Deep-learning based method for breech face comparisons

When a bullet is fired from a barrel, micro impression marks caused by the breech face on cartridge cases are one of the most critical factors in ballistic identification. This paper focuses on breech face impression images and introduces a deep-learning based algorithm, superpoint to extract interest point features and offer the local descriptor for each keypoint. Superpoint is a self-supervised framework for interest point detectors and descriptors. The classical brute-force matching, distance ratio matching and RANSAC methods are used to find out the correct matches. Validation experiments were performed on an image set with a total of 40 breech face impression samples, giving 63 pairs of known matching (KM) and 717 pairs of known non-matching (KNM) image comparisons The proposed method can still figure out the matching points for breech face impressions with random biases. The results illustrate that the superpoint and the feature matching methods are feasible for breech face impression image comparisons. Moreover, compared with SIFT, the proposed method performs better.