AN INTELLIGENT FRAMEWORK FOR AERIAL IMAGE DETECTION USING U-NET AND YOLOV10

Abstract

We provide a unique method that combines a U-Net architecture with YOLOv10 to handle the difficulties of few-shot aerial picture semantic segmentation, where unseen-category items in query aerial images must be identified using only a few annotated support images. Typically, few-shot segmentation involves segmenting query images pixel-bypixel using category prototypes that are extracted from support samples. However, objects in aerial photos frequently have irregular spatial distributions and arbitrary orientations, which results in notable differences in their properties. Low confidence scores and incorrect classification of same-category items with varying orientations are common outcomes of conventional approaches that do not take orientation changes into account. In order to get over these restrictions, our method combines YOLOv10 for quick and accurate object identification with U-Net for precise semantic segmentation, allowing for reliable item localisation even in challenging aerial scenarios. Because of YOLOv10's sophisticated detection capabilities, the system can recognise objects in a variety of sizes and orientations, producing trustworthy bounding boxes that direct the U-Net segmentation procedure. Combining these two architectures improves the network's ability to recognise rotated and dispersed aerial objects and reduces oscillation in confidence scores by reliably segmenting same-category objects independent of rotation or size. This U-Net + YOLOv10 framework successfully bridges the gap between effective object recognition and high-fidelity semantic segmentation in difficult aerial imagery by providing a scalable, rotation-robust few-shot aerial image segmentation solution.