·1150· 智能系统学报 第14卷 表3 MS COCO数据集检测结果 Table 3 Object detection results on MS COCO test-dev set. 方法 基础网络 AP0s:0.95 APo.s AP0.75 APs APM APL AR1 AR10 AR100 ARs ARM ARL Faster[4] VGG16 21.9 42.7 ION[7] VGG16 23.6 43.2 23.66.424.138.323.232.733.210.137.753.6 R-FCN[15] Residual-101 29.2 51.5 10.332.443.3 DSOD[16] DS/64/192/4 29.3 47.3 30.6 9.431.547 27.340.743 16.747.165 Yolov2[17刀 Darknet 21.6 44.0 19.2 9.028.941.924.837.5 39.814.043.559.0 sSD300[6 VGG16 25.1 43.1 25.8 6.625.941.423.7 37.211.240.458.4 DSSD321[9例 Residual-101 28.0 46.1 2 47.6 25. 39.412.742 62.6 STDN321[18] DenseNet20 28.0 45.6 29 29.1 45. 24.4 h 38.4 12.542.760.1 Ours320 VGG16 28.2 47.7 29. 10.331.443.725.838.9 41.216.947.261.0 SSD512[6 VGG16 28.8 48.5 30.3 10.931843.526.139.5 42 16.546.660.8 DSSD513[9] Residual-101 33.2 53.3 35.2 1335.451.128.943.5 46.221.849.166.4 STDN513[18] DenseNet 31.8 51.0 33.614.436.143.427.040.141.918.348.357.3 Ours512 VGG16 33.1 52.3 32.415.634.642.728.342.645.625.950.860.1 3结束语 [5]REDMON J,DIVVALA S,GIRSHICK R,et al.You only look once:unified,real-time object detection[Cl//Proceed- 针对小目标检测准确率较低的问题，本文提 ings of 2016 IEEE Conference on Computer Vision and 出了一种基于跳跃连接金字塔的小目标检测模 Pattern Recognition(CVPR).Las Vegas,NV,USA,2016: 型。通过跳跃连接的特征金字塔融合高层与低层 779-788. 特征图信息，并且利用不同大小卷积和不同步长 [6]LIU Wei,ANGUELOV D,ERHAN D,et al.SSD:single 空洞卷积的横向结构来提取全局特征信息，有效 shot MultiBox detector[C]//Proceedings of the 14th 弥补因连续池化而造成的信息丢失。整个网络模 European Conference on Computer Vision.Amsterdam, 型以端到端方式进行训练，并且在PASCAL VOC the Netherlands,2016:21-37. 和MS COCO数据集上进行了实验，实验结果表 [7]BELL S,ZITNICK CL,BALA K,et al.Inside-outside 明本文提出的模型在小目标的检测准确率方面明 net:detecting objects in context with skip pooling and re- 显优于其他算法模型。 current neural networks[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition 参考文献： (CVPR).Las Vegas,NV,USA,2016:2874-2883. [8]LIN T Y,DOLLAR P,GIRSHICK R,et al.Feature pyram- [1]GIRSHICK R,DONAHUE J,DARRELL T,et al.Rich id networks for object detection[Cl//Proceedings of 2017 feature hierarchies for accurate object detection and se- mantic segmentation[C]//Proceedings of 2014 IEEE Con- IEEE Conference on Computer Vision and Pattern Recog- ference on Computer Vision and Pattern Recognition(CV- nition (CVPR).Honolulu,HI,USA,2017:936-944. PR).Columbus,OH,USA,2014:580-587. [9]FU Chengyang,LIU Wei,RANGA A,et al.DSSD:decon- [2]GIRSHICK R.Fast R-CNN[Cl/Proceedings of 2015 IEEE volutional single shot detector[J].arXiv:1701.06659, International Conference on Computer Vision (ICCV). 2017 Santiago,Chile,2015:1440-1448. [10]HE Kaiming,ZHANG Xiangyu,REN Shaoqing,et al. [3]UIJLINGS J RR,VAN DE SANDE K E A,GEVERS T, Deep residual learning for image recognition[C]//Proceed- et al.Selective search for object recognition[].Internation- ings of 2016 IEEE Conference on Computer Vision and al journal of computer vision,2013,104(2):154-171. Pattern Recognition(CVPR).Las Vegas,NV,USA,2016: [4]REN Shaoqing,HE Kaiming,GIRSHICK R,et al.Faster 770-778. R-CNN:towards real-time object detection with region [11]YU F,KOLTUN V.Multi-scale context aggregation by proposal networks[J].IEEE transactions on pattern analys- dilated convolutions[J].arXiv:1511.07122,2015. is and machine intelligence,2017,39(6):1137-1149. [12]SIMONYAN K.ZISSERMAN A.Very Deep Convolu-3 结束语 针对小目标检测准确率较低的问题，本文提 出了一种基于跳跃连接金字塔的小目标检测模 型。通过跳跃连接的特征金字塔融合高层与低层 特征图信息，并且利用不同大小卷积和不同步长 空洞卷积的横向结构来提取全局特征信息，有效 弥补因连续池化而造成的信息丢失。整个网络模 型以端到端方式进行训练，并且在 PASCAL VOC 和 MS COCO 数据集上进行了实验，实验结果表 明本文提出的模型在小目标的检测准确率方面明 显优于其他算法模型。 参考文献： GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and se￾mantic segmentation[C]//Proceedings of 2014 IEEE Con￾ference on Computer Vision and Pattern Recognition (CV￾PR). Columbus, OH, USA, 2014: 580–587. [1] GIRSHICK R. Fast R-CNN[C]//Proceedings of 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, Chile, 2015: 1440–1448. [2] UIJLINGS J R R, VAN DE SANDE K E A, GEVERS T, et al. Selective search for object recognition[J]. Internation￾al journal of computer vision, 2013, 104(2): 154–171. [3] REN Shaoqing, HE Kaiming, GIRSHICK R, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE transactions on pattern analys￾is and machine intelligence, 2017, 39(6): 1137–1149. [4] REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceed￾ings of 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA, 2016: 779–788. [5] LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: single shot MultiBox detector[C]//Proceedings of the 14th European Conference on Computer Vision. Amsterdam, the Netherlands, 2016: 21–37. [6] BELL S, ZITNICK C L, BALA K, et al. Inside-outside net: detecting objects in context with skip pooling and re￾current neural networks[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA, 2016: 2874–2883. [7] LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyram￾id networks for object detection[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recog￾nition (CVPR). Honolulu, HI, USA, 2017: 936–944. [8] FU Chengyang, LIU Wei, RANGA A, et al. DSSD: decon￾volutional single shot detector[J]. arXiv: 1701.06659, 2017. [9] HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]//Proceed￾ings of 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA, 2016: 770–778. [10] YU F, KOLTUN V. Multi-scale context aggregation by dilated convolutions[J]. arXiv:1511.07122, 2015. [11] [12] SIMONYAN K, ZISSERMAN A. Very Deep Convolu- 表 3 MS COCO 数据集检测结果 Table 3 Object detection results on MS COCO test-dev set. 方法 基础网络 AP0.5∶0.95 AP0.5 AP0.75 APS APM APL AR1 AR10 AR100 ARS ARM ARL Faster[4] VGG16 21.9 42.7 − − − − − − − − − − ION[7] VGG16 23.6 43.2 23.6 6.4 24.1 38.3 23.2 32.7 33.2 10.1 37.7 53.6 R-FCN[15] Residual-101 29.2 51.5 − 10.3 32.4 43.3 − − − − − − DSOD[16] DS/64/192/4 29.3 47.3 30.6 9.4 31.5 47 27.3 40.7 43 16.7 47.1 65 Yolov2[17] Darknet 21.6 44.0 19.2 9.0 28.9 41.9 24.8 37.5 39.8 14.0 43.5 59.0 SSD300[6] VGG16 25.1 43.1 25.8 6.6 25.9 41.4 23.7 35.1 37.2 11.2 40.4 58.4 DSSD321[9] Residual-101 28.0 46.1 29.2 7.4 28.1 47.6 25.5 37.1 39.4 12.7 42 62.6 STDN321[18] DenseNet[20] 28.0 45.6 29.4 7.9 29.7 45.1 24.4 36.1 38.4 12.5 42.7 60.1 Ours320 VGG16 28.2 47.7 29.1 10.3 31.4 43.7 25.8 38.9 41.2 16.9 47.2 61.0 SSD512[6] VGG16 28.8 48.5 30.3 10.9 31.8 43.5 26.1 39.5 42 16.5 46.6 60.8 DSSD513[9] Residual-101 33.2 53.3 35.2 13 35.4 51.1 28.9 43.5 46.2 21.8 49.1 66.4 STDN513[18] DenseNet 31.8 51.0 33.6 14.4 36.1 43.4 27.0 40.1 41.9 18.3 48.3 57.3 Ours512 VGG16 33.1 52.3 32.4 15.6 34.6 42.7 28.3 42.6 45.6 25.9 50.8 60.1 ·1150· 智 能 系 统 学 报 第 14 卷