A system for forensic analysis of large image sets
Simske, Steven J.; Sturgill, Margaret; Everest, Paul; Guillory, George
Keyword(s): Image classification, image forensics, counterfeiting, security printing
Abstract: Variable Data Printing (VDP) offers the ability to uniquely tag each item in a serialized list, which increases product security. However, it is often not cost-effective or practical to employ VDP, in which case each item is printed with the same set of images. The brand owner should be able to use such "static" printed areas to gauge if the printing was performed by an authentic or a counterfeit printer/label converter. In this paper, we describe a system that uses a small set of pre-classified images (either authentic or counterfeit images from the same source) for initial training, and thereafter adaptively classifies and aggregates images from multiple sources as they join the population to be classified. Authentic images and multiple sources of counterfeit images are identified, and secondary links between the non-compliant samples are provided. The system currently uses a set of 420 metrics which are filtered to a smaller set of metrics that can reliably describe our known set. This filtered set of metrics, or metric signature, is used for the search and clustering thereafter. We describe the use of this system to streamline and enhance investigations for a global brand protection program. For such an application, the system is tuned for high precision.
Additional Publication Information: Presented at IEEE WIFS 2009, London, UK, Dec. 6-9, 2009.
External Posting Date: December 17, 2009 [Fulltext]. Approved for External Publication
Internal Posting Date: December 17, 2009 [Fulltext]