Methodology for Armaments Identification Using a Neural Network

Authors propose to use artificial neural networks (ANNs) for armaments identification based on the analysis of digital images. They describe that this problem is caused by an increase in: the number of armaments samples, requirements to efficiency and the need to automate the process of armaments identification on the basis of digital images analysis. The authors also propose a methodology for armaments identification using ANN such as GoogleNet, NasNet-Large, Xception, Places365-Google, DenseNet-201 and others. They give an example of this methodology implementation using GoogleNet ANN, pre-trained on the ImageNet dataset. In order to solve the problem of increasing the efficiency of armaments identification based on the analysis of digital images, the authors carried out the modification of the initial full-coherent layer of pre-trained GoogleNet by 5 classes from 1000 to 5 neurons and additional training of the GoogleNet ANN. ANN training was carried out using a set of digital images by five classes: BMD-4 (airborne combat vehicle); BTR-82 (armored personnel carrier); LINSA (armored ambulance), T-72 (T-72 tank), T-90 (T-90 tank). The selected optimal training parameters were: speed (step) - 0.00005, number of epochs - 20, optimization algorithm - Adam; validation frequency - 10, packet (batch) size - 25. The effectiveness of the proposed model was tested on a set of 571 images, the total number of classes - 5. The authors selected accuracy and learning error as the main indicators of the the neural network effectiveness. The result is a new trained model with identification (classification) accuracy of validation (test) sampling of 90% and with the probability of identification error - 0.1 that proves the selection of the ANN architecture and training parameters. The use of the proposed methodology allows to automate the process of armaments identification based on the analysis of digital images of elements.