Классификация изображений следов бойков по экземплярам оружия с помощью полносвязной нейронной сети

Abstract

Introduction. The aim of the work is to increase the efficiency of identification of firearms by images of firing pin marks in the automatic mode. The relevance of the task is determined by the low efficiency of the known methods of automatic identification of firearm by the firing pin marks with individual topological types of individualizing features. This affects the investigation of crimes related to the use of firearms. Formation of clone images. A training sample was formed; it included 140 original images of firing pin marks from 50 classes, on the basis of which about 1000 clone images were made with slightly modified individualizing features. In this case a specific specimen of a firearm is meant as a class. Neural network training. A fully connected neural network with the following architecture was used as a classifier: an input layer of neurons; two hidden layers; an output layer. The input layer included 2500 neurons, the first hidden layer was made up of 625 neurons, the second hidden layer contained 156 neurons; the output layer consisted of 50 neurons (in accordance with the number of the classes). Evaluation of the calculation results. The prediction accuracy of the trained neural network was estimated according to the Accuracy metric, which is the ratio of the number of correct predictions to the total number of predictions. The prediction accuracy for the maximum signal on one output neuron was 81%, and when the maximum signals on three output neurons were taken into account, the accuracy was about 91%. Conclusions. The research has shown the possibility of classification of the images of firing pin marks by weapons using a fully connected neural network, as well as the effectiveness of using artificially generated clone images of firing pin marks for training a fully connected neural network in cases with a small number of initial objects.