An Improved Version of a Tool Mark Comparison Algorithm

Chumbley et al. (2010) described a statistically based algorithm for comparing pairs of tool marks. They presented empirical evidence that the algorithm produces well-separated similarity score values for "matching" and "non-matching" pairs of tool marks. However, the algorithm has two substantial weaknesses. First, it is "uncalibrated" in the sense that error rates can be determined only through empirical investigation. Second, it relies on a randomized test and can lead to different similarity scores when the algorithm is repeatedly applied to the same pair of tool marks. We present an improved version of the procedure, which eliminates the randomized scores and yields more consistent and predictable error rate control. This is accomplished by replacement of a random sampling step from the original algorithm with a deterministic process. We demonstrate the improved algorithm and compare its performance to the original by applying to known "matching" and "non-matching" pairs of tool marks.