Identification automatisée de marques descriptives sur des images de douilles de cartouche

Abstract

Crime scenes involving firearms need to be carefully analyzed to present a body of evidence at trial. It is useful for investigations to be able to associate a firearm with a crime scene. When a cartridge is fired, the mechanisms of the firearm leave descriptive marks on the bullet and the cartridge case, and when cartridge cases are found at a crime scene, these marks are analyzed in a laboratory. These marks may be class characteristics, which are common to firearms in the same family; or individual characteristics, which are theoretically unique and enable a specimen to be linked to a specific firearm. Before individual characteristics are compared to propose matches, manual sorting is generally carried out by considering class characteristics, to avoid unnecessary calculation of comparisons with specimens produced by different gun families. An automatic analysis would assist technicians when entering information, by suggesting the categories present. The main objective of this thesis is to determine a machine learning method for identifying microscopic marks present in images of cartridge cases. The first contribution consists in classifying the marks according to seven categories: parallel, arch, crosshatch, circular, granular, smooth and unknown. We trained and evaluated multiclass, multilabel and binary models. Evaluation of the ENB3 multilabel model without augmentation shows an F1 measure of 54.47%, and a loss of 0.36. The evaluations of the binary models without augmentation show values between 70.00% and 93.00% for the F1 measure, and values between 0.22 and 2.01 for the loss function. During these experiments, we noticed inconsistencies in the ground-truth labels and suggest relabelling the data. The second contribution consists in evaluating clusters of cartridge case images. These clusters were created by traditional algorithms and a deep clustering network, based on the descriptive features of the images extracted by the networks trained in the previous objective. We observed that algorithms favouring the creation of clusters of various shapes and sizes seemed better suited to our data, and we appreciated the fuzzy clustering technique, which calculates a degree of membership of points to each group. We suggest future work combining these two approaches. The final contribution study inter-observer agreement between six human observers and supervised deep learning methods. We observed average Kappa coefficients ranging from low for the granular and smooth categories, to high for the circular category. We subsequently built a ground truth set of 1,000 samples verified by two experts. We repeated the multilabel experiment with this set, without data augmentation. We observed an improvement, with a training curve reaching 90.00% accuracy. In the evaluation, we observed an improvement in the weighted average for the F1 measure to 77.12%. We believe that this model demonstrates better capabilities for the classification of cartridge case images, and we suggest improving the ground truth set by adding additional samples, particularly those in the more uncommon categories. In summary, this thesis shows that deep learning could be used to automatically identify descriptive marks on images of cartridge cases. Eventually, classification models could be included in automated identification systems. As well as improving the performance of matching algorithms by reducing computation times, they could provide information to investigators promptly, which could help to advance ongoing investigations faster. The models will, however, need to be improved to include rarer categories, as well as certain exceptions (such as marks produced at times other than when the weapon was being used). We recommend continuing research with binary classifiers while increasing the verified ground truth set.

Date	23 Oct 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Luc Duong (Supervisor) & Sylvie Ratté (Co-supervisor)