Abstract
This paper addresses the problem of collecting and preprocessing data acquired from SWIR cameras for training a neural network to recognize objects under limited-visibility conditions. The relevance of the study is обусловлена the high informativeness of short-wave infrared images in fog, smoke, and low-light environments; however, such data are affected by InGaAs sensor noise, defective pixels, and reduced contrast. A specialized dataset containing images of pedestrians and vehicles captured in various outdoor conditions was created. A two-stage preprocessing pipeline was proposed, including non-uniformity correction (NUC) and contrast-limited adaptive histogram equalization (CLAHE). The YOLOv8 architecture was used to evaluate the effectiveness of the proposed approach. Experimental results showed that the method increased PSNR by 18.4% and SSIM by 12.6%, while improving object recognition accuracy up to 0.87 mAP@0.5 for vehicles and 0.79 for pedestrians. The results demonstrate that the developed pipeline can be applied in real-time all-weather monitoring systems and autonomous transport solutions.