SSCW-YOLO: A Lightweight and High-Precision Model for Small Object Detection in UAV Scenarios

Zhuolun He, Rui She, Bo Tan, Jiajian Li, Xiaolong LeiDrones 2026 · DOI

TL;DR

This paper introduces SSCW-YOLO, a novel object detection model built on YOLOv8, designed to significantly improve the detection of small objects in challenging UAV scenarios. By introducing a Spatial Cosine Convolution (SCoConv) module for enhanced feature extraction, a C2f_ScConv module for efficient feature fusion, and the Wise-IoU (WIoU) loss function, the model achieves state-of-the-art performance with improved accuracy and speed, making it well-suited for real-world drone applications.

What Problem Does This Paper Address?

Detecting small objects in images captured by Unmanned Aerial Vehicles (UAVs) is a notoriously difficult task. UAVs often fly at high altitudes, resulting in small objects appearing with indistinct features, making them prone to being missed or misclassified. Furthermore, complex backgrounds, occlusions, and the sheer number of overlapping objects in aerial imagery exacerbate these challenges. Existing object detection models, even powerful ones like YOLOv8, struggle to maintain high accuracy and speed when faced with these specific difficulties, leading to missed detections and false positives. This paper aims to bridge this gap by developing a more robust and precise model specifically for UAV-based small object detection.

Methods

The authors build upon the YOLOv8 architecture, introducing three key innovations:

  • Spatial Cosine Convolution (SCoConv): This novel module is designed to enhance spatial feature representation. It combines a spatial reconstruction mechanism with cosine similarity to better capture fine-grained spatial structures and suppress noisy features. This "Filter-then-Enhance" pipeline aims to preserve delicate details crucial for identifying small objects.
  • C2f_ScConv Module: An improved feature fusion module that builds upon YOLOv8's C2f module. It incorporates spatial-channel reconstruction convolution to selectively fuse features, effectively reducing redundancy in both spatial and channel dimensions. This leads to a more efficient model with lower complexity and computational cost.
  • Wise-IoU (WIoU) Loss Function: The standard CIoU loss is replaced with WIoU, which uses a dynamic non-monotonic focusing mechanism. This helps to mitigate the issue of uneven quality in anchor boxes, particularly prevalent in small object datasets, by allowing the model to focus more on low-quality anchors and improving bounding box regression.

Datasets

The primary dataset used for training and evaluation is VisDrone2019. This dataset is described as a comprehensive benchmark for UAV-based object detection, containing a large number of frames and images with ten object categories. It is noted for its characteristic long-tail distribution of classes and the prevalence of small-scale objects and complex backgrounds.

For generalization experiments, the authors also utilize:

  • DOTA (Dataset for Object Detection in Aerial Images): This dataset contains objects of 16 categories with varying sizes, orientations, and shapes, captured from different sensors and platforms, making it suitable for evaluating performance on aerial imagery.
  • SSDD (Ship Detection Dataset): This dataset focuses on ships under various sea conditions, lighting, types, and sizes, allowing for evaluation in maritime scenarios.

Key Insights and Ideas

  • The "Filter-then-Enhance" Synergy: The SCoConv module's effectiveness stems from its sequential application: first filtering out noise and then enhancing directional cues. This is more effective than applying a single type of convolution alone.
  • Selective Fusion is Key: Simply concatenating features, as in some standard fusion modules, can introduce noise. The C2f_ScConv's ability to selectively fuse features based on spatial-channel reconstruction is crucial for reducing redundancy and improving efficiency.
  • Loss Function Matters for Small Objects: Standard IoU-based losses can be penalized by geometric factors that are less relevant for small, often distorted objects. WIoU's dynamic focusing mechanism helps the model better handle the inherent challenges of small object anchor boxes.
  • Holistic Optimization Pays Off: The paper emphasizes that the significant improvements are not just from individual components but from their synergistic integration, creating a feedback loop where enhanced features lead to better loss guidance, and vice-versa.

Key Results

  • VisDrone Performance: SSCW-YOLO achieves a mAP50 of 37.8% on the VisDrone dataset, a 5.4% improvement over the baseline YOLOv8n. It also achieves 115.1 FPS, indicating a speedup.
  • Small Object Specialization: The model shows a dominant improvement on small objects, achieving an AP50 of 22.4% for small objects on VisDrone, a 5.6% increase over YOLOv8n.
  • Generalization: SSCW-YOLO demonstrates strong generalization, achieving 98.7% mAP50 on the SSDD dataset and 42.3% mAP50 on the DOTA dataset, outperforming baseline models.
  • Efficiency: The model has 2.61 million parameters, which is fewer than the baseline YOLOv8n (3.01 million), and a computational cost of 8.7 GFLOPS, only slightly higher than YOLOv8n (8.1 GFLOPS).

Limitations and Future Work

  • Computational Cost: While efficient, the proposed model incurs a modest increase in FLOPs (8.1 G to 8.7 G) compared to YOLOv8n, which might be a consideration for ultra-low-power micro-drones.
  • Extreme Conditions: The current evaluation does not cover extreme conditions like low-light or adverse weather.
  • Future Work: The authors suggest:
  • Channel pruning of the C2f_ScConv module for further lightweighting.
  • Augmenting training data with specialized UAV scenarios (e.g., agricultural pest monitoring in low-light).
  • Optimizing inference pipelines for edge devices.
  • Expanding the model's application to specialized UAV tasks like nighttime agriculture and high-altitude reconnaissance.

Summary updated

Abstract

To address the problems of missed and false detections caused by insufficient feature quality in small object detection from UAV perspectives, this paper proposes a UAV small object detection algorithm based on YOLOv8 feature optimization. A spatial cosine convolution module is introduced into the backbone network to optimize spatial features, thereby alleviating the problem of small object feature loss and improving the detection accuracy and speed of the model. An improved C2f_SCConv feature fusion module is employed for feature integration, which effectively reduces feature redundancy in spatial and channel dimensions, thereby lowering model complexity and computational cost. Meanwhile, the WIoU loss function is used to replace the original CIoU loss function, reducing the interference of geometric factors in anchor box regression, enabling the model to focus more on low-quality anchor boxes, and enhancing its small object detection capability. Ablation and comparative experiments on the VisDrone dataset validate the effectiveness of the proposed algorithm for small object detection from UAV perspectives, while generalization experiments on the DOTA and SSDD datasets demonstrate that the algorithm possesses strong generalization performance.
BibTeX
@inproceedings{he2026sscw,
  author = {Zhuolun He and Rui She and Bo Tan and Jiajian Li and Xiaolong Lei},
  title = {{SSCW-YOLO: A Lightweight and High-Precision Model for Small Object Detection in UAV Scenarios}},
  booktitle = {Drones},
  year = {2026},
  doi = {10.3390/drones10010041},
}
Figure