基于条件生成对抗网络的交通环境多任务语义分割方法研究
|
摘要
道路交通环境具有复杂、强干扰、多遮挡、检测物体尺度变化大、光线不均匀、难以预测的特点,传统基于全卷积神经网络的分割方法,由于采用单一的检测结果评价标准,缺乏对分割结果一致性的检验,忽略了像素与像素的相互关系,造成误识别很可能导致交通事故发生。本文在传统交并比评价指标的基础之上,采用交通环境语义分割复合评价指标,提出基于条件生成对抗网络的交通环境多任务语义分割方法,采用对抗损失拟合语义分割结果像素之间的作用关系,使得结果更具备一致性和可用性,更利于实际应用,同时对比了三种典型交通环境检测任务,验证了算法的有效性,并对三种任务进行多任务学习,在不增加计算开销的基础上,获得相近的性能。
The road traffic environment has complex,strong interference,multiple occlusion,large changes in the scale of the detected object,uneven light,and difficult to predict. The traditional segmentation method based on the full convolutional neural network lacks consistency in results due to the adoption of a single evaluation standard.The test ignores the relationship between pixels and pixels,which causes misidentification and is likely to cause traffic accidents. Based on the cross-comparison ratio,this paper adopts the traffic environment semantic segmentation composite evaluation index,and proposes a multi-task semantic segmentation method based on conditional generation confrontation network,which uses the action relationship between the pixels to resist the loss and makes the result more suitable. Moreover,three typical traffic environment detection tasks are compared,and the effectiveness of the algorithm is verified. On this basis,multi-task learning is performed on three tasks,and similar performance is obtained without increasing computational overhead.
The road traffic environment has complex,strong interference,multiple occlusion,large changes in the scale of the detected object,uneven light,and difficult to predict. The traditional segmentation method based on the full convolutional neural network lacks consistency in results due to the adoption of a single evaluation standard.The test ignores the relationship between pixels and pixels,which causes misidentification and is likely to cause traffic accidents. Based on the cross-comparison ratio,this paper adopts the traffic environment semantic segmentation composite evaluation index,and proposes a multi-task semantic segmentation method based on conditional generation confrontation network,which uses the action relationship between the pixels to resist the loss and makes the result more suitable. Moreover,three typical traffic environment detection tasks are compared,and the effectiveness of the algorithm is verified. On this basis,multi-task learning is performed on three tasks,and similar performance is obtained without increasing computational overhead.
引文
[1]梁博,于蕾,李爽.基于卷积神经网络的多任务图像语义分割[J].无线电工程,2019,07:85~91.
[2]侯津京.基于GAN的城市环境图像语义分割[D].吉林:吉林大学,2018:1~7.
[3]Cun Y L,Boser B,Denker J S,et al. Handwritten digit recognition with a back-propagation network[J]. Advances in Neural Information Processing Systems,1990,2(2):396~404.
[4]Srivastava N,Hinton G,Krizhevsky A,et al. Dropout:a simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research,2014,15(1):1929~1958.
[5]Goodfellow I. NIPS 2016 Tutorial:Generative Adversarial Networks[J]. 2016:19~25.
[6]Brock A,Donahue J,Simonyan K. Large Scale GAN Training for High Fidelity Natural Image Synthesis[J]. 2018:396~404.
[7]Arjovsky M,Chintala S,Bottou,Léon. Wasserstein GAN[J].2nd IEEE International Conference on Computational Intelligence and Applications(ICCIA),2017:288~312.
[8]Huang X,Wang P,Cheng X,et al. The ApolloScape Open Dataset for Autonomous Driving and its Application[J]. 2018:23~45.
[2]侯津京.基于GAN的城市环境图像语义分割[D].吉林:吉林大学,2018:1~7.
[3]Cun Y L,Boser B,Denker J S,et al. Handwritten digit recognition with a back-propagation network[J]. Advances in Neural Information Processing Systems,1990,2(2):396~404.
[4]Srivastava N,Hinton G,Krizhevsky A,et al. Dropout:a simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research,2014,15(1):1929~1958.
[5]Goodfellow I. NIPS 2016 Tutorial:Generative Adversarial Networks[J]. 2016:19~25.
[6]Brock A,Donahue J,Simonyan K. Large Scale GAN Training for High Fidelity Natural Image Synthesis[J]. 2018:396~404.
[7]Arjovsky M,Chintala S,Bottou,Léon. Wasserstein GAN[J].2nd IEEE International Conference on Computational Intelligence and Applications(ICCIA),2017:288~312.
[8]Huang X,Wang P,Cheng X,et al. The ApolloScape Open Dataset for Autonomous Driving and its Application[J]. 2018:23~45.
© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号 地址：北京市海淀区学院路29号 邮编：100083 电话：办公室：(+86 10)66554848；文献借阅、咨询服务、科技查新：66554700 |