Abstract
Unmanned aerial vehicles (UAVs) often operate under stringent resource constraints while requiring real-time object detection, which can lead to failures in cluttered backgrounds or when targets are small or partially occluded. To address these challenges, we introduce AAB-FusionNet, a real-time detection model specifically designed for UAV edge computing platforms. At its core is the Adaptive Attention Block (AAB), which employs an Adaptive Saliency-based Attention (ASA) mechanism to highlight the most discriminative tokens while a lightweight MBConv sub-layer refines local spatial features. This saliency-driven framework ensures the network remains focused on critical cues despite complex aerial imagery. To further boost performance, AAB-FusionNet utilizes a Multi-layer Feature Fusion Network that integrates three key components: Attentive Inverted Bottleneck Aggregation (AIBA) to restore significant details at multiple scales, DySample for preserving spatial fidelity during feature alignment, and the Dual-Attention Noise Mitigation (DNM) module to suppress environmental noise through complementary channel and spatial attention. Experiments on diverse aerial datasets confirm that AAB-FusionNet achieves robust detection, especially for small or partially occluded objects, while offering real-time inference on low-power hardware. Overall, AAB-FusionNet effectively balances accuracy, computational efficiency, and adaptability, making it ideally suited for UAV scenarios demanding fast, reliable object detection and robust and consistent performance.•Incorporates an Adaptive Saliency-based Attention mechanism to emphasize critical visual cues.•Introduces a Multi-layer Feature Fusion Network for detail restoration, feature alignment, and noise mitigation.•Demonstrates real-time, high-accuracy detection on low-power UAV platforms, particularly for small or occluded targets.