Evaluation of the target identification effectiveness by an intelligent thermal imaging device based on a pre-trained convolutional neural network

Abstract

Background. The rapid development of technology in recent years has significantly expanded the range of the proposed element base of thermal imaging devices and allowed the use of various automated target identification systems, including those based on convolutional neural networks. The existing methods for assessing the range of thermal imaging devices are based on the Johnson criterion used to evaluate the “target – thermal imaging device – operator” system; the issue of their application to the “target – thermal imaging device – recognition system” systems is currently unresolved. In the conditions of increasing the share of tasks performed by intelligent vision systems based on convolutional neural networks, an important task remains to substantiate rational requirements for them, using criteria that take into account the peculiarities of image perception by neural networks. The purpose of the study is to determine criteria that allow to reliably assess the ability to identify a target with an intelligent thermal imaging device based on convolutional neural networks. Materials and methods. Based on the most effective models of convolutional neural networks, previously trained on a sufficient number of thermal imaging images of seven types of targets, such parameters of the image containing the target images as SNR, PSNR and SCR were analyzed by statistical methods. On the basis of the obtained statistical base, the correlation coefficients of the recognition probability from the image parameters for targets located at different ranges and having different areas relative to the entire image are analyzed. Results. Regression equations are obtained that allow, based on the geometric parameters of the target – Sц and the peak signal–to-noise ratio of the image - PSNR, to determine the probability of 50% and 100% identification of the target by a convolutional neural network. The methodology for estimating 50% and 100% of the target identification range by an intelligent thermal imaging device based on a pre-trained convolutional neural network has been refined. Conclusions. The obtained dependencies make it possible, based on the so-called “geometric parameters”, namely the focal length of the lens, the pixel size of the radiation receiver matrix, as well as the size of the target, to accurately calculate the values of the minimum target identification range with a probability of 50% and 100%, the error does not exceed 14%, and also to determine the minimum required PSNR ratio of the image, in which the neural network is able to provide stability when identifying targets.

Keywords

• thermal imaging devices
• identification range
• focal length
• pixel size of the photodetector
• convolutional neural network
• image quality