Lightweight UAV object-detection method based on efficient multidimensional global feature adaptive fusion and knowledge distillation

Unmanned aerial vehicles (UAVs) equipped with remote-sensing object-detection devices are increasingly employed across diverse domains. However, the detection of small, densely-packed objects against complex backgrounds and at various scales presents a formidable challenge to conventional detection algorithms, exacerbated by the computational constraints of UAV-embedded systems that necessitate a delicate balance between detection speed and accuracy. To address these issues, this paper proposes the Efficient Multidimensional Global Feature Adaptive Fusion Network (MGFAFNET), an innovative detection method for UAV platforms. The novelties of our approach are threefold: Firstly, we introduce the Dual-Branch Multidimensional Aggregation Backbone Network (DBMA), an efficient architectural innovation that captures multidimensional global spatial interactions, significantly enhancing feature distinguishability for complex and occluded targets. Simultaneously, it reduces the computational burden typically associated with processing high-resolution imagery. Secondly, we construct the Dynamic Spatial Perception Feature Fusion Network (DSPF), which is tailored specifically to accommodate the notable scale variances encountered during UAV operation. By implementing a multi-layer dynamic spatial fusion coupled with feature-refinement modules, the network adeptly minimizes informational redundancy, leading to more efficient feature representation. Finally, our novel Localized Compensation Dual-Mask Distillation (LCDD) strategy is devised to adeptly translate the rich local and global features from the higher-capacity teacher network to the more resource-constrained student network, capturing both low-level spatial details and high-level semantic cues with unprecedented efficacy. The practicability and superior performance of our MGFAFNET are corroborated by a dedicated UAV detection platform, showcasing remarkable improvements over state-of-the-art object-detection methods, as demonstrated by rigorous evaluations conducted using the VisDrone2021 benchmark and a meticulously assembled proprietary dataset.