To solve the problem of existing deep learning-based object detection algorithms being unsuitable for onboard deployment due to their complexity, a multi-scale feature enhanced lightweight detection algorithm (MFLDet) for ship recognition tasks in optical remote sensing images is proposed. Firstly, to reduce the algorithm's parameter and computation load, a lightweight network architecture, PG-HGNet, is constructed as the backbone network, and constructs a lightweight cross-scale feature fusion network for interactive feature integration. Secondly, a multi-scale feature enhancement module (MFEM) is designed to accommodate the scale variability of ship targets in remote sensing images, thereby enhancing detection accuracy. Finally, the MPDIoU bounding box loss function is introduced to adjust for cases where the predicted and actual bounding boxes share the same aspect ratio but differ in absolute dimensions. Comparative experiments conducted on the HRSC2016 dataset demonstrate that, compared to the baseline model, MFLDet achieves a 55% reduction in parameters and a 23.5% reduction in computation, with only a 0.2 percentage points decrease in average precision, thus effectively balancing complexity and accuracy. Overall, the proposed method surpasses other comparative methods in terms of both lightweight level and detection precision.
虽然CIoU引入了边框的长宽值,但是以长宽比的形式使用,而不是绝对值。因此,在遥感图像舰船检测场景中,由于舰船目标的尺度变化比较大,容易出现预测框与真实框长宽比一致但长宽绝对值不相等的情况。此时,=0失去惩罚作用,损失函数难以被优化,使得在非极大值抑制过程中容易造成漏检现象,降低了边界框回归的准确性。为解决此问题,本文引入基于最小点距离的交并比(Minimum point distance based IoU, MPDIoU)[15]损失函数替换CIoU计算边框损失,其示意图如图7所示。
DongChao, LiuJing-hong, XuFang, et al. Fast ship detection in optical remote sensing images[J].Journal of Jilin University (Engineering and Technology Edition), 2019, 49(4): 1369-1376.
[3]
LiuW, AnguelovD, ErhanD, et al. SSD: Single shot multibox detector[C]∥14th European Conference on Computer Vision-ECCV 2016. Cham, Switzerland: Springer Internation Publishing, 2016: 21-37.
[4]
RenS Q, HeK, GirshickR, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[5]
CaiZ W, VasconcelosN.Cascade R-CNN: delving into high quality object detection[C]∥2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2018: 6154-6162.
LiYong-gang, ZhuWei-gang, HuangQiong-nan, et al. Near shore ship target detection with SAR images in complex background[J].Systems Engineering and Electronics, 2022, 44(10): 3096-3103.
ChengQian, LiJia, DuJuan.Ship target detection algorithm of optical remote sensing image based on YOLOv5[J]. Systems Engineering and Electronics, 2023, 45(5): 1270-1276.
[10]
YiH, LiuB, ZhaoB, et al. Small object detection algorithm based on improved YOLOv8 for remote sensing[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 17: 1734-1747.
HanK, WangY H, TianQ, et al. GhostNet: More features from cheap operations[C]∥Proceedings of the 2020 IEEE/CVF Conferencef on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2020: 1580-1589.
[14]
LeH, BorjiA. What are the receptive, effective receptive, and projective fields of neurons in convolutional neural networks?[J/OL].[2024-09-11].
[15]
LuoW, LiY, UrtasunR, et al. Understanding the effective receptive field in deep convolutional neural networks[J/OL]. [2024-09-11].
[16]
DingX H, ZhangX Y, HanJ G, et al. Scaling up your kernels to 31×31: revisiting large kernel design in CNNs[C]∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway, NJ: IEEE, 2022: 11953-11965.
[17]
MaN N, ZhangX Y, ZhengH T, et al. ShuffleNet V2: practical guidelines for efficient CNN architecture design[C]∥Proceedings of the European Conference on Computer Vision. Cham, Switzerland: Springer Interaction Publishing, 2018: 116-131.
[18]
XieX R, SunF X, LinJ, et al. Fast-ABC: a fast architecture for bottleneck-like based convolutional neural networks[C]∥2019 IEEE Computer Society Annual Symposium on VLSI. Piscataway, NJ: IEEE, 2019: 1-6.
[19]
MaS L, XuY. MPDIoU: a loss for effcient and accurate bounding box regression[J/OL]. [2024-09-11].
[20]
LiuZ K, YuanL, WengL B, et al. A high resolution optical satellite image dataset for ship recognition and some new baselines[C]∥International Conference on Pattern Recognition Applications and Methods. New York: INSTICC, 2017, 2: 324-331.
[21]
PangY H, ZhangY M, KongQ L, et al. SOCDet: a lightweight and accurate oriented object detection network for satellite on-orbit computing[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 1-15.
[22]
ChengG, WangJ B, LiK, et al. Anchor-free oriented proposal generator for object detection[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-11.
[23]
LiH, ZhangS, LiX, et al. DetectorNet: Transformer enhanced spatial temporal graph neural network for traffic prediction[C]∥Proceedings of the 29th International Conference on Advances in Geographic Information Systems. New York: ACM, 2021: 133-136.
[24]
ZhouL, WeiH R, LiH, et al. Arbitrary-oriented object detection in remote sensing images based on polar coordinates[J]. IEEE Access, 2020, 8: 223373-223384.
[25]
YangJ R, WuB, ZhangH Y, et al. Rotating frame labeling and improved YOLOv7-based algorithm for vehicle object detection[C]∥Proceedings of the 2023 IEEE 6th International Conference on Pattern Recognition and ArtificialIntelligence. Piscataway, NJ: IEEE, 2023: 170-175.