Small object detection for UAV using bio-inspired spiking neural network
Article Text (Baidu Translation)
-
摘要: 针对无人机小目标检测方法难以在检测精度和计算复杂度之间实现有效权衡，进而难以实际部署的问题，本文提出一种基于生物启发脉冲神经网络(SNN)框架的SpikeSOD模型，用于无人机小目标检测。该模型方案基于目标检测YOLOv8模型进行改进，引入生物神经元多突触结构启发的泄露积分-多级点火(LIMF)神经元减小脉冲量化误差，缓解了SNN模型稀疏性加剧小目标信息丢失问题；使用轻量化的生物视觉神经元横向抑制机制启发的脉冲特征增强模块来增强主干网络对小目标局部信息及其周围环境的感知能力；采用脉冲特征融合模块来增强颈部网络的多尺度特征融合和互补表征能力。基于无人机航拍数据集VisDrone-DET2019进行验证，结果表明：相较于基线YOLOv8n模型，所提出的SpikeSOD模型平均精度均值提升了22.7%，参数量减少了16.7%，能耗降低了12.3 mJ，其中，对小目标检测性能的改善尤为显著；所设计的LIMF神经元相较于最具竞争力的神经元平均精度均值提升了8.8%，有效缓解了传统SNN模型在小目标信息处理中的局限性；相较于现有目标检测模型，SpikeSOD模型在检测精度、轻量化和低功耗3个关键指标间实现了有效平衡，在无人机平台的实际部署方面具有显著可行性和应用潜力。Abstract: To address the problem that small object detection methods for unmanned aerial vehicles struggle to achieve an effective trade-off between detection accuracy and computational complexity, which hinders practical deployment, a SpikeSOD model based on the bio-inspired spiking neural network (SNN) framework was proposed for small object detection of unmanned aerial vehicles. The model scheme was improved based on the YOLOv8 object detection model. A leaky integrate and multi fire (LIMF) neuron inspired by the multi-synaptic structure of biological neurons was introduced to reduce the spike quantization error and alleviate the small object information loss problem exacerbated by the sparsity of the SNN model. A lightweight spiking feature enhancement module inspired by the lateral inhibition mechanism of biological visual neurons was used to enhance the perception ability of the backbone network for the local information of small objects and their surrounding environment. A spiking feature fusion module was adopted to enhance the multi-scale feature fusion and complementary representation abilities of the neck network. The proposed model was validated on the unmanned aerial vehicle dataset VisDrone-DET2019. The results indicate that compared with the baseline YOLOv8n model, the proposed SpikeSOD model increases the mean average precision by 22.7%, reduces the parameter amount by 16.7%, and decreases the energy consumption by 12.3 mJ, among which the improvement in small object detection performance is particularly significant. The designed LIMF neuron increases the mean average precision by 8.8% compared with the most competitive neuron and effectively alleviates the limitations of traditional SNN models in small object information processing. Compared with existing object detection models, the SpikeSOD model achieves an effective balance among three key indicators, i.e., detection accuracy, lightweight design, and low power consumption, and has significant feasibility and application potential for practical deployment on unmanned aerial vehicle platforms.
-
图 1 无人机小目标检测模型SpikeSOD的总体框架
Figure 1. Overall framework of SpikeSOD model for UAV small object detection
图 2 不同脉冲神经元模型的训练和推理过程示意
Figure 2. Schematic of the training and inference processes for different spiking neuron models
表 1 VisDrone-DET2019数据集主要目标检测精度
Table 1. Detection accuracy of target objects on VisDrone-DET2019 dataset
% 检测目标 基线(YOLOv8n) SpikeSOD AP@50 AP@50:95 AP@50 AP@50:95 全部目标 26.9 14.9 49.6(↑22.7) 30.3(↑15.4) 人 12.3 3.9 58.1(↑35.0) 28.5(↑24.6) 行人 23.1 8.6 47.7(↑35.4) 20.0(↑11.4) 自行车 5.9 2.3 21.8(↑15.9) 10.2(↑7.9) 汽车 66.3 40.9 86.7(↑20.4) 63.4(↑22.5) 面包车 29.3 18.8 54.1(↑24.8) 39.4(↑20.6) 卡车 31.1 18.5 42.9(↑11.8) 28.6(↑10.1) 三轮车 10.9 5.3 37.9(↑27.0) 21.6(↑16.3) 遮阳蓬三轮车 15.3 7.9 22.2(↑6.9) 13.9(↑6.0) 公交车 50.4 33.5 66.4(↑16.0) 49.8(↑16.3) 摩托车 24.0 8.9 58.1(↑34.1) 27.8(↑18.9) 注：“(↑)”表示相比于基线模型，对应指标增加的百分比。 
下载: 导出CSV
表 2 LIMF模型与现有脉冲神经元模型的对比结果
Table 2. Comparative results of LIMF model and existing spiking neuron models
下载: 导出CSV
表 3 SpikeSOD模型与现有模型对比结果
Table 3. Comparative results of SpikeSOD model and existing models
类型 序号 模型 基线模型 分辨率/ (像素×像素) mAP @50/% 参数量/ 106 FPS 能耗/ mJ FLOP (SOP)/109 平均SFR ANN ① YOLOv8n YOLO 640×640 26.9 3.0 62.4 18.9 8.2 ② YOLOv10n YOLO 640×640 32.6 2.7 61.5 19.3 8.4 ③ YOLO11n YOLO 640×640 33.4 2.6 65.3 15.0 6.5 ④ RPLFDet[15] YOLO 640×640 53.8 1.8 54.1 23.5 ⑤ SMA-YOLO[10] YOLO 640×640 45.9 3.3 50.1 57.3 24.9 ⑥ FDB-YOLO[34] YOLO 640×640 39.9 8.5 53.0 143.5 62.4 ⑦ LSOD-YOLO[12] YOLO 640×640 37.0 3.8 93.0 78.0 33.9 ⑧ ESOD[6] ESOD 59.7 36.4 274.9 119.5 ⑨ HR-FPN[8] FPN 1 024×1 024 50.8 32.1 23.9 929.2 404.0 ⑩ PS-YOLO[35] YOLO 640×640 32.3 5.5 46.0 20.0 ⑪ DREB-Net[36] CenterNet 1 024×1 024 34.4 30.3 11.7 475.9 206.9 ⑫ CPDD-YOLOv8[37] YOLO 41.0 206.0 22.0 326.4 141.9 ⑬ AEFFNet[38] YOLO 800×800 34.4 32.9 68.0 327.3 142.3 ⑭ Deformable DETR[18] DETR 640×640 43.1 40.0 14.0 450.8 196.0 ⑮ RT-DETR[19] DETR 640×640 45.2 20.0 96.2 138.0 60.0 ⑯ APNet[20] RT-DETR 640×640 48.7 21.3 65.3 142.4 61.9 ⑰ SDFA-Net[21] RT-DETR 640×640 49.7 18.1 63.4 127.0 55.2 ⑱ Drone-DETR[22] RT-DETR 640×640 42.4 28.7 30.0 295.1 128.3 SNN ⑲ EMS-YOLO[28] YOLO 640×640 21.5 28.7 39.5 22.6 72.1 0.35 ⑳ SpikeYOLO[30] YOLO 640×640 33.9 65.5 45.8 45.9 404.0 0.13 ㉑ Spiking Trans-YOLO[32] YOLO 640×640 37.6 21.5 50.8 19.7 141.9 0.15 ㉒ SpikeSOD YOLO 640×640 49.6 2.5 48.1 6.6 50.3 0.15 SpikeSOD YOLO 800×800 51.3 2.5 34.8 12.0 78.6 0.17 SpikeSOD YOLO 1 024×1 024 55.0 2.5 24.0 20.9 128.8 0.18
下载: 导出CSV
表 4 SpikeSOD模型不同成分的消融验证结果
Table 4. Ablation experimental results of different components of SpikeSOD
编号 LIMF SFFM SFFM1 SFFM2 mAP @50/% AP@50/% 参数量/106 能耗/ mJ FPS 全部目标 人 行人 自行车 汽车 面包车 卡车 三轮车 遮阳蓬三轮车 公交车 摩托车 S0 26.7 28.4 31.8 4.1 72.7 28.1 17.1 15.3 7.0 32.0 30.8 4.1 9.1 65.4 S1 √ 42.9 41.7 50.2 13.7 83.5 48.9 35.0 29.3 17.7 58.6 50.0 4.1 6.5 68.0 S2 √ √ 45.1 44.0 52.6 16.8 84.9 49.3 38.7 31.2 17.9 62.1 53.5 1.8 3.8 59.5 S3 √ √ 44.8 43.4 52.4 16.2 84.4 50.2 37.5 33.2 17.6 59.7 53.8 4.6 8.0 56.2 S4 √ √ 46.0 45.4 54.3 20.1 84.4 46.7 40.1 32.7 19.1 61.7 55.5 4.4 9.5 77.5 S5 √ √ √ 46.8 45.6 56.2 19.0 85.8 51.8 38.0 34.3 20.9 60.6 56.3 2.2 5.7 48.8 S6 √ √ √ 47.9 46.6 56.3 19.9 85.7 52.2 41.0 33.4 20.8 66.7 56.3 2.0 4.7 33.4 S7 √ √ √ 47.2 45.6 55.9 19.2 85.6 50.8 39.9 36.6 19.4 63.2 55.3 4.9 12.9 53.7 本文模型 √ √ √ √ 49.6 47.7 58.1 21.8 86.7 54.1 42.9 37.9 22.2 66.4 58.1 2.5 6.6 48.1 注：“√”指示对应模型引入该模块。
下载: 导出CSV
-
[1] WEI W, CHENG Y, HE J F, et al. A review of small object detection based on deep learning[J]. Neural Computing and Applications, 2024, 36(12): 6283-6303. doi: 10.1007/s00521-024-09422-6 [2] 余磊, 施柏鑫, 王威, 等. 类脑赋能视觉增强: 原理、方法与前沿进展[J]. 中国图象图形学报, 2025, 30(6): 1593-1615.YU Lei, SHI Bo-xin, WANG Wei, et al. Neuromorphic-enabled visual enhancement: Principles, methods and recent advances[J]. Journal of Image and Graphics, 2025, 30(6): 1593-1615. [3] NEFTCI E O, MOSTAFA H, ZENKE F. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks[J]. IEEE Signal Processing Magazine, 2019, 36(6): 51-63. doi: 10.1109/MSP.2019.2931595 [4] CHENG G, YUAN X, YAO X W, et al. Towards large-scale small object detection: Survey and benchmarks[J]. IEEE Transactions on Pattern Analysis and Machine Intel-ligence, 2023, 45(11): 13467-13488. [5] 蒋仕新, 邹小雪, 杨建喜, 等. 复杂背景下基于改进YOLOv8s的混凝土桥梁裂缝检测方法[J]. 交通运输工程学报, 2024, 24(6): 135-147.JIANG Shi-xin, ZOU Xiao-xue, YANG Jian-xi, et al. Concrete bridge crack detection method based on improved YOLOv8s in complex backgrounds[J]. Journal of Traffic and Trans-portation Engineering, 2024, 24(6): 135-147. [6] LIU K, FU Z H, JIN S, et al. ESOD: Efficient small object detection on high-resolution images[J]. IEEE Transactions on Image Processing, 2025, 34: 183-195. doi: 10.1109/TIP.2024.3501853 [7] ZOPH B, CUBUK E D, GHIASI G, et al. Learning data augmentation strategies for object detection[C]//Springer. Computer Vision-ECCV 2020. Berlin: Springer, 2020: 566-583. . [8] CHEN Z D, JI H B, ZHANG Y Q, et al. High-resolution feature pyramid network for small object detection on drone view[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2024, 34(1): 475-489. doi: 10.1109/TCSVT.2023.3286896 [9] 陈婷, 姚大春, 高涛, 等. 基于PReNet和YOLOv4融合的雨天交通目标检测网络[J]. 交通运输工程学报, 2022, 22(3): 225-237. doi: 10.19818/j.cnki.1671-1637.2022.03.018CHEN Ting, YAO Da-chun, GAO Tao, et al. A fused network based on PReNet and YOLOv4 for traffic object detection in rainy environment[J]. Journal of Traffic and Transportation Engineering, 2022, 22(3): 225-237. doi: 10.19818/j.cnki.1671-1637.2022.03.018 [10] ZHOU S L, ZHOU H J, QIAN L. A multi-scale small object detection algorithm SMA-YOLO for UAV remote sensing images[J]. Scientific Reports, 2025, 15: 9255. doi: 10.1038/s41598-025-92344-7 [11] FAN Q S, LI Y T, DEVECI M, et al. LUD-YOLO: A novel lightweight object detection network for unmanned aerial vehicle[J]. Information Sciences, 2025, 686: 121366. doi: 10.1016/j.ins.2024.121366 [12] WANG H Z, LIU J H, ZHAO J, et al. Precision and speed: LSOD-YOLO for lightweight small object detection[J]. Expert Systems with Applications, 2025, 269: 126440. doi: 10.1016/j.eswa.2025.126440 [13] 刘一诺, 张琪, 王蓉, 等. 针对航拍小目标检测的YOLOv7改进方法[J]. 北京航空航天大学学报, 2025, 51(7): 2506-2512.LIU Yi-nuo, ZHANG Qi, WANG Rong, et al. Improved YOLOv7 method for aerial small target detection in aerial photography[J]. Journal of Beijing University of Aeronautics and Astronautics, 2025, 51(7): 2506-2512. [14] WU Y M, MU X F, SHI H, et al. An object detection model AAPW-YOLO for UAV remote sensing images based on adaptive convolution and reconstructed feature fusion[J]. Scientific Reports, 2025, 15: 16214. doi: 10.1038/s41598-025-00239-4 [15] WANG R P, LIN C, LI Y J. RPLFDet: A lightweight small object detection network for UAV aerial images with rational preservation of low-level features[J]. IEEE Transactions on Instrumentation and Measurement, 2025, 74: 5013514. [16] 王迎龙, 孙备, 丁冰, 等. BG-YOLO: 复杂大视场下低慢小无人机目标检测方法[J]. 仪器仪表学报, 2025, 46(2): 255-266.WANG Ying-long, SUN Bei, DING Bing, et al. BG-YOLO: A low-altitude slow-moving small UAV targets detection method in complex large field of view[J]. Chinese Journal of Scientific Instrument, 2025, 46(2): 255-266. [17] SHEN C C, QIAN J B, WANG C, et al. Dynamic sensing and correlation loss detector for small object detection in remote sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5627212. [18] ZHU X Z, SU W J, LU L W, et al. Deformable DETR: Deformable transformers for end-to-end object detection[J/OL]. ICLR, 2020, https://arxiv.org/abs/2010.04159 .[19] ZHAO Y A, LV W Y, XU S L, et al. DETRs beat YOLOs on real-time object detection[C]//IEEE. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2024: 16965-16974. [20] ZHANG P R, ZHANG G X, YANG K H. APNet: Accurate positioning deformable convolution for UAV image object detection[J]. IEEE Latin America Transactions, 2024, 22(4): 304-311. doi: 10.1109/TLA.2024.10472961 [21] QI Y S, CAO H W. SDFA-net: Synergistic dynamic fusion architecture with deformable attention for UAV small target detection[J]. IEEE Access, 2025, 13: 110636-110647. doi: 10.1109/ACCESS.2025.3578737 [22] KONG Y N, SHANG X F, JIA S J. Drone-DETR: Efficient small object detection for remote sensing image using enhanced RT-DETR model[J]. Sensors, 2024, 24(17): 5496. doi: 10.3390/s24175496 [23] LIANG C, GAO S B, HE L T, et al. Biological vision inspired context-awareness network for various non-generic object detection[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2025, 35(7): 6726-6739. doi: 10.1109/TCSVT.2025.3540495 [24] PANG X T, LIN C, LI F Z, et al. Bio-inspired XYW parallel pathway edge detection network[J]. Expert Systems with Applications, 2024, 237: 121649. doi: 10.1016/j.eswa.2023.121649 [25] HUANG K J, LIN C, PENG J S. Bio-inspired visual mecha-nism lightweight network for edge detection[J]. Signal, Image and Video Processing, 2025, 19(6): 450. doi: 10.1007/s11760-025-04050-6 [26] IABONI C, ABICHANDANI P. Event-based spiking neural networks for object detection: A review of datasets, architec-tures, learning rules, and implementation[J]. IEEE Access, 2024, 12: 180532-180596. doi: 10.1109/ACCESS.2024.3479968 [27] KIM S, PARK S, NA B, et al. Spiking-YOLO: Spiking neural network for energy-efficient object detection[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(7): 11270-11277. doi: 10.1609/aaai.v34i07.6787 [28] SU Q Y, CHOU Y H, HU Y F, et al. Deep directly-trained spiking neural networks for object detection[C]//IEEE. 2023 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE, 2023: 6532-6542. [29] FAN Y M, LIU C S, LI M Y, et al. SpikeDet: Better firing patterns for accurate and energy-efficient object detection with spiking neural networks[J/OL]. ICLR, 2025, https://arxiv.org/abs/2501.15151 .[30] LUO X H, YAO M, CHOU Y H, et al. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection[C]// Springer. Computer Vision-ECCV 2024. Berlin: Springer, 2024: 253-272. [31] YAO M, HU J K, HU T X, et al. Spike-driven transformer V2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips[J/OL]. ICLR, 2024, https://arxiv.org/abs/2404.03663 .[32] HUO Y S, GE H W, TANG G Z, et al. Spiking Trans-YOLO: A range-adaptive energy-efficient bridge between YOLO and Transformer[J]. Neurocomputing, 2025, 645: 130407. doi: 10.1016/j.neucom.2025.130407 [33] DU D W, ZHU P F, WEN L Y, et al. VisDrone-DET2019: The vision meets drone object detection in image challenge results[C]//IEEE. 2019 IEEE/CVF International Confe-rence on Computer Vision Workshops (ICCVW). New York: IEEE, 2019: 213-226. [34] 宋耀莲, 王粲, 李大焱, 等. 基于改进YOLOv5s的无人机小目标检测算法[J]. 浙江大学学报(工学版), 2024, 58(12): 2417-2426.SONG Yao-lian, WANG Can, LI Da-yan, et al. UAV small target detection algorithm based on improved YOLOv5s[J]. Journal of Zhejiang University (Engineering Science), 2024, 58(12): 2417-2426. [35] ZHONG H, ZHANG Y, SHI Z G, et al. PS-YOLO: A lighter and faster network for UAV object detection[J]. Remote Sensing, 2025, 17(9): 1641. doi: 10.3390/rs17091641 [36] LI Q P, ZHANG Y X, FANG L Y, et al. DREB-net: Dual-stream restoration embedding blur-feature fusion network for high-mobility UAV object detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 5621218. [37] WANG J Y, GAO J Y, ZHANG B. A small object detection model in aerial images based on CPDD-YOLOv8[J]. Scien-tific Reports, 2025, 15: 770. doi: 10.1038/s41598-024-84938-4 [38] NIAN Z Y, YANG W Z, CHEN H. AEFFNet: Attention enhanced feature fusion network for small object detection in UAV imagery[J]. IEEE Access, 2025, 13: 26494-26505. doi: 10.1109/ACCESS.2025.3538873 -
图(8) / 表(4)
计量
- 文章访问数: 14
- HTML全文浏览量: 6
- PDF下载量: 3
- 被引次数: 0
