Gunshot entrance recognition by artificial intelligence using computer vision

The use of firearms as a means of facilitating crimes, such as robberies and homicides, has grown in several places around the world. However, recognizing this type of evidence is not a trivial task. Therefore, trace examinations are increasingly crucial to obtain information about a crime scene and criminal dynamics. Given this scenario, this work aimed to use resources based on computer vision to recognize different entries caused by caliber type on a white cotton T-shirt. The algorithm used was YOLOv11 (Ultralytics), based on convolutional neural networks. The samples comprised images of three firearms: a.38 caliber revolver and 9 mm and.357 caliber pistols. These were obtained with the Leica DVM6 digital microscope, totaling 110 images divided into 53 images of 9 mm caliber, 29 of.357 caliber, and 28 of.38 caliber. Due to the limited quantity, a methodology known as data augmentation was used, which increased the number of samples (totaling 436) without introducing new information into the system. These samples were divided into training (336 images) and validation (100 images). The training results indicate robustness for the prediction and stability of the model. The model quality parameters were all satisfactory. All samples were classified, and based on the confusion matrix, a 3 × 3 contingency table was constructed, and its analysis indicated parameters average above 90 %. Computer vision applied to forensic science problems is still in its infancy compared to other approaches. Still, it is growing and can provide complementary information with less subjective interpretation procedures.