Ensemble Teacher: Semi-Supervised Semantic Segmentation with Teacher’s Cross-Pseudo Supervision
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Semi-supervised learning utilizing both labeled and unlabeled data has recently made significant progress; however, while semi-supervised image classification research has been actively conducted and solved many problems, semisupervised semantic segmentation research still has some problems that need to be solved. In this work, semi-supervised semantic segmentation, which requires more annotation work than other tasks such as classification, was investigated because labels are required on a pixel-by-pixel basis. Consistency regularization based on input image perturbation, feature perturbation, and network perturbation in the semi-supervised learning approach showed significant performance improvement. In particular, consistency regularization through network perturbation showed excellent performance in semi-supervised semantic segmentation. However, important challenges remain. Due to pseudo-label bias, inaccurate pseudo-labels for unlabeled data can negatively affect training, which causes more serious problems in semi supervised segmentation tasks. This study mitigates the pseudo-label bias issue by expanding the network perturbation-based method, which has shown excellent results, through a simple ensemble strategy. In addition, this simple method shows improved prediction performance. The results of this study outperform previous state-of-the-art methods and achieve state-of-the-art semi-supervised segmentation performance. The proposed method can be practically applied to the handson data science domain that hardly obtains sufficient class-labels (e.g. medicine, manufacturing, etc.).
Abstract
Semi-supervised learning utilizing both labeled and unlabeled data has recently made significant progress; however, while semi-supervised image classification research has been actively conducted and solved many problems, semisupervised semantic segmentation research still has some problems that need to be solved. In this work, semi-supervised semantic segmentation, which requires more annotation work than other tasks such as classification, was investigated because labels are required on a pixel-by-pixel basis. Consistency regularization based on input image perturbation, feature perturbation, and network perturbation in the semi-supervised learning approach showed significant performance improvement. In particular, consistency regularization through network perturbation showed excellent performance in semi-supervised semantic segmentation. However, important challenges remain. Due to pseudo-label bias, inaccurate pseudo-labels for unlabeled data can negatively affect training, which causes more serious problems in semi supervised segmentation tasks. This study mitigates the pseudo-label bias issue by expanding the network perturbation-based method, which has shown excellent results, through a simple ensemble strategy. In addition, this simple method shows improved prediction performance. The results of this study outperform previous state-of-the-art methods and achieve state-of-the-art semi-supervised segmentation performance. The proposed method can be practically applied to the handson data science domain that hardly obtains sufficient class-labels (e.g. medicine, manufacturing, etc.).
참고문헌 (55)
- Krizhevsky A, "Imagenet classification with deep convolutional neural networks", NIPS, pp.25~ (2012)
- He K, "Deep residual learning for image recognition", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.770~778 (2016)
- Ren S, "Faster r-cnn : Towards real-time object detection with region proposal networks", NIPS, pp.28~ (2015)
- Redmon J, "You only look once : Unified, real-time object detection", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.779~788 (2016)
- Long J, "Fully convolutional networks for semantic segmentation", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.3431~3440 (2015)
- Chen LC, "Semantic image segmentation with deep convolutional nets and fully connected crfs", arXiv preprint arXiv:1412.7062, ~ (2014)
- Suh SM, "Weighted mask R-CNN for improving adjacent boundary segmentation", J Sens, 2021, pp.1~8 (2021)
- Deng J, "Imagenet: A large-scale hierarchical image database", 2009 IEEE conference on computer vision and pattern recognition, pp.248~255 (2009), Ieee
- Park YE, "Impacts of heavy rain and floodwater on floating debris entering an artificial lake(Daecheong Reservoir, Korea)during the summer", Desalination Water Treat, 219, pp.399~404 (2021)
- Ko KM, "Squeeze Face : Integrative face recognition methods with LiDAR sensors", J Sens, 2021, pp.1~8 (2021)
- Zhai X, "S4l : Self-supervised semi-supervised learning", Proceedings of the IEEE, 2019, 1476~1485
- 고경민, "Action Recognition Model to Monitor Illegal Dumping using Zoom-In Image Sequence Data", Quantitative Bio-Science, 40(2), pp.95~102 (2021), 자연과학연구소[DOI]
- 민진홍, "Grammatical Error Correction Models for Korean Language via Pre-trained Denoising", Quantitative Bio-Science, 39(1), pp.17~24 (2020), 자연과학연구소[DOI]
- Chen X, "Semi-supervised semantic segmentation with cross pseudo supervision", Proceedings of the IEEE, 2021, 2613~2622
- Jung SH, "Estimation of particulate levels using deep dehazing network and temporal prior", J Sens, 2020, pp.1~9 (2020)
- Filipiak D, "n-CPS: Generalising cross pseudo supervision to n networks for semi-supervised semantic segmentation", arXiv preprint arXiv:2112.07528, ~ (2021)
- French G, "Semisupervised semantic segmentation needs strong, high-dimensional perturbations", ~ (2019)
- Sohn KH, "Fixmatch : Simplifying semi-supervised learning with consistency and confidence", NIPS 2020, 33, pp.596~608 (2020)
- 이현, "AI-Based Dental Prostheses Fabrication Using Generative Adversarial Networks", Quantitative Bio-Science, 40(1), pp.39~44 (2021), 자연과학연구소[DOI]
- Park YE, "A machine learning approach for early warning of cyanobacterial bloom outbreaks in a freshwater reservoir", J Environ Manage, 288, pp.112415~ (2021)
- Han JS, "Reinforcement learning guided by double replay memory", J Sens, 2021, pp.1~8 (2021)
- Zou Y, "Pseudoseg: Designing pseudo labels for semantic segmentation", arXiv preprint arXiv:2010.09713, ~ (2020)
- Ke Z, "Guided collaborative training for pixel-wise semi-supervised learning", European Conference on Computer Vision, pp.429~445 (2020), Springer
- Ouali Y, "Semi-supervised semantic segmentation with cross-consistency training", Proceedings of the IEEE, 2020, 12674~12684
- Lee DH, "Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks", Workshop on challenges in representation learning, ICML, 3, pp.896~ (2013)
- Arazo E, "Pseudo-labeling and confirmation bias in deep semi-supervised learning", 2020 International Joint Conference on Neural Networks (IJCNN), pp.1~8 (2020), IEEE
- Berthelot D, "Mixmatch : A holistic approach to semi-supervised learning", NIPS 2019, pp.32~ (2019)
- Oksuz K, "Imbalance problems in object detection : A review", IEEE Trans Pattern Anal Mach Intell, 43, pp.3388~3415 (2020)
- 김성환, "Instance Segmentation Guided by Weight Map with Application to Tooth Boundary Detection", Quantitative Bio-Science, 39(2), pp.159~167 (2020), 자연과학연구소[DOI]
- Everingham M, "The pascal visual object classes challenge : A retrospective", IJCVI, 111, pp.98~136 (2015)
- Cordts M, "The cityscapes dataset for semantic urban scene understanding", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.3213~3223 (2016)
- Van Engelen JE, "A survey on semi-supervised learning", Mach Learn, 109, pp.373~440 (2020)
- Tarvainen A, "Mean teachers are better role models : Weight-averaged consistency targets improve semi-supervised deep learning results", NIPS 2017, pp.30~ (2017)
- Xie Q, "Self-training with noisy student improves imagenet classification", Proceedings of the IEEE, 2020, 10687~10698
- Pham H, "Meta pseudo labels", Proceedings of the IEEE, 2021, 11557~11568
- Liu YC, "Unbiased teacher for semi-supervised object detection", arXiv preprint arXiv:2102.09480, ~ (2021)
- Lin TY, "Focal loss for dense object detection", Proceedings of the IEEE International Conference on Computer Vision, pp.2980~2988 (2017)
- Geifman Y, "Selectivenet: A deep neural network with an integrated reject option", International Conference on Mach. Learn. PMLR, pp.2151~2159 (2019)
- Xu M, "End-to-end semi-supervised object detection with soft teacher", Proceedings of the IEEE, 2021, 3060~3069
- Yun SD, "Cutmix : Regularization strategy to train strong classifiers with localizable features", Proceedings of the IEEE, 2019, 6023~6032
- Olsson V, "Classmix : Segmentation-based data augmentation for semi-supervised learning", Proceedings of the IEEE, 2021, 1369~1378
- Zhang H, "mixup: Beyond empirical risk minimization", arXiv preprint arXiv:1710.09412, ~ (2017)
- Laine S, "Temporal ensembling for semi-supervised learning", arXiv preprint arXiv:1610.02242, ~ (2016)
- Chen LC, "Deeplab : Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs", IEEE Trans Pattern Anal Mach Intell, 40, pp.834~848 (2017)
- Zhao H, "Pyramid scene parsing network", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.2881~2890 (2017)
- Yu F, "Multi-scale context aggregation by dilated convolutions", arXiv preprint arXiv:1511.07122, ~ (2015)
- Vaswani A, "Attention is all you need", NIPS 2017, pp.30~ (2017)
- Chen LC, "Attention to scale : Scale-aware semantic image segmentation", Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.3640~3649 (2016)
- Chen LC, "Encoderdecoder with atrous separable convolution for semantic image segmentation", Proceedings of the European Conference on Computer Vision(ECCV), pp.801~818 (2018)
- Kingma DP, "Adam: A method for stochastic optimization", arXiv preprint arXiv:1412.6980, ~ (2014)
- Ioffe S, "Batch normalization: Accelerating deep network training by reducing internal covariate shift", International Conference on Mach. Learn. PMLR, pp.448~456 (2015)
- Kwon DH, "Semi-supervised semantic segmentation with error localization network", arXiv preprint arXiv:2204.02078, ~ (2022)
- Hariharan B, "Semantic contours from inverse detectors", 2011 International Conference on Computer Vision, pp.991~998 (2011), IEEE
- He K, "Delving deep into rectifiers : Surpassing human-level performance on imagenet classification", Proceedings of the IEEE International Conference on Computer Vision, pp.1026~1034 (2015)
- Lai X, "Semisupervised semantic segmentation with directional context-aware consistency", Proceedings of the IEEE, 2021, 1205~1214