的角膜为实验对象,通过高清数码摄像机连续拍摄48小时,对家兔死后角膜随时间的变化进行直接的观察和记录。在角膜区域图像分割和图像特征提取的基础上,使用分类计算的方法建立死亡时间推断模型。本实验采用的方法步骤简单,运算方法可靠,预测结果较为满意。[目的]1.建立0-48小时内,连续高清摄像,间隔15min截取兔眼数码照片以观察角膜变化的方法。2.应用matlab软件,在角膜图像分割方法的基础上和建立进行参数提取和运算的方法,取得角膜混浊图片特征库。3.建立用分类器进行死亡时间推断的模型。4.对不同分类器预测效果的进行比较。[方法和步骤]1.预实验：健康家兔1只,空气栓塞法处死。尸体置于温度控制在(25±1)℃,相对湿度20%-60%的暗室中,门窗关紧遮光,台灯照明,用止血钳使角膜暴露于空气中。从死后即时到48小时内,使用高清摄像机(SONY,HDR-SR12)连续拍摄48小时,取得视频每15分钟截图,取死后15分钟至47小时的图片依次标号为1-188。2.对兔眼图片进行直方图分割,将图片分割为角膜瞳孔区域和其他区域,并提取出角膜瞳孔区域。3.利用matlab软件提取一下9种视觉特征：GF, EL, G, S, C, J, Mean, Var和Ske。其中,GF、MEAN/VAR和SKE为颜色特征量,G、S、C和J为纹理特征量。9个特征参数分别为：GF：描述两区域灰度值的区别；MEAN:反应图像颜色的平均值；VAR：表达颜色在图像上的分布均匀程度；SKE：表达图像颜色分布的不对称性；EL：描述图像纹理的光滑程度；G：描述图像纹理清晰程度；S：反映图像纹理的多少和质地；C：描述矩阵中行或列元素之间相似程度的；J对图像灰度分布均匀性的度量。4.用K最近邻(K-Nearest Neighbor, KNN)分类器对兔子的所有图片进行分类计算,采用4-折的交叉验证运算结果取平均值,进行5轮运算。将15分钟到47小时依次分为3个、4个、5个、6个、8个、10个和14个时间段(分类数为3时,则表示将死后时间分隔成三段,每个时间段15小时左右),观察各个时间段分类的准确率和时间精度对结果的影响。5.取家兔4只。依次标为兔1,兔2,兔3,兔4,图片摄取和分析处理的方法以及KNN运算方法同预实验。6.取9个参数联合使用,分别用KNN分类器、Adaboost (Adaptive Boosting)分类器和SVM (Support Vector Machine)分类器对4只兔子的所有图片进行分类计算,采用4-折的交叉验证运算结果取平均值,进行5轮运算。将15分钟到47小时依次分为3个、4个、5个、6个、8个、10个和14个时间段(类别数),比较三种种分类器的分类准确率。[结果]1.建立的分类模型能较好的完成死亡时间的分类运算以进行预测。2.所用角膜分割方法均能较好的完成的兔眼图像分割。3.9种视觉特征均可用于推断模型,单独使用的分类能力较弱,联合使用结果较好。4.同类别数下,9个参数联合运用KNN分类器,单只兔子和4只联合运算均取得了较为满意的结果。5.联合使用各特征,随着死后时间分段数的增加,各时间间隔内的分类准确率下降。对单只兔子数据使用KNN分类器,在分类数为3时,单只兔子的分类准确率平均值为97.1%,分类数为8时准确率平均值为88.5%,分类数为14时平均值为81.5%；取四只兔子图片数据一起进行分类,其准确率在分类数3时为96.9%,分类数为8时准确率为87.6%,分类数14时准确率为80.9%。6.使用Adaboost分类器,9参数联合使用,单只兔子的分类准确率平均值在分类数为3时,为94.4%；分类数为8时,平均值为85.3%,分类数为14时,平均值为72.9%。四只兔子联合运算,分类数为3时准确率为85.1%；分类数为14时准确率为64.7%统计分析结果显示,较之于KNN分类器,单只兔子运算,分类数较少时两者无明显差异,而分类数较多(10个和14个)有统计学差异,Adaboost分类器低于KNN分类器的结果。而多只兔子联合运算时Adaboost明显低于KNN分类器。7.使用SVM分类器,9参数联合使用,分类数为3时,单只兔子的平均分类准确率为88.3%,四只兔子联合运算时准确率为78.9%；分类数为14时,单只兔子的平均分类准确率为50.9%,四只兔子联合运算时准确率为30.2%。远低于前两种分类器。[结论]1.成功建立了用9个图像特征对死亡时间进行推断的分类模型。2.基于灰度直方图的方法可以对角膜瞳孔区域进行有效地截取。3.本实验所选取的9个视觉特征可以对兔的死亡时间的进行较准确的推断。4.建立的KNN分类器推断模型能较好的完成本实验数据的运算,得出较为稳定的结果。5.在本实验中,KNN分类器优于SVM分类器和Adaboost分类器。背景：中毒在全球均有较高的发生率,危害人类健康,对人类的生产和生活均有比较大的影响。本文旨在描述华中地区中毒死亡案件的分布特点和变化趋势,为案件调查和公共防治提供参考。材料和方法：取1999年-2008年华中科技大学同济医学院法医学系暨湖北同济法医学司法鉴定中心218例中毒案件,并对其进行回顾性研究。结果：年龄主要集中分布于20-49岁,占69.7%,男女比率为1.7:1。其中最为常见的为杀鼠剂中毒,占19.7%,农药及除草剂中毒占17.9%,一氧化碳中毒占16.5%,而药物、酒精中毒分别为13.8%、12.4%。口服为主要的中毒途径,占65.1%,其次依次为吸入、注射及皮肤接触。死亡方式大多数为为意外中毒,占64.7%；自杀占25.2%,而他杀中毒死亡占3.7%,未定性的占4.1%。与本单位1956-1984年和1983-1999年两次的研究资料相比,杀鼠剂、CO、酒精和药物中毒的比例升高,意外中毒死亡的案件比例也升高。结论：农药中毒在中国仍为突出的威胁公共安全和健康的问题,政府实行有效的管理措施,进一步对农药,尤其是杀鼠剂的应用,进行有效的限制及管理将尤为重要。更多还原

【Abstract】 BackgroundThe estimation of postmortem interval has been one of the most important and difficult issues in the field of criminal science and forensic medicine. It is of great importance to estimate PMI accurately and quickly both for criminal cases and civil cases. During the past decades, many new methods have been proposed to estimate PMI, both by theoretical and technologic means involving physics, chemistry, biology, immunohistochemistry and other basic sciences. But as yet, all the proposed methods and techniques are still far away from practical application, which is supported by the fact that studies on the reliability and precision of death time estimation are still scarce. Moreover, operating error tends to occur with complicated procedures.We therefore prefer to go back to the direct observation method and try to find quantified data by image analysis. Image processing and analyzing technique are employed to find features representing the quantified relationship between changes of corneal opacity and postmortem intervals. We applied K-Nearest Neighbor classifier, Adaptive Boosting classifier and Support Vector Machine classifier to evaluate the estimating result. This new technique is convenient and reliable, which is promising to solve the problem of lack of proper data for PMI estimation. Objective1. Establish a observation method to record the postmortem changes of rabbit’s cornea with high definition camera continuously which could be cut by every 15 minutes to get digital images.2. With matlab software, use a gray-level histogram based method for segmenting corneal-pupil region (object region) from eye images and extract nine features to reflect the changes of cornea at different postmortem intervals, including four color-based features and five texture-based ones which could build up a feature database for corneal images3. Analyze the features primarily and establish models with classifiers.4. Compare the three classifiers with 9 features.Material and methods1. Preliminary experiment:One healthy rabbit was sentenced to death by air embolism. The body was put in a dark and closed room with temperature of about (25±1)℃and the relative humidity of 20%-60%. Meanwhile, the cornea was exposed to air by stretching eyelids by using haemostatic forceps and was lighted by a desk lamp. From the time point of pronounced death up to 47 h, videos was taken continuously at cornea by employing digital camera (SONY, HDR-SR12, Japan) and was cut by every 15 minutes to obtain digital images of rabbit’s eye. The images gotten from 15 minutes to 47 hours after death were labeled with number one to 188.2.The eye images were segmented with gray level histogram method to divide into object region and nonobject region, then the object region was used.3.9 features were extracted with Matlab software, including GF, EL, G, S, C, J, Mean, Var and Ske. GF, gray-level based feature, presents the gray value difference of two regions in an images. MEAN, Means, the mean of gray value in the images, describes the average brightness of the whole images. VAR, Variance, the variance of images, in which, the lower of the variance means the changing of images is much smaller. SKE, Skewness,the Skewness of images, the greater asymmetry of distribution of images, the value of Skewness will be much bigger. EL, ratio of low-frequency energy to high-frequency energy, the higher El implies more energy is collected in low-frequency. G, contrast, textual features, the Contrast of images. S, Entropy, the informational entropy of images. The bigger value of entropy illuminates the distribution of texture in the image is more irregular. C, correlation, the correlation between the pixels in the images. J, Energy, the measure of uniformity of the gray-value distribution. The higher of the Energy’s value, the scale of texture is much bigger.4. The features of all the images from all the rabbits were classified with K-Nearest’ Neighbor classifier under 4-folds cross validations where the mean value was taken as accuracy rate. For each mission the experiment went through 5 rounds. The postmortem period from 15 minutes to 47 hours was divided into 3 intervals,4 intervals,5 intervals,6 intervals,8 intervals,10 intervals and 14 intervals (i.e. with 3 intervals there was about 15 hour for each interval in this 47 hours). The accuracy rate and predicting precision were evaluated under different interval length.5. Four healthy rabbits were labeled with R1, R2, R3 and R4, Their videos were taken and the images were obtained and analyzed with the same method stating in preliminary experiment. KNN classifier were applied as mentioned above.6. The predicting results were compared by using KNN classifier, Adaptive Boosting classifier and Support Vector Machine respectively with all features from images of four rabbits.4-folds cross validation methods were used and each experiment went 5 rounds.Results1. The classification model showed good results on the estimation of postmortem interval.2. The gray-level histogram based method effectively carried out segmentation in our experiment. Few images should be selected out and modified manually since the shadow under eye and the region of haemostatic forceps presents.3. All the nine features were capable for classification. The classification result of using single feature was weak. The combination use of 9 features improved the performance greatly.4. With the different amounts of intervals, there were good classification results both by using features from one rabbits alone and by features from the collection of four rabbits.5. As the amount of intervals increased, the accuracy rate decreased in each postmortem period by KNN classifier. While using features from each rabbit, the average accuracy rate was 97.1% with 3 intervals,88.5% with 8 intervals and 81.5% with 14 intervals. While using features from 4 rabbits, the accuracy rate was 96.9% with 3 intervals,87.6% with 8 intervals and 80.9% with 14 intervals.6. By adapting Adaboost classifier and using 9 features from each rabbit, the average accuracy rate was 94.4% with 3 intervals,85.3% with 8 intervals and 72.9% with 14 intervals; which were a little lower than KNN classifier. While using features from 4 rabbits, the accuracy rate was 85.1% with 3 intervals and 64.7% with 14 intervals, which were lower than KNN classifier.7. By adapting SVM classifier and using 9 features from each rabbit, the average accuracy rate was 88.3% with 3 intervals and 50.9% with 14 intervals; While using features from 4 rabbits, the accuracy rate was 78.9% with 3 intervals and 30.2% with 14 intervals. They were all much lower than the former two classifiers.Conclusions1. The classification model by using 9 features have been established successfully to estimate PMI.2. The gray-level based histogram method could segment the object region effectively.3. The 9 features extracted in our experiment could carry out PMI estimation reliably.4. In these experiments, KNN classifier is better than Adaboost classifier and SVM classifier. Background:Poisonings cause considerable morbidity and mortality, which influences the safety and healthy of humans beings worldwide. We collected the poisoning cases from our department and analyzed them in order to reflect the characteristic issues and the changing trends in central china with hope of helping improve public prevention and forensic examination.Material and methods:The records of 218 poisoning deaths in Hubei province of China from the Department of Forensic Medicine located at the middle region of China, Tongji Center for Medicolegal Expertise in Hubei (TCMEH), from 1999 to 2008 were retrospectively reviewed.Results:The majority (69.7%) of poisoning victims aged between the interval of 20 years old and 49 years old, and there was a male preponderance (male:female=1.7:1). The most common category of substances involved in poisoning deaths were rodenticide (19.7%), pesticide& herbicide (17.9%), carbon monoxide (16.5%), drugs (13.8%) and alcohols (12.4%). The manner of death, of the vast majority (64.7%), was accidental; suicidal intent was present in 25.2% of cases, homicide in 3.7%, and undetermined 4.1%. Ingestion was the predominant route of exposure (65.1%), followed in frequency by inhalation, injection and dermal. When compared with the former reports from the same institution, one for 1956-1984 and another for 1983-1999, an increase was found in the proportion of deaths due to rodenticides, CO, alcohols and drugs, as well as in accidental poisoning deaths. Conclusions:Poisoning deaths due to pesticides remain the major public health problem in China. Our government should carry out further regulatory enforcement to manage and restrict the application of pesticides and rodenticides which are were most dangerous to humans.更多还原