Assessment of the resistance of neural network models to noise and artifacts in complex operating conditions

Abstract

Assessment of robustness of neural network models to noise and artifacts in complex operating conditions. Neural networks are widely used to solve a variety of complex tasks: from image recognition to natural language processing. Many modern scientific works are aimed at increasing the efficiency of the use of neural networks in various applications and software operation. In the course of this work, an analysis of the issue of robustness of neural network models to noise and artifacts in difficult operating conditions was carried out and its assessment was provided. It was found that in any system there are noises created by the internal physical properties of both memory devices and peripheral circuits. The influence of these noises increases even more in difficult operating conditions, which cause additional artifacts. It is mentioned that image denoising is one of the fundamental problems in the field of image processing, as it is a mandatory step in many applications related to computer vision. This is especially relevant in the medical and military fields, where noise reduction algorithms are used to obtain high-quality X-ray images in computer tomography systems and in satellite imaging. The issue of preserving the quality of the image provided by reconnaissance drones is also important, since some means of radio-electronic warfare can spoil the quality of the image or make it completely illegible. To protect clarity, neural network models are now actively used, as they show quite good results and are capable of subsequent adaptation and training. The study highlights the potential of neural networks to synthesize and supplement sensitive medical data in the context of clinical respiratory disease detection. The obtained results of the current study can be involved in the future for similar tasks in computer vision systems, for example, during the analysis of images in video surveillance systems