From StirMark to StirTrace: Benchmarking pattern recognition based printed fingerprint detection

Abstract

Artificial sweat printed fingerprints need to be detected during crime scene investigations of latent fingerprints. Several detection approaches have been suggested on a rather small test set. In this paper we use the findings from StirMark applied to exemplar fingerprints to build a new StirTrace tool for simulating different printer effects and enhancing test sets for benchmarking detection approaches. We show how different influence factors during the printing process and acquisition of the scan sample can be simulated. Furthermore, two new feature classes are suggested to improve detection performance of banding and rotation effects during printing. The results are compared with original existing detection feature space. Our evaluation based on 6000 samples indicates that StirTrace is suitable to simulate influence factors resulting into overall 195000 simulated samples. Furthermore, the original and our extended feature set show resistance towards image manipulations with the exception of scaling (to 50 and 200%) and cropping to 25%. The new feature space enhancement is capable for handling banding, rotation as well as removal of lines and columns and shearing artifacts, while the original feature space performs better for additive noise, median cut and stretching in X-direction.