ECSFNet-DET: An Anti-Drone Object Detection Algorithm Based on Visual State Space Modeling and Spectral Enhancement
Abstract To address the core challenges of anti-drone detection in complex low-altitude environments, including sparse target pixels, strong background clutter interference, and insufficient utilization of frequency-domain information by existing methods, this paper proposes ECSFNet-DET, a novel anti-drone object detection algorithm based on visual state space modeling and spectral enhancement. Built upon YOLOv11n, ECSFNet-DET introduces three synergistic improvements across the backbone, feature fusion neck, and deep semantic enhancement stages. First, a Dual-branch Visual State-space Feature Encoder (DVSFE) is proposed, which deeply integrates a lightweight state space sequence modeling mechanism with a depth-wise convolutional activation gated feed-forward unit within a dual-branch architecture, achieving joint local detail perception and global long-range context modeling at near-linear computational complexity. Second, an Adaptive Bidirectional Multi-Scale Semantic Fusion Neck (ABMSFN) is designed to construct a bidirectional multi-scale feature flow pathway through global context aggregation, recursive asymmetric strip channel-spatial calibration, dynamic cross-scale semantic injection, and gated adaptive fusion, effectively mitigating the semantic gap across scales. Third, a Polarized Attention Multi-Granularity Spectral Enhancement module (PAMGSE) is introduced, which embeds a multi-granularity hybrid convolutional enhancement network and a spectral adaptive modulation gated feed-forward network into a polarized self-attention backbone via learnable scalar-gated residuals, establishing a hierarchical joint spatial-frequency feature representation. Furthermore, a self-constructed anti-drone detection dataset SUAV-Bird is built, comprising 7,874 images across drone and bird categories, with over 86% of target instances below 16\((\times)\)16 pixels. Experimental results demonstrate that ECSFNet-DET achieves improvements of 4.1%, 3.5%, 3.8%, and 2.0% over the YOLOv11n baseline in precision, recall, mAP50, and mAP50-95, respectively, attaining an optimal accuracy-efficiency trade-off at 8.8 GFLOPs and 3.28M parameters, with robust cross-domain generalization validated on the multi-source Mixdata benchmark.
Authors
- 穆远超 Mu Yuanchao
- Fan Gao (ORCID: https://orcid.org/0000-0001-9207-7635)
- Shibo Liu
- Chengming Sun
- Yonghang Yu
- Yu Zhang
Institutions
- Aerospace Information Research Institute (CN)
- Zhejiang University of Technology (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-10429850/v1
- Primary Topic
- Advanced Neural Network Applications
- Type
- preprint