第5期 王照国，等：基于F1值的非极大值抑制阈值自动选取方法 ·1011· image processing,2015,24(12):5706-5722 [5]REDMON J.DIVVALA S.GIRSHICK R,et al.You only (a)NMS look once:unified,real-time object detection[C]//Proceed- ings of 2016 IEEE Conference on Computer Vision and Pattern Recognition.Las Vegas,USA,2016:779-788. [6]LIU Wei,ANGUELOV D,ERHAN D,et al.SSD:single (b)Soft-NMS shot MultiBox detector[C]//Proceedings of the 14th European Conference on Computer Vision.Amsterdam, the Netherlands.2016:21-37. (c)本文算法 [7]REDMON J,FARHADI A.YOLOv3:an incremental im- 图8本文提出算法与其他算法实验效果对比(COC0数 provement[J].arXiv preprint arXiv:1804.02767,2018. 据集) [8]GIRSHICK R,DONAHUE J,DARRELL T,et al.Rich Fig.8 Comparison of experimental results between the proposed algorithm and other algorithms (COCO feature hierarchies for accurate object detection and se- dataset) mantic segmentation[C]//Proceedings of 2014 IEEE Con- 4结束语 ference on Computer Vision and Pattern Recognition. Columbus,USA,2014:580-587 传统的非极大值抑制算法需要大量实验调整 [9]GIRSHICK R.Fast R-CNN[C//Proceedings of 2015 IEEE 阈值，阈值设定的不合理会导致漏检和误检，阈 International Conference on Computer Vision.Santiago, 值调整需要耗费大量时间。基于此，本文提出了 USA,2015:1440-1448. 基于F1值的非极大值抑制阈值自动选取方法，减 [10]REN Shaoqing,HE Kaiming,GIRSHICK R,et al.Faster 弱了检测算法对于人为设定阈值的依赖，减少了 R-CNN:Towards real-time object detection with region 阈值调整耗费的时间。利用YOLOv.3模型在Pas cal VOC2007数据集上面进行训练，通过与传统 proposal networks[J].IEEE transactions on pattern ana- lysis and machine intelligence,2017,39(6):1137-1149. NMS、Soft-NMS以及Softer-NMS算法进行性能 对比，证明了本算法在检测准确率方面的提高。 [11]ROSENFELD A,THURSTON M.Edge and curve detec- 本文可以采用更多的映射模型来获取最佳阈值， tion for visual scene analysis[J].IEEE transactions on 期望未来能够使用卷积神经网络获取更加精准的 computers,1971,C-20(5):562-569. 映射关系模型。 [12]BODLA N,SINGH B,CHELLAPPA R,et al.Soft- 参考文献： NMS-improving object detection with one line of code[Cl//Proceedings of 2017 IEEE International Confer- [1]FELZENSZWALB P F,GIRSHICK R B,MCALLESTER ence on Computer Vision.Venice,Italy,2017: D,et al.Object detection with discriminatively trained 5561-5569 part-based models[J].IEEE transactions on pattern analys- [13]HE Yihui,ZHU Chenchen,WANG Jianren,et al.Bound- is and machine intelligence,2010,32(9):1627-1645 ing box regression with uncertainty for accurate object de- [2]HAN Junwei,ZHANG Dingwen,CHENG Gong,et al.Ad- tection[C]//Proceedings of 2019 IEEE/CVF Conference vanced deep-learning techniques for salient and category- on Computer Vision and Pattern Recognition.Long specific object detection:a survey[J].IEEE signal pro- Beach,USA,2019:2888-2897. cessing magazine,2018,35(1):84-100. [14]侯志强，刘晓义，余旺盛，等.基于双阈值-非极大值抑制 [3]XIE Cihang,WANG Jianyu,ZHANG Zhishuai,et al.Ad- 的Faster R-CNN改进算法[.光电工程，2019,46(12)」 versarial examples for semantic segmentation and object 190159. detection[C]//Proceedings of 2017 IEEE International Con- HOU Zhiqiang,LIU Xiaoyi,YU Wangsheng,et al.Im- ference on Computer Vision.Venice,Italy,2017: proved algorithm of Faster R-CNN based on double 1369-1378. threshold-non-maximum suppression[J].Opto-electronic [4]BORJI A,CHENG Mingming,JIANG Huaizu,et al.Sali- engineering,2019,46(12):190159 ent object detection:a benchmark[J].IEEE transactions on [15]赵文清，严海，邵绪强.改进的非极大值抑制算法的目(a) NMS (b) Soft-NMS (c) 本文算法 图 8 本文提出算法与其他算法实验效果对比 (COCO 数 据集) Fig. 8 Comparison of experimental results between the proposed algorithm and other algorithms (COCO dataset) 4 结束语 传统的非极大值抑制算法需要大量实验调整 阈值，阈值设定的不合理会导致漏检和误检，阈 值调整需要耗费大量时间。基于此，本文提出了 基于 F1 值的非极大值抑制阈值自动选取方法，减 弱了检测算法对于人为设定阈值的依赖，减少了 阈值调整耗费的时间。利用 YOLOv3 模型在 Pas￾cal VOC 2007 数据集上面进行训练，通过与传统 NMS、Soft-NMS 以及 Softer-NMS 算法进行性能 对比，证明了本算法在检测准确率方面的提高。 本文可以采用更多的映射模型来获取最佳阈值， 期望未来能够使用卷积神经网络获取更加精准的 映射关系模型。 参考文献： FELZENSZWALB P F, GIRSHICK R B, MCALLESTER D, et al. Object detection with discriminatively trained part-based models[J]. IEEE transactions on pattern analys￾is and machine intelligence, 2010, 32(9): 1627–1645. [1] HAN Junwei, ZHANG Dingwen, CHENG Gong, et al. Ad￾vanced deep-learning techniques for salient and category￾specific object detection: a survey[J]. IEEE signal pro￾cessing magazine, 2018, 35(1): 84–100. [2] XIE Cihang, WANG Jianyu, ZHANG Zhishuai, et al. Ad￾versarial examples for semantic segmentation and object detection[C]//Proceedings of 2017 IEEE International Con￾ference on Computer Vision. Venice, Italy, 2017: 1369−1378. [3] BORJI A, CHENG Mingming, JIANG Huaizu, et al. Sali￾ent object detection: a benchmark[J]. IEEE transactions on [4] image processing, 2015, 24(12): 5706–5722. 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. Las Vegas, 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] REDMON J, FARHADI A. YOLOv3: an incremental im￾provement[J]. arXiv preprint arXiv: 1804.02767, 2018. [7] 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. Columbus, USA, 2014: 580−587. [8] GIRSHICK R. Fast R-CNN[C]//Proceedings of 2015 IEEE International Conference on Computer Vision. Santiago, USA, 2015: 1440−1448. [9] 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 ana￾lysis and machine intelligence, 2017, 39(6): 1137–1149. [10] ROSENFELD A, THURSTON M. Edge and curve detec￾tion for visual scene analysis[J]. IEEE transactions on computers, 1971, C-20(5): 562–569. [11] BODLA N, SINGH B, CHELLAPPA R, et al. Soft￾NMS—improving object detection with one line of code[C]//Proceedings of 2017 IEEE International Confer￾ence on Computer Vision. Venice, Italy, 2017: 5561−5569. [12] HE Yihui, ZHU Chenchen, WANG Jianren, et al. Bound￾ing box regression with uncertainty for accurate object de￾tection[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach, USA, 2019: 2888−2897. [13] 侯志强, 刘晓义, 余旺盛, 等. 基于双阈值-非极大值抑制 的 Faster R-CNN 改进算法 [J]. 光电工程, 2019, 46(12): 190159. HOU Zhiqiang, LIU Xiaoyi, YU Wangsheng, et al. Im￾proved algorithm of Faster R-CNN based on double threshold-non-maximum suppression[J]. Opto-electronic engineering, 2019, 46(12): 190159. [14] [15] 赵文清, 严海, 邵绪强. 改进的非极大值抑制算法的目 第 5 期 王照国，等：基于 F1 值的非极大值抑制阈值自动选取方法 ·1011·