neural network

Night vision devices (NVD) and thermal imaging cameras are widely used in the fields of surveillance, security and monitoring, search and rescue, emergency response, industrial control and maintenance, ground operations and unmanned aerial vehicles. Night vision devices operate on the principle of amplifying residual light (starlight, moonlight, city lighting) in the visible and near-infrared ranges (0.4–0.9 μm). Thermal imaging devices record thermal radiation of objects in the medium (3–5 μm) or far (8–14 μm) infrared range.

The paper analyses the system characteristics and functional capabilities of multi-bit pyramidal adders that can be used in the structures of discrete perceptron of modern neural networks for summing weight coefficients and input signals. The methodology for designing multi-bit adders using flow and spatial-temporal graphs is described. Algorithmic and recursive pyramidal adders have been developed and their main system characteristics have been determined.

The paper addresses various aspects of optimizing deep video enhancement models for efficient execution on modern hardware. The focus is on a multi-frame generative network with multi-scale structure and frame-by-frame smoothing (MST-GAN). A comprehensive hardware acceleration strategy is proposed, which includes structural thinning, quantization (FP16/INT8), pipeline, parallelization, and model compilation using TensorRT. A comparative analysis is performed before and after optimizations, including changes in FPS, latency, memory consumption, and FLOPs.

The paper considers the problem of enhancing image quality using neural networks with different types of multi-level distortions - contrast degradation, adding noise of various natures, image compression, etc. The widespread TID2013 database, which contains both original images and images modified using various types of distortions (25 basic images, 24 types of distortions, and 5 of their levels), was used as the image database for training neural networks. This database was divided into training (480 images), validation (360), and test (120) images.

This paper presents an approach to improving mobile robot localization by filtering dynamic objects using camera image segmentation. The proposed algorithm integrates a Particle Filter with state-of-the-art computer vision techniques, specifically employing the YOLO model for segmentation, which effectively differentiates static elements of the environment from moving objects. This approach reduces the impact of noisy data and enhances localization accuracy in dynamic conditions, which is crucial for the reliable autonomous operation of mobile robots.

The paper considers problems for the tasks of improving the quality of digital video images for cloud environments, as well as on the client side using generative adversarial neural networks (GANs) adapted for work in the browser. A method is proposed that uses WebGPU for accelerated execution of convolutional calculations, which allows to increase the resolution and improve the quality of low-quality video in real time without significant load on servers.

The article is devoted to the theoretical foundations and practical aspects of facial recognition using neural networks. The paper analyzes various approaches to feature extraction from facial images, as well as methods for training neural networks for recognition. Special attention is paid to problems related to variations in lighting, facial expressions, and other factors affecting recog- nition accuracy. The paper describes the creation and optimization of a neural network for real-time human face recognition.