To address the challenges of high model complexity, insufficient detection performance, and weak generalization ability in aircraft target detection in synthetic aperture radar (SAR) images, this paper proposed a lightweight improvement to the YOLOv8 model from both network design and model compression perspectives. In terms of network design, the original C2f module was replaced with a CSPRELAN-C2f module, which generated redundant feature maps via 1×1 convolutions to optimize the feature extraction process. Meanwhile, the traditional BottleNeck residual block was discarded, and a RepConv module was introduced into the gradient flow branch to enhance feature extraction capability and gradient flow. Additionally, given the small size of aircraft targets in SAR images, the large-target detection layer was removed to strengthen the identification of small targets such as aircraft. Finally, the Inner-SIoU loss function was adopted to further improve model training performance. In terms of model compression, the CWD method was introduced for knowledge distillation to boost the detection performance of the lightweight model. Experimental results on the SAR-AIRcraft-1.0 dataset show that the improved algorithm reduces model size, parameter count, and computational load by 77.63%, 80.41%, and 69.14%, respectively.
YOLOv8的网络结构由主干网络、特征提取网络和检测头三大模块组成(图1)。主干网络采用了CBS(卷积、批归一化、激活函数)模块作为基础,结合C2f(CSP Bottleneck with 2 Convolutions)模块,后者受YOLOv7中高效层聚合网络(ELAN)结构的启发,通过精简卷积层来减少计算复杂度,并增强梯度流。特征提取网络则结合了传统的特征金字塔网络(FPN)和路径聚合网络(PAN),通过自底向上的高分辨率特征融合和自顶向下的语义信息传递,增强了不同尺度目标的检测能力,生成了高质量的特征图。最后,检测头部分将YOLOv5中的Anchor-based耦合头结构替换为Anchor-free解耦头结构,将分类与回归任务分开处理,从而减少了模型的复杂度并提升了推理效率。
本研究将缩放因子设置为0.5,是基于“控制变量实验”筛选出来的一个最优平衡点。本文测试了不同缩放因子(0.25、0.5、0.75)对模型的参数量计算量和平均精度值MAP(mean average precision)的影响,实验结果见表1。缩放因子大于0.5(如0.75)时,冗余分支保留了过多的通道,这不仅增加了计算量,还可能引入无效的背景特征,干扰模型对飞机关键特征的聚焦,尤其在小目标检测中易导致误检;缩放因子小于0.5(如0.25)时,通道过度压缩导致飞机稀疏的关键特征丢失,导致MAP值低于94.5%。因此选择缩放因子为0.5,是在“补偿性能损失”和“保持轻量化”之间找到的最优平衡点。
DUDGEOND E, LACOSSR T. An overview of automatic target recognition [J]. 1993, 6(1): 3-10.
[2]
DENGY K, YUW D, ZHANGH, et al. Forthcoming spaceborne SAR development[J]. Journal of Radars, 2020, 9(1): 1-33.
[3]
GUOQ, WANGH P, XUF. Research progress on aircraft detection and recognition in SAR imagery[J]. 2020, 9(3): 497-513.
[4]
GIRSHICKR, DONAHUEJ, DARRELLT, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Columbus: IEEE, 2014: 580-587.
[5]
GIRSHICKR. Fast R-CNN[C] //Proceedings of the IEEE International Conference on Computer Vision. Los Alamitos: IEEE Computer Society Press, 2015: 1440-1448.
[6]
RENS, 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.
[7]
TERVENJ, CORDOVA-ESPARZAD M, ROMERO-GONZALEZJ A. A comprehensive review of YOLO architectures in computer vision: From YOLOv1 to YOLOv8 and YOLO-NAS[J]. Machine Learning and Knowledge Extraction, 2023, 5(4): 1680-1716.
[8]
WEIL, DRAGOMIRA, DUMITRUE, et al. SSD: Single shot multibox detector[J]. arXiv preprint arXiv, 2016: 1512. 02325.
[9]
CARIONN, MASSAF, SYNNEVEG, et al. End-to-end object detection with transformers[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2020: 1529-1539.
WANGC Y, LIAOH Y, WUY H, et al. CSPNet: A new backbone that can enhance learning capability of CNN[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Seattle: IEEE, 2020: 390-398.
[14]
HANK, WANGY H, TIANQ, et al. GhostNet: More features from cheap operations[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2020: 1577-1586.
[15]
GEVORGYANZ. Siou loss: More powerful learning for bounding box regression[J].arXiv preprint arXiv, 2022: 2205.12740.
[16]
ZHENGZ H, WANGP, LIUW, et al. Distance-IoU loss: Faster and better learning for bounding box regression[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34(7): 12993-13000.
[17]
ZHANGH, XUC, ZHANGS. Inner-IoU: More effective intersection over union loss with auxiliary bounding box[J]. arXiv preprint arXiv, 2023: 2311. 02877.
[18]
SHUC, LIUY, GAOJ, et al. Channel-wise knowledge distillation for dense prediction[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 5291-5300.
ZHANGP, XUH, TIANT, et al. SEFEPNet: Scale expansion and feature enhancement pyramid network for SAR aircraft detection with small sample dataset[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2022,15: 3365-3375.
[21]
WANGC Y, BOCHKOVSKIYA, LIAOH Y M. YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]// Proceedings of IEEE/ CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 7464-7475.