AED-YOLO11: A small object detection model based on YOLO11

Small object detection suffers from inherent challenges including noise susceptibility, frequent occlusions, low spatial feature saliency, and imbalanced data distribution. While You Only Look Once 11 (YOLO11) maintains real-time processing capabilities, its detection efficacy on small objects is compromised by insufficient frequency-domain analysis and redundant computational operations in shallow network layers. To overcome these challenges, this study introduces Adaptive Efficient and Dynamic-YOLO11 (AED-YOLO11), a novel detection framework built upon the YOLO11 architecture with specialized enhancements for small object recognition. Specifically, the model introduces the following innovations: First, the Adaptive Frequency Domain Aggregation (AFDA) module dynamically aggregates features using frequency-domain information and channel-wise weighting, resolving frequency inconsistencies in small object images. Second, the Efficient Attention Compression (EAC) module significantly reduces computational costs by compressing channel dimensions and fusing features, thereby improving feature extraction capabilities. Third, the Dynamic Upsampling (DySample) module enhances spatial transformation capabilities through dynamic sampling of input feature maps. Finally, the Wise-IoU(WIoU) loss function is applied to improve detection performance on low-quality samples. Additionally, the detection head structure is optimized to better suit small object detection needs. Collectively, these improvements enhance the model's accuracy and computational efficiency, demonstrating superior performance in complex scenarios. Benchmark tests on VisDrone2019 indicate AED-YOLO11 yields a 4.2% mAP enhancement over baseline approaches while surpassing existing YOLO-series models in small object recognition tasks.