Objects in aerial images, such as vehicles, ships, buildings, etc., may appear in any direction with significant differences in their scale. Traditional non oriented bounding boxes are difficult to accurately describe the direction and shape of objects, and the position, direction, and shape of objects in the image may deviate from the actual situation, resulting in high regression loss uncertainty and low detection accuracy and efficiency in small object detection. Therefore, Improved RetinaNet is applied to achieve precise detection of oriented objects in aerial images. Designed a feature interaction fusion network to obtain rich contextual information of oriented objects. A dynamic allocation method for training samples has been introduced, which dynamically allocates training samples based on the spatial position relationship of the object to coordinate the imbalance between training samples and predict object categories through classification sub-network. Based on this, a regression sub-network is used to regress the offset between each oriented bounding box and the adjacent ground truth bounding box, and the long edge definition method is used to represent the oriented bounding box. The KL divergence between the oriented bounding box and the ground truth bounding box is calculated as the regression loss to solve the angle sensitivity problem of large aspect-ratio oriented objects. Extensive experiments on publicly available aerial image datasets have demonstrated the effectiveness and superiority of Improved RetinaNet in detecting oriented objects.
目前,主流的定向目标检测器通常利用角度定义法(中心点坐标、宽度、高度、旋转角度)来表示任意方向的目标。虽然这些检测器已经取得了良好的性能,但是当它们在航空图像中检测定向目标时,仍然存在着以下不足:①特征旋转增强的方法[5⁃9]精心设计了各种特征增强模块,并在每个位置上设置了不同角度、比例和高宽比的锚框,使得网络变得沉重且复杂,导致了大量的计算开销和内存占用;②样本动态分配的方法[10⁃12]设置了一些超参数,如基于锚框的方法的IoU(Intersection over Union,交并比)阈值、无锚框的方法的区域范围等,超参数的不同设置将会产生非常不同的检测结果;③回归损失优化的方法[13⁃16]不能根据定向目标的高度和宽度动态调整梯度,难以应对具有挑战性的情况,例如具有大高宽比的定向目标对角度的偏移很敏感,轻微的角度误差就会导致检测精度严重下降。
为验证Improved RetinaNet的优越性,进行了一系列消融实验。选择Rotated⁃RetinaNet作为消融实验的基线,采用平均精度均值(Mean average precision,mAP)来评估目标的检测精度,所有的实验条件和参数设置均严格一致。在DOTA⁃v1.0数据集上研究了特征交互融合网络(FIFN)、训练样本动态分配(TSDA)方法和KL散度损失对实验结果的影响,如表1所示。FIFN表示是否使用特征交互融合网络,TSDA表示是否使用训练样本动态分配方法,KL表示是否使用KL散度损失。实验结果表明,每种结构对检测性能都有不同程度的提高。
HanJia-ming, DingJian, XueNan, et al. ReDet: a rotation-equivariant detector for aerial object detection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Virtual, 2021: 2786-2795.
[2]
LiWen-tong, ChenYi-jie, HuKai-xuan, et al. Oriented reppoints for aerial object detection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, America, 2022: 1829-1838.
[3]
LuDong-chen. OSKDet: towards orientation-sensitive keypoint localization for rotated object detection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, America, 2022: 1182-1192.
[4]
DolorielC, CajoteR. Improving the detection of small oriented objects in aerial images[C]∥IEEE Workshop on Applications of Computer Vision, Hawaii, America, 2023: 176-185.
[5]
YangXue, YanJun-chi, FengZi-ming, et al. R3Det: refined single-stage detector with feature refinement for rotating object[C]∥AAAI Conference on Artificial Intelligence, Virtual, 2021: 3163-3171.
[6]
HanJia-ming, DingJian, LiJie, et al. Align deep features for oriented object detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: No. 5602511.
[7]
DingJian, XueNan, LongYang, et al. Learning roi transformer for oriented object detection in aerial images[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, America, 2019: 2849-2858.
[8]
XieXing-xing, ChengGong, WangJia-bao, et al. Oriented R-CNN for object detection[C]∥IEEE International Conference on Computer Vision, Virtual, 2021: 3520-3529.
[9]
WangWen-yong, CaiYuan-zheng, LuoZhi-ming, et al. SA3Det: detecting rotated objects via pixel-level attention and adaptive labels assignment[J]. Remote Sensing, 2024, 16: No.2496.
[10]
Lin, GoyalP, GirshickR, et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(2): 318-327.
[11]
HouLi-ping, LuKe, XueJian, et al. Shape-adaptive selection and measurement for oriented object detection[C]∥AAAI Conference on Artificial Intelligence, Vancouver, Canada, 2022: 923-932.
YangXue, YanJun-chi. Arbitrary-oriented object detection with circular smooth label[C]∥European Conference on Computer Vision, Virtual, 2020: 677-694.
[14]
XuYong-chao, FuMing-tao, WangQi-meng, et al. Gliding vertex on the horizontal bounding box for multi-oriented object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(4): 1452-1459.
[15]
YangXue, YanJun-chi, MingQi, et al. Rethinking rotated object detection with gaussian wasserstein distance loss[C]∥International Conference on Machine Learning, Chejudo, Korea, 2021: 11830-11841.
[16]
ZhuHu-fei, HuangYong-hui, XuYing, et al. Unmanned aerial vehicle(UAV) object detection algorithm based on keypoints representation and rotated distance-IoU loss[J]. Journal of Real-Time Image Processing, 2024, 21: No.58.
[17]
XieS, GirshickR, DollarP, et al. Aggregated residual transformations for deep neural networks[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Hawaii, America, 2017: 5431-5440.
[18]
ZhangShi-feng, ChiCheng, YaoYong-qiang, et al. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Seattle, America, 2020: 9756-9765.
[19]
KullbackS, LeiblerR. On information and sufficiency[J]. The Annals of Mathematical Statistics, 1951, 22(1): 79-86.
[20]
XiaGui-song, BaiXiang, DingJian, et al. Dota: a large-scale dataset for object detection in aerial images[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, America, 2018: 3974-3983.
[21]
SunYi-ming, CaoBing, ZhuPeng-fei, et al. Drone-based RGB-infrared cross-modality vehicle detection via uncertainty-aware learning[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 32(10): 6700-6713.
[22]
ZhouYue, YangXue, ZhangGe-fan, et al. Mmrotate: a rotated object detection benchmark using pytorch[C]∥ACM International Conference on Multimedia, Lisbon, Portugal, 2022: 7331-7334.
[23]
LinTsung-yi, DollarP, GirshickR, et al. Feature pyramid networks for object detection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Hawaii, America, 2017: 2117-2125.
[24]
LiuShu, QiLu, QinHai-fang, et al. Path aggregation network for instance segmentation[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, America, 2018: 1534-1544.
[25]
GhiasiG, LinTsung-yi, PangRuo-ming, et al. NAS-FPN: learning scalable feature pyramid architecture for object detection[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, America, 2019: 7392-7401.
[26]
ZhangJiang-fan, ZhangYan, ShiZhi-guang, et al. Unmanned aerial vehicle object detection based on information-preserving and fine-grained feature aggregation[J]. Remote Sensing, 2024, 16(14): No. 2590.