Gun-related cases pose extreme societal harm and urgently demand efficient and precise detection methods. This study aims to integrate the reflectance transformation imaging (RTI) technique with deep learning to apply it to the field of gun and bullet recognition. In the experiment, a total of 1500 samples of fired cartridge cases were selected from five QSZ92 9 mm pistols. Detailed images of the markings on the base of the cartridge cases were captured using the DTV3.1 intelligent imaging system to obtain their normal maps. Thereafter, the pre-trained ResNet-50 network extracted features from the normal maps and underwent classification training. The model’s performance was evaluated by outputting AUC values, accuracy on the test set, and a confusion matrix. The experimental results reveal a total AUC value of 0.98 across the five guns, with gun No. 2 achieving the highest accuracy of 97.66% and gun No. 1 the lowest at 93.75%. This study demonstrates that the automatic recognition method of cartridge case marks based on RTI technology and deep learning yields significant results, offering valuable reference for the identification of other traces in the field of trace inspection.

The presence of close non-matches poses inherent risks to fingerprint identification, and exploring their occurrence patterns in different regions of the fingerprint can enhance the risk awareness of identification personnel. To investigate this, 30 fingerprints from the central and branch regions of loops were selected as research subjects. Each fingerprint was queried 15 times within a database of ten million individuals to examine the occurrence of homologous fingerprints and close non-matches when 6 to 20 minutiae were marked. The results revealed that close non-matches could be found in both the central and branch regions of loops when 6 to 20 minutiae were marked, and notably, the number of close non-matches in the branch region was significantly higher than in the central region. The proportion of homologous fingerprints ranked higher than close non-matches in the central and branch regions of loops, with respective percentages of 98.4% and 60.9%. Even when 20 minutiae were marked, there were still instances where homologous fingerprints did not appear in the branch region of the loop. Frequency analysis of similar regions indicated that the side of the central region facing the loop and the side of the branch region facing the branch were more susceptible to having close non-matches. Additionally, high-level close non-matches with over 10 matching points were identified in both the central and branch regions of loops. Therefore, fingerprint identification personnel must handle close non-matches with caution to prevent them from interfering with accurate identification.