HarmoNet: Balancing Computational Efficiency and Detection Accuracy for Drone-Based Object Recognition

Drone-based object detection faces significant challenges in aerial imagery due to the prevalence of clustered and occluded objects which severely impact detection accuracy and computational efficiency. Clustered objects create overlapping bounding boxes and ambiguous boundaries, while occlusion leads to missed detections and false merging of multiple objects. To address these challenges while maintaining computational efficiency, we propose HarmoNet, an innovative framework that balances detection accuracy with reduced model complexity. HarmoNet incorporates three key modules: (1) Harmonized Feature Pyramid Network (HarmoFPN), a streamlined neck structure that enhances information flow and feature fusion across different scale levels while maintaining architectural simplicity; (2) Spatial Channel Attention Module (SCAM), which captures crucial positional and edge information to improve bounding box precision for better object localization; and (3) Reparameterizable Detection Head (RepDetHead), designed to reduce computational complexity in the detection head while preserving detection performance. We conduct extensive experiments on multiple public UAV datasets to validate our approach. The proposed HarmoNet demonstrates superior performance on UAVDT, CARPK, and DIOR datasets, effectively handling both clustered and occluded objects in aerial scenarios. Our experimental results show that HarmoNet achieves an optimal balance between computational efficiency and detection accuracy, making it particularly suitable for real time drone-based object recognition applications where both performance and resource constraints are critical considerations.