NEURAL NETWORKS FOR HUMAN IDENTIFICATION IN SEARCH AND RESCUE OPERATIONS

Abstract

The paper proposes an approach to improving a human identification system on the images for application in search and rescue operations by using convolutional neural networks. A deep convolutional neural network with cascaded arrangement of convolutional and pooling layers has been developed. Experimental studies were conducted with various training optimizers, including Nadam, RMSprop, Adam, SGD, Adamax, and Adagrad.

Experiments with the different convolution kernel size showed that increasing the size from 3×3 to 5×5 improves model accuracy, while reducing to 2×2 leads to the accuracy degradation. The influence of the number of filters in convolutional layers was investigated: doubling the filters (64, 128, 256, 512 accordingly) increased efficiency, while their reduction (16, 32, 64, 128 accordingly) decreased model accuracy.

The best results were achieved with the Nadam optimizer, kernel size 5×5, and increased number of filters: F1-score 0.91 and accuracy 0.92. Tha application of data augmentation techniques enhanced the robustness of the proposed model against overfitting and improved model’s accuracy.