Lightweight Detection Network for Small Ground Targets from UAV Based on Multi-Dimensional Feature Representation and Reconstruction

Approaches that aim for high accuracy typically rely on large input sizes, deep network architectures, and intricate multi-scale fusion mechanisms, leading to excessive computational burden and memory consumption. Conversely, lightweight strategies such as network pruning, channel sparsification, or input down-sampling inevitably sacrifice spatial details and semantic information, causing small object features to be overwhelmed by background noise and resulting in severe missed detections and false alarms.

Method

To address these challenges, this paper proposes a lightweight UAV ground small object detection network, termed Difficulty Adaptive Frequency Enhanced Network (DAF-Net), which integrates multi-dimensional feature representation and reconstruction strategies to achieve high detection accuracy while maintaining low computational overhead. The proposed DAF-Net consists of three core modules designed to tackle the key bottlenecks in lightweight small object detection.

First, a multi-feature fusion difficulty-adaptive loss weighting module (DALW) is introduced to address the imbalance in training resource allocation caused by inherent sample difficulty variations. Specifically, DALW quantifies sample difficulty from four complementary perspectives: image complexity, object distribution, detection uncertainty, and matching quality. A comprehensive difficulty score is computed via weighted geometric mean, and samples are discretized into difficulty levels using the Jenks natural breaks method.

Loss weights are then dynamically assigned based on these levels, enabling the model to focus on valuable hard samples such as small, occluded, or low contrast objects during optimization. Second, a high-resolution pyramid detection head (HRPH) is constructed to enhance small object feature representation. Building upon the conventional P3–P5 feature pyramid, HRPH introduces an additional P2 layer that retains high-resolution spatial details.

Bidirectional path aggregation is employed—top-down semantic information flows into shallow layers while bottom-up fine-grained details propagate to deeper levels—effectively integrating global context with local edge textures and significantly improving localization accuracy for small objects. Third, a frequency-domain multi-scale enhancement module (FME) is developed to expand the receptive field without incurring substantial parameter overhead.

Leveraging discrete wavelet transform (DWT), FME decomposes input feature maps into low-frequency contour components and high-frequency detail components. Small kernel convolutions are independently applied to each frequency band, followed by inverse wavelet transform (IWT) to reconstruct the enhanced features. This design enables receptive field expansion with negligible parameter increase, allowing the network to capture broad contextual information while preserving fine-grained edge details.

These three modules synergistically optimize the feature representation and reconstruction process from the perspectives of sample difficulty perception, spatial detail reconstruction, and receptive field expansion. Result Extensive experiments are conducted on the VisDrone dataset, which comprises over 10,000 aerial images with diverse scenes and challenging small object instances. Quantitative results demonstrate that the proposed DAF-Net achieves 43.1% mAP50 and 26.5% mAP, with only 9.6 million parameters and 33.4 GFLOPs.

Compared to mainstream general purpose detectors including YOLOv8, RT-DETR, and YOLOv11, DAF-Net achieves 5.7% higher mAP50 while maintaining comparable model size. Compared to state-of-the-art specialized lightweight UAV detectors such as MSFE-YOLO-s (41.4% mAP50, 31.6M parameters) and LWUAVDet-S (33.5% mAP50, 5.2M parameters), DAF-Net strikes a superior balance between accuracy and efficiency, achieving the highest detection accuracy with moderate parameter count.

Ablation studies systematically validate the contribution of each module: integrating DALW improves mAP50 by 1.4% over the baseline, HRPH contributes a 2.7% gain, and FME yields a 0.9% gain. The full combination of all three modules achieves a cumulative 5.7% improvement, demonstrating their synergistic effect. Furthermore, visualization analysis using class activation maps reveals that HRPH expands the model's perceptual range, FME deepens responses in salient regions, and DALW adjusts attention distribution to balance easy and hard samples—collectively resulting in both broad and focused activation patterns.

Confusion matrix analysis confirms that the proposed modules progressively reduce inter-class confusion and background misclassification, particularly for small and easily confused categories. Generalization experiments on four unseen datasets—ERA (Event Recognition in Aerial videos), SDD (Stanford Drone Dataset), UAV123, and AU AIR—with varying resolutions and scene characteristics further demonstrate the robustness and adaptability of the proposed method, with most in-class objects successfully detected under diverse conditions including viewpoint changes, varying sensor parameters, and resolution variations.

Conclusion This paper presents a lightweight UAV ground small object detection network, DAF-Net, which effectively balances detection accuracy and computational cost through the synergistic optimization of difficulty-aware weighting, spatial detail reconstruction, and frequency-domain multi-scale enhancement. The proposed DALW module enables adaptive allocation of training resources toward hard samples, the HRPH module preserves fine-grained spatial details crucial for small object localization, and the FME module expands receptive field with minimal parameter overhead.

Extensive experiments on multiple datasets demonstrate that DAF-Net achieves state-of-the-art performance among lightweight UAV detectors, with superior generalization capabilities across diverse scenarios. The method provides a practical and effective solution for real-time ground small object detection on resource-constrained UAV platforms, with potential applications in traffic monitoring, battlefield reconnaissance, and emergency response systems.

Future work will focus on further reducing model complexity while maintaining detection accuracy, exploring adaptive mechanisms for dynamic resource allocation in edge deployment scenarios. The code of this paper is publicly available at: github.com/fzkicw/DAF-Net.