Automated sex identification from 3D skulls based on patch representations for forensic examinations

This thesis introduces new perspective of addressing skull sex determination by investigating the possibility of using computer vision based solutions for sex determination from skull data, and developing new algorithms tailored for mitigating the underlying problems experienced in the conventional forensic sex determination methods. First, this thesis proposes the use of state-of-the-art 3D local shape descriptors to derive patch based representations from the skull. This is achieved by geometrically partitioning the skull into local regions, and aggregating the features from the regions into compact representation, which summarize the regional shape properties. Several rigorous experiments are conducted using this approach, and the results show that the proposed patch based representation is comparable to conventional forensic methods with a sex prediction rate of 86%. Traditional morphometric analysis on the same dataset produced prediction rate of 86.4%, while PCA/LDA based method produced prediction rate of 75% which indicates that the proposed patch based representation is competitive and an effective method of sex determination. Second, this thesis proposes a new approach of partitioning the skull into pseudo-anatomical regions, which is reminiscent of the anatomical regions commonly examined in forensic morphological assessment. In order to achieve this, the skull samples are registered to establish correspondence across the dataset using an improved version of coherent point drift (CPD) algorithm named Ensemble CPD. The correspondence established is exploited to partition the skulls using Fuzzy cmean (FCM) into regions which share visual resemblance with the anatomical regions in forensic morphological assessment. Further, 3D local shape descriptors are computed from each region, aggregated into compact representation, and eventually used for sex determination. The best result in this experimental evaluation is 86.5% which is also comparable to conventional forensic methods. Finally, this thesis proposes two group variable selection algorithms which are (1) novel group regression model for manifold data and (2) novel group regression model for manifold data with split augmented lagrangian algorithm, for automatically discovering discriminant (important) skull regions. The main objective of these methods is to prune out the salient regions of the skull that contribute positively to sex prediction performance. Forensic methods have demonstrated that some anatomical parts are more sexually dimorphic than others, however such analysis is performed using stepwise variable selection. Thus, this thesis introduces new techniques for automatically learning such information directly from the data. The best result in this experiment is 84.5% which is comparable to conventional forensic methods