第4期 王晓林，等：一种基于级联神经网络的飞机检测方法 ·703· 不均衡而提出的，而在P-Net和R-Net的训练样本 [6]LIN T Y,GOYAL P,GIRSHICK R,et al.Focal loss for 生成过程中，我们已经将正负样本的比例控制在 dense object detection[C]//Proceedings of the IEEE Inter- 1:3左右，因此Focal Loss没有带来性能提升。而 national Conference on Computer Vision.Venice,Italy, 在边框回归任务中，使用L2损失得到了最佳性能。 2017:2999-3007. [7]LIN T Y,MAIRE M,MAIRE M,et al.Microsoft COCO: 表7损失函数的影响 common objects in context[C]//Proceedings of the 14th Table 7 Comparison between different loss function European Conference on Computer Vision.Zurich, 任务 方法 mAP/% Switzerland.2014:740-755. 分类 Cross Entropy 90.42 [8]EVERINGHAM M,GOOL L J V.The PASCAL visual Focal Loss 90.40 object classes challenge[J].International journal of com- 12 puter vision,2010,88(2):303-338. 90.42 [9]DALAL N,TRIGGS B.Histograms of oriented gradients 回归 LI 88.6 for human detection[C]//Proceedings of the IEEE Confer- smooth LI 90.16 ence on Computer Vision and Pattern Recognition.San Diego,USA,2005:886-893. 5结束语 [10]LOWE D G.Distinctive image features from scale-invari- ant keypoints[J].International journal of computer vision, 飞机检测任务是目前遥感图像领域的一个重 2004,60(2:91-110. 要研究方向。针对目前的检测算法依赖复杂主干 [11]CORTES C,VAPNIK V.Support-vector networks[J]. 网络的问题，本文提出了一种级联两个简单神经 Machine learning,1995,20(3):273-297 网络来解决飞机检测任务的模型，在检测精度上 [12]UIJLINGS J.SANDE K,GEVERS T,et al.Selective 接近了目前主流方法的水平，而本方法的模型参 search for object recognition[J].International journal of 数量远少于主流方法，在检测速度上也取得了一 computer vision,2013,104(2):154-171. 定的领先。此外，还制作了一个新的遥感图像数 [13]GIRSHICK R.DONAHUE J,DARRELL T,et al.Rich 据集，该数据集针对飞机检测任务标注了超过 feature hierarchies for accurate object detection and se- mantic segmentation[C]//Proceedings of the IEEE Confer- 9000个飞机实例，能够较好地验证检测算法的有 ence on Computer Vision and Pattern Recognition. 效性。 Columbus.USA.2014:580-587. 参考文献： [14]GIRSHICK R.Fast R-CNN[C]//Proceedings of the IEEE International Conference on Computer Vision.Santiago, [1]REN S,HE K.GIRSHICK R.et al.Faster R-CNN:to- Chile,.2015:1440-1448. wards real-time object detection with region proposal net- [15]AN Z,SHI Z,TENG X,et al.An automated airplane de- works [Cl//Proceedings of the 28th International Confer- tection system for large panchromatic image with high ence on Neural Information Processing System.Montreal, spatial resolution[J].Optik,2014,125(12):2768-2775. Canada2015:91-99. [16]LI W,XIANG S,WANG H,et al.Robust airplane detec- [2]REDMON J.DIVVALA S.GIRSHICK R.et al.You only tion in satellite images[C]//Proceedings of International look once:unified,real-time object detection[Cl//Proceedi Conference on Image Processing.Brussels,Belgium, ngs of the IEEE Conference on Computer Vision and Pat- 2011:2821-2824. tern Recognition.Las Vegas,USA,2015:779-788. [17]HSIEH M R,LIN Y L.HSU W H.Drone-based object [3]REDMON J,FARHADI A.YOLO9000:better,faster, counting by spatially regularized regional proposal[C/ strong-er[C]//Proceedings of the IEEE Conference on Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition.Honolulu,USA. Computer Vision.Venice,Italy,2017:4165-4173. 2017:6517-6525 [18]LIU L,PAN Z,LEI B.Learning a rotation invariant de- [4]REDMON J,FARHADI A.YOLOv3:an incremental im- tector with rotatable bounding box[J].ar Xiv preprint arX- provement[J].arXiv preprint arXiv:1804.02767,2018. iv:1711.09405,2017 [5]LIU W.ANGUELOY D,ERHAN D,et al.SSD:single [19]YANG Y,ZHUANG Y,BI F,et al.M-FCN:effective shot multibox detector[C]//Proceedings of the 14th fully convolutional network-based airplane detection European Conference on Computer Vision.Amsterdam, Frame-work[J].IEEE geoscience and remoting sensing The Netherlands,2015:21-37. letters,.2017,148):1293-1297不均衡而提出的，而在 P-Net 和 R-Net 的训练样本 生成过程中，我们已经将正负样本的比例控制在 1∶3 左右，因此 Focal Loss 没有带来性能提升。而 在边框回归任务中，使用 L2 损失得到了最佳性能。 表 7 损失函数的影响 Table 7 Comparison between different loss function 任务 方法 mAP/% 分类 Cross Entropy 90.42 Focal Loss 90.40 回归 L2 90.42 L1 88.6 smooth L1 90.16 5 结束语 飞机检测任务是目前遥感图像领域的一个重 要研究方向。针对目前的检测算法依赖复杂主干 网络的问题，本文提出了一种级联两个简单神经 网络来解决飞机检测任务的模型，在检测精度上 接近了目前主流方法的水平，而本方法的模型参 数量远少于主流方法，在检测速度上也取得了一 定的领先。此外，还制作了一个新的遥感图像数 据集，该数据集针对飞机检测任务标注了超过 9 000 个飞机实例，能够较好地验证检测算法的有 效性。 参考文献： REN S, HE K, GIRSHICK R, et al. Faster R-CNN: to￾wards real-time object detection with region proposal net￾works [C]//Proceedings of the 28th International Confer￾ence on Neural Information Processing System. Montreal, Canada, 2015: 91−99. [1] REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceedi ngs of the IEEE Conference on Computer Vision and Pat￾tern Recognition. Las Vegas, USA, 2015: 779−788. [2] REDMON J, FARHADI A. YOLO9000: better, faster, strong-er[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, USA. 2017: 6517−6525. [3] REDMON J, FARHADI A. YOLOv3: an incremental im￾provement[J]. arXiv preprint arXiv: 1804.02767, 2018. [4] LIU W, ANGUELOY D, ERHAN D, et al. SSD: single shot multibox detector[C]//Proceedings of the 14th European Conference on Computer Vision. Amsterdam, The Netherlands, 2015: 21−37. [5] LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//Proceedings of the IEEE Inter￾national Conference on Computer Vision. Venice, Italy, 2017: 2999−3007. [6] LIN T Y, MAIRE M, MAIRE M, et al. Microsoft COCO: common objects in context[C]// Proceedings of the 14th European Conference on Computer Vision. Zurich, Switzerland, 2014: 740−755. [7] EVERINGHAM M, GOOL L J V. The PASCAL visual object classes challenge[J]. International journal of com￾puter vision, 2010, 88(2): 303–338. [8] DALAL N, TRIGGS B. Histograms of oriented gradients for human detection[C]// Proceedings of the IEEE Confer￾ence on Computer Vision and Pattern Recognition. San Diego, USA, 2005: 886−893. [9] LOWE D G. Distinctive image features from scale-invari￾ant keypoints[J]. International journal of computer vision, 2004, 60(2): 91–110. [10] CORTES C, VAPNIK V. Support-vector networks[J]. Machine learning, 1995, 20(3): 273–297. [11] UIJLINGS J, SANDE K, GEVERS T, et al. Selective search for object recognition[J]. International journal of computer vision, 2013, 104(2): 154–171. [12] GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and se￾mantic segmentation[C]//Proceedings of the IEEE Confer￾ence on Computer Vision and Pattern Recognition. Columbus, USA, 2014: 580−587. [13] GIRSHICK R. Fast R-CNN[C]//Proceedings of the IEEE International Conference on Computer Vision. Santiago, Chile, 2015: 1440–1448. [14] AN Z, SHI Z, TENG X, et al. An automated airplane de￾tection system for large panchromatic image with high spatial resolution[J]. Optik, 2014, 125(12): 2768–2775. [15] LI W, XIANG S, WANG H, et al. Robust airplane detec￾tion in satellite images[C]//Proceedings of International Conference on Image Processing. Brussels, Belgium, 2011: 2821−2824. [16] HSIEH M R, LIN Y L, HSU W H. Drone-based object counting by spatially regularized regional proposal[C]// Proceedings of the IEEE International Conference on Computer Vision. Venice, Italy, 2017: 4165−4173. [17] LIU L, PAN Z, LEI B. Learning a rotation invariant de￾tector with rotatable bounding box[J]. arXiv preprint arX￾iv: 1711.09405, 2017. [18] YANG Y, ZHUANG Y, BI F, et al. M-FCN: effective fully convolutional network-based airplane detection Frame-work[J]. IEEE geoscience and remoting sensing letters, 2017, 14(8): 1293–1297. [19] 第 4 期 王晓林，等：一种基于级联神经网络的飞机检测方法 ·703·