Matching Forensic Sketches to Mug Shot Photos Using a Population of Sketches Generated by Combining Geometrical Facial Changes and Genetic Algorithms

摘要

Matching mug shot photos to forensic sketches drawn according to verbal descriptions of eyewitnesses is a decisive point for criminal investigations. However, the incapability of a witness to precisely describe the appearance of a suspect and his/her reliance on a subjective aspect of the description often lead to imprecise and inadequate sketches. This necessitates the development of robust automated matching methods such that least dependency exists on the quality of original sketches. The focus of the paper is on enhancing the preprocessing phase, before the matching phase is applied, by generating a population of sketches out of each initial sketch via applying geometrical changes in facial areas. The population is then optimized using Genetic Algorithms (GA) by adopting the Structural SIMilarity (SSIM) index as the fitness function. The matching is finally applied to the best sketch produced by GA by employing the Local Feature-based Discriminant Analysis (LFDA) framework. The efficiency of the proposed hybrid approach in achieving correct matchings is evaluated against 88 sketch/photo pairs provided by the Michigan State Police Department and Forensic Art Essentials, and 100 sketch/black-and-white photo pairs from FERET database. The experimental results indicate that our proposed approach obtains fairly better results relative to the LFDA framework. Furthermore, we notice a significant improvement in the retrieval rate if sketch/photo pairs are first cropped to central facial areas before a matching technique is applied.