Learning a similarity measure for striated toolmarks using convolutional neural networks

We propose TripNet a method for calculating similarities between striated toolmarks. The objective for this system is to distinguish the individual characteristics of tools while being invariant to class and sub-class characteristics, and varying parameters like angle of attack. Instead of designing a handcrafted feature extractor customized for this task we propose the use of a Convolutional Neural Network (CNN). With the proposed system 1D profiles extracted from images of striated toolmarks are mapped into an embedding. The system is trained by minimizing a triplet loss function so that a similarity measure is defined by the L2 distance in this embedding. The performance is evaluated on the NFI Toolmark database containing 300 striated toolmarks of screwdrivers published by the National Forensic Institute of the Netherlands. The proposed system is able to adapt to a large range of angles of attack between 15°and 75°, achieving a Mean Average Precision (MAP) of 0.95 for toolmark comparisons with differences in angle of attack of 15°to 45°; for differences of 15°to 60°a MAP of 0.78 is achieved.