Adaptive super-resolution integration to enhance object detection on low-quality unmanned aerial vehicle imagery

Abstract

The article addresses the problem of improving the accuracy of object detection in images captured by unmanned aerial vehicles under conditions of reduced spatial resolution and the presence of noise artifacts. The relevance of this research is driven by the practical need to maintain the reliability of computer vision systems in challenging field environments, where conventional detection algorithms tend to lose effectiveness. The aim of the study is to enhance the robustness of object detection in low-quality unmanned aerial vehicles imagery through the development of an adaptive preprocessing mechanism based on deep neural network–driven image super-resolution. The proposed approach involves the dynamic activation of the super-resolution module only in cases where image quality or detector confidence is insufficient. Within the framework of the study, a combination of the high-accuracy two-stage model Faster R-CNN and prior image upscaling using Real-ESRGAN is employed. An adaptive logic for triggering the image enhancement module is introduced, which is activated solely when the detector's confidence level falls below a defined threshold, thereby reducing computational overhead without compromising recognition performance. An experimental evaluation of the proposed method was conducted using unmanned aerial vehicles imagery degraded by various distortions, including blur, noise, and compression artifacts. The results demonstrate consistent improvements in detection accuracy across all tested image degradation types while maintaining acceptable processing time. The practical value of this research lies in its applicability to autonomous monitoring systems, search-and-rescue missions, and situational analysis tasks based on unmanned aerial vehicles video streams. The proposed approach opens up opportunities for further optimization by incorporating additional components, such as lightweight preliminary object filtering modules.