Centro de Documentação da PJ
Monografia
 CD143 |
Improving facial reproduction using empirical modeling [Documento electrónico] : final technical report / J. Wesley Hines .. [et al.].- Knoxville : The University of Tennessee, 2010.- 1 CD-ROM ; 12 cm. - (Report No. 2008-DN-BX-K183 /UTNE/2009-1)
Resumo extraído do próprio documento. Ficheiro de 4,49 Mb em formato pdf (105 p.).
RECONSTITUIÇÃO FACIAL, ANÁLISE DE VESTÍGIOS, ANÁLISE LABORATORIAL, CIÊNCIA FORENSE, IDENTIFICAÇÃO DE PESSOAS, ESTADOS UNIDOS
"Forensic facial reconstruction has been used for many years to identify skeletal remains. The face of the unknown person can be reproduced based upon the soft facial tissue thickness, which overlays the bony structure of the skull. Currently, forensic artists place average facial tissue markers at 21 specific anatomical locations on the skull and use clay to model the face based on the length of the markers. The purpose of this study was to develop a new technique for estimating the facial soft tissue thickness at the 21 traditional craniometrical landmarks used in forensic facial reconstruction. The soft tissue thicknesses or marker lengths used in the forensic facial reconstruction are currently the average tissue depths of various examined cadavers of different ethnicity, sex, and body type; whereas this new technique uses a non-parametric modeling technique to predict the facial tissue depths based on a unique skull input. The development of the methodology for this pilot study has been completed. The 100 live male subjects' Computed Tomography (CT) images have been identified. Extended amounts of time were spent researching feasible ways of accurately measuring the tissue thicknesses at the desired craniometrical locations. Problems occurred in trying to align and orient, called registering, the database of male skulls with a base skull that has markers at the 21 locations where the tissue thickness must be measured. A viable option to automatically register CT images and consistently take accurate measurements for the entire database was found in a software package, IDAS. It is being developed by Dr. Mohamed Mahfouz at the Center for Musculoskeletal Research (CMR) at The University of Tennessee. Computed Tomography (CT) images of 100 Caucasian male subjects' skulls were used to build a database of facial tissue thicknesses and input predictors for the non-parametric model. The inputs to the model are various cranial bone thicknesses and measurements along specific anatomical lines, which then are used to predict the facial tissue thicknesses at the traditional landmarks using a Non-Parametric Kernel Regression model. The tissue and bone measurements were performed using the software package IDAS. Hetero-Associative Kernel Regression (HAKR) and Inferential Kernel Regression models were built using the measurements from the 100 male subjects. The performance of the empirical model was initially judged by comparing the predicted facial tissue depths to the actual known tissue depths for each skull in the database using a Leave One out Cross Validation (LOOCV) technique. Two results were computed for each model; one including the query’s demographics as a predictor and the other with demographics removed from the model. The Root Mean Squared Error (RMSE) when not using the demographics as an input to the model was 2.21mm for the HAKR architecture and 2.19 mm for the Inferential. When including the demographics, the RMSE for the HAKR architecture was 2.04mm and 1.89mm for the inferential architecture. The HAKR and Inferential model's RMSE were both less than the currently used tabled tissue thickness RMSE from the actual measured tissue thicknesses of 3.07 mm. The newly developed inferential model provided forensic facial tissue thickness approximations with an average of 38% less error when using demographics or 29% less error when not using demographics. The error reduction is based on the tabled tissue thicknesses that are used in facial reconstructions today. The similarity metric for the models' predictions, after a LOOCV, was computed to be 0.19 and 0.81 for the HAKR and inferential models, respectively. This metric displays how close the query input skull is to one of the memory skulls based on their input parameters and can be used to give an estimate of the model’s prediction confidence. The average prediction uncertainty from the LOOCV was computed to be 19.7% for the HAKR model and 20.5% for the inferential model. The above quantitative results supporting the empirical model’s ability to accurately predict facial tissue thickness were put to the test on 3 male skulls from the William Bass Donated Collection at The University of Tennessee. A certified forensic artist was used to construct the facial reconstructions. The facial reconstructions using tabled tissue thicknesses were compared to the reconstructions of the same subject using the inferential models’ predicted tissue thicknesses. Certain landmarks, particularly the zygomatics, were estimated too large, but overall the model seemed to do a better job of estimating the face in each case."