Content-Based Information Retrieval from Forensic Databases

Abstract

In forensic science, the number of image databases is growing rapidly. For this reason, it is necessary to have a proper procedure for searching in these images databases based on content. The use of image databases results in more solved crimes; furthermore, statistical information can be obtained for forensic identification. This thesis investigates the applicability of image-matching algorithms in forensic image database retrieval. From the experiments described in this thesis with the databases of toolmarks, cartridge cases, shoe marks and drugs tablets, it appeared that a reproducible image acquisition method should be used. Further, a pre-processing step is necessary for filtering the relevant regions of the image for the comparison. From the study in this thesis, is concluded that there is no single system to treat all image databases. For each database a different approach is necessary, depending on the marks that are used for the comparison. In order to control the use of these databases, methods are needed that predict the performance on larger databases. Representative test sets should be available for evaluating the performance of matching algorithms in the forensic databases. The performance of matching on a specific types of marks depends also on future developments in products that cause these marks. For instance, if firearms are used where the breech face marks are hardly visible or if the breech face marks are not characteristic anymore, this will reduce the results of using the cartridge case database. Forensic scientists often give a conclusion based on experience, not on statistics. The forensic databases could provide a reliability measure on the uniqueness of certain identifying characteristics., which results in usefulness in the interpretation of the evidence. For the shoe mark and drug tablet systems the identifying characteristics are often small irregularities. They are not visible in the low resolution images that have been used in our research. For the shoe mark and drug tablet systems, higher resolution images are necessary to obtain the characteristics. The cartridge case system and the tool mark system have these identifying characteristics available. More research on statistical measures from forensic image databases is desirable. The results of using the databases can not only result in data for court, but also for operational activities. By using the new investigative power due to fast treatment of image database, it is possible to derive information on the operational activities. Examples are new waves of drugs on the market and extensions of the traffic of those drugs, or the use of a specific kind of firearm in a group of criminals. In order to optimize information retrieval from the various databases, a data searching strategy should be developed that combines the information in the multiple databases. This is a way to alleviate the problem that often only partial profiles are found at the scene of the crime. In some cases, e.g. weak evidence from a shoe mark could be combined with weak evidence from a fingerprint which results in stronger evidence.