Improving Real-Time UAV Target Recognition via Super-Resolution, InceptionNeXt, and Attention-Enhanced YOLOv8

Abstract

— Accurate detection and recognition of clustered targets in aerial imagery remains difficult in Unmanned Aerial Vehicle (UAV) applications because targets are often small, blurred by low spatial resolution, affected by scale variation, partially occluded, and captured under uneven lighting. To address these issues, we present You Only Look Once (YOLO)v8-SR++, an improved UAV target recognition framework built on YOLOv8 that improves detection quality while preserving real-time speed. Rather than introducing entirely new component designs, the framework integrates and adapts several proven techniques within a single UAV-oriented pipeline. First, a Wide Activation Super-Resolution (WDSR) module enhances low-resolution inputs and restores fine details, helping the detector learn sharper features. Next, InceptionNeXt blocks are introduced in the backbone to strengthen multi-scale representation without heavy computation. In addition, a Separate and Enhanced Attention Module (SEAM) is placed in the neck to highlight important target regions and suppress background noise. Finally, to improve localization in dense and overlapping scenes, we use a hybrid loss that combines Complete Intersection-over-Union (CIoU) with Normalized Wasserstein Distance (NWD). Experiments on the VisDrone dataset show that YOLOv8-SR++ achieves 49.6% mAP@0.5 and 32.1% mAP@0.5:0.95, with 76.2% precision, 69.4% recall, and an F1-Score of 72.6%, while maintaining real-time inference at 43 FPS. These results indicate that YOLOv8-SR++ improves detection reliability for UAV aerial surveillance without sacrificing efficiency.