1.School of Surveying and Geoinformation Engineering,East China University of Technology,Nanchang 330013,China
2.School of Electrical Engineering,Jiangxi University of Water Resources and Electric Power,Nanchang 330099,China
3.Key Laboratory of Mine Environmental Monitoring and Improving around Poyang Lake of Ministry of Natural Resources,East China University of Technology,Nanchang 330013,China
4.Jiangxi Province Engineering Research Center of Surveying,Mapping and Geographic Information,Nanchang 330025,China
To address the issue of degraded feature extraction capability and unstable regression accuracy in lightweight remote sensing rotational object detection due to computational constraints, this paper proposes a lightweight rotational object detection algorithm based on scale-invariant feature fusion. In the backbone network, an adaptive downsampling (Adown) module replaces the conventional downsampling strategy, effectively preserving more critical information with lower computational cost, reducing feature loss while improving detection accuracy. In the neck network, a high-efficiency SDI-BiFPN multi-scale feature fusion module is designed, integrating the scale-invariant feature fusion (SDI) mechanism and bidirectional feature propagation (BiFPN) strategy, which enhances multi-scale feature representation while reducing computational complexity. Furthermore, to improve the stability of rotated bounding box regression, the model incorporates the CSL (Circular Smooth Label) angle encoding strategy and adopts LPIoU (Logarithmic Polar IoU) as the regression loss function, effectively mitigating the periodicity issue in angle regression, thus improving regression accuracy and detection robustness. Experimental results on the public DOTA remote sensing dataset demonstrate that the proposed method achieves a 43.34% reduction in model parameters while maintaining detection accuracy, and improves mAP by 4.5% over the baseline model, reaching 73.19%. Additionally, in comparison with state-of-the-art methods such as R³Det, DAL, and CenterMap, the proposed approach maintains competitive detection performance while achieving a lightweight design.
传统交并比(Intersection over Union,IoU)损失在旋转目标检测任务中存在局限性,主要表现为:①角度不敏感,即IoU仅关注目标框的区域重叠,无法正确反映角度偏差;②尺度不平衡,当目标框尺寸差距较大时,IoU对小目标的梯度贡献较小。为此,本文采用极坐标对数IoU(Logarithmic Polar IoU,LPIoU)损失函数,在极坐标空间下计算IoU,以增强旋转框的匹配度。预测框和真实框分别表示为Bp=(xp,yp,wp,hp,θp)和Bt=(xt,yt,wt,ht,θt),LPIoU计算过程如下:
本实验采用DOTA公开遥感图像目标检测数据集进行模型性能评估。其中,DOTA(Dataset for object detection in aerial images)是一个高分辨率遥感图像数据集,包含15类目标,如飞机、船舶、车辆等,其标注方式为旋转边界框(OBB)。DOTA数据集包含多个版本,本实验选用DOTA-v1.0,其中训练集、验证集和测试集分别包含1 400、458和937张图像,分辨率从800×800到4 000×4 000不等。
2.2 评价标准
为了全面评估模型的检测性能,本实验采用以下指标进行分析:
P(Precision,精确率):衡量预测的目标框中正确检测目标的比例,定义为:
式中:TP为正确检测的目标数;FP为错误检测的目标数。
R(Recall,召回率):衡量所有真实目标中被正确检测的比例,定义为:
式中:FN为未被检测出的目标数。
mAP(Mean Average Precision,均值平均精度):表示所有类别的平均精度,计算公式如下:
式中:N为目标类别数;APi 为第i类的平均精度。
GFLOPS(Giga Floating-point Operations Per Second):衡量模型的计算复杂度,即每秒执行的浮点运算次数,GFLOPS越小,模型计算成本越低。Params(参数量):表示模型的总参数数量,通常用于衡量模型的复杂度和存储需求。
LiuChun-juan, XinYu-qiang, WuXiao-suo, et al. Remote sensing image change detection network based on multi-scale dense fusion[J]. Journal of Sichuan University (Natural Science Edition), 2024, 61(6): 115-125.
[3]
WangJ, YangW, LiH C, et al. Learning center probability map for detecting objects in aerial images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 59(5): 4307-4323.
[4]
NieH, PangH, MaM, et al. A lightweight remote sensing small target image detection algorithm based on improved YOLOv8[J]. Sensors, 2024, 24(9): s24092952.
LiKun-ya, OuOu, LiuGuang-bin, et al. Target detection algorithm of remote sensing image based on improved YOLOv5[J]. Computer Engineering and Applications, 2023, 59(9): 207-214.
[15]
YangX, YanJ, FengZ, et al. R3det: refined single-stage detector with feature refinement for rotating object[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(4): 3163-3171.
JiaKe-bin, HeYan, WeiZhi-hao. Fusing multi-scale features for segmentation of high resolution forest remote sensing images[J]. Journal of Beijing University of Technology, 2024, 50(9): 1089-1099.
LiZi-yi, QuTing-ting, ChongQian-peng, et al. Remote sensing image segmentation network based on fuzzy multiscale features[J]. Journal of Computer Applications, 2024, 44(11): 3581-3586.
[22]
YangX, YanJ. On the arbitrary-oriented object detection: classification based approaches revisited[J]. International Journal of Computer Vision, 2022, 130(5): 1340-1365.