{"title":"基于迁移学习的半监督实例分割算法","authors":"Bing Liu, Ren Yi, Zhongquan Yu, Shiyu Wang, Xuewen Yang, Fuwen Wang","doi":"10.1080/10589759.2023.2274013","DOIUrl":null,"url":null,"abstract":"ABSTRACTSemi-supervised instance segmentation algorithms are mainly divided into algorithms based on pseudo-label generation and algorithms based on transfer learning. The algorithms based on pseudo-label generation need to design a specific pseudo-label generation process, but the process is not scalable for different types of source tasks. The algorithms based on transfer learning that started late have relatively high scalability, but the algorithm research ideas are relatively simple. To expand the research on semi-supervised instance segmentation based on transfer learning, this paper proposes a feature transfer-based semi-supervised instance segmentation algorithm Feature Transfer Mask R-CNN (FT-Mask). The FT-Mask algorithm is more scalable than algorithms based on pseudo-label generation and can be used to transfer knowledge from different types of source tasks. Compared with other semi-supervised instance segmentation algorithms based on transfer learning, FT-Mask uses the feature transfer method to achieve semi-supervised instance segmentation for the first time. The experimental results show that the FT-Mask model improves the semi-supervised instance segmentation accuracy of the Mask R-CNN benchmark model through the semi-supervised learning process, and can achieve effective transfer learning.KEYWORDS: Semi-supervised learninginstance segmentationtransfer learning Disclosure statementNo potential conflict of interest was reported by the authors.Additional informationFundingThis work was supported in part by the National Natural Science Foundation of China under Grant 62072319; the Sichuan Science and Technology Program under Grant 2023YFQ0022 and 2022YFG0041; the Luzhou Science and Technology Innovation R&D Program (No. 2022CDLZ-6).","PeriodicalId":49746,"journal":{"name":"Nondestructive Testing and Evaluation","volume":" 3","pages":"0"},"PeriodicalIF":3.0000,"publicationDate":"2023-11-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Semi-supervised instance segmentation algorithm based on transfer learning\",\"authors\":\"Bing Liu, Ren Yi, Zhongquan Yu, Shiyu Wang, Xuewen Yang, Fuwen Wang\",\"doi\":\"10.1080/10589759.2023.2274013\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"ABSTRACTSemi-supervised instance segmentation algorithms are mainly divided into algorithms based on pseudo-label generation and algorithms based on transfer learning. The algorithms based on pseudo-label generation need to design a specific pseudo-label generation process, but the process is not scalable for different types of source tasks. The algorithms based on transfer learning that started late have relatively high scalability, but the algorithm research ideas are relatively simple. To expand the research on semi-supervised instance segmentation based on transfer learning, this paper proposes a feature transfer-based semi-supervised instance segmentation algorithm Feature Transfer Mask R-CNN (FT-Mask). The FT-Mask algorithm is more scalable than algorithms based on pseudo-label generation and can be used to transfer knowledge from different types of source tasks. Compared with other semi-supervised instance segmentation algorithms based on transfer learning, FT-Mask uses the feature transfer method to achieve semi-supervised instance segmentation for the first time. The experimental results show that the FT-Mask model improves the semi-supervised instance segmentation accuracy of the Mask R-CNN benchmark model through the semi-supervised learning process, and can achieve effective transfer learning.KEYWORDS: Semi-supervised learninginstance segmentationtransfer learning Disclosure statementNo potential conflict of interest was reported by the authors.Additional informationFundingThis work was supported in part by the National Natural Science Foundation of China under Grant 62072319; the Sichuan Science and Technology Program under Grant 2023YFQ0022 and 2022YFG0041; the Luzhou Science and Technology Innovation R&D Program (No. 2022CDLZ-6).\",\"PeriodicalId\":49746,\"journal\":{\"name\":\"Nondestructive Testing and Evaluation\",\"volume\":\" 3\",\"pages\":\"0\"},\"PeriodicalIF\":3.0000,\"publicationDate\":\"2023-11-09\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Nondestructive Testing and Evaluation\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1080/10589759.2023.2274013\",\"RegionNum\":3,\"RegionCategory\":\"材料科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"MATERIALS SCIENCE, CHARACTERIZATION & TESTING\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Nondestructive Testing and Evaluation","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1080/10589759.2023.2274013","RegionNum":3,"RegionCategory":"材料科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"MATERIALS SCIENCE, CHARACTERIZATION & TESTING","Score":null,"Total":0}
引用次数: 0
摘要
摘要半监督实例分割算法主要分为基于伪标签生成的算法和基于迁移学习的算法。基于伪标签生成的算法需要设计特定的伪标签生成过程,但该过程不能针对不同类型的源任务进行扩展。起步较晚的基于迁移学习的算法具有较高的可扩展性,但算法研究思路相对简单。为了拓展基于迁移学习的半监督实例分割研究,本文提出了一种基于特征迁移的半监督实例分割算法feature transfer Mask R-CNN (FT-Mask)。FT-Mask算法比基于伪标签生成的算法更具可扩展性,可用于从不同类型的源任务转移知识。与其他基于迁移学习的半监督实例分割算法相比,FT-Mask首次利用特征转移方法实现了半监督实例分割。实验结果表明,FT-Mask模型通过半监督学习过程提高了Mask R-CNN基准模型的半监督实例分割精度,并能实现有效的迁移学习。关键词:半监督学习实例分割迁移学习披露声明作者未报告潜在利益冲突。本研究得到国家自然科学基金项目(62072319)的部分资助;四川省科技计划项目2023YFQ0022和2022YFG0041;泸州市科技创新发展计划(2022CDLZ-6)。
Semi-supervised instance segmentation algorithm based on transfer learning
ABSTRACTSemi-supervised instance segmentation algorithms are mainly divided into algorithms based on pseudo-label generation and algorithms based on transfer learning. The algorithms based on pseudo-label generation need to design a specific pseudo-label generation process, but the process is not scalable for different types of source tasks. The algorithms based on transfer learning that started late have relatively high scalability, but the algorithm research ideas are relatively simple. To expand the research on semi-supervised instance segmentation based on transfer learning, this paper proposes a feature transfer-based semi-supervised instance segmentation algorithm Feature Transfer Mask R-CNN (FT-Mask). The FT-Mask algorithm is more scalable than algorithms based on pseudo-label generation and can be used to transfer knowledge from different types of source tasks. Compared with other semi-supervised instance segmentation algorithms based on transfer learning, FT-Mask uses the feature transfer method to achieve semi-supervised instance segmentation for the first time. The experimental results show that the FT-Mask model improves the semi-supervised instance segmentation accuracy of the Mask R-CNN benchmark model through the semi-supervised learning process, and can achieve effective transfer learning.KEYWORDS: Semi-supervised learninginstance segmentationtransfer learning Disclosure statementNo potential conflict of interest was reported by the authors.Additional informationFundingThis work was supported in part by the National Natural Science Foundation of China under Grant 62072319; the Sichuan Science and Technology Program under Grant 2023YFQ0022 and 2022YFG0041; the Luzhou Science and Technology Innovation R&D Program (No. 2022CDLZ-6).
期刊介绍:
Nondestructive Testing and Evaluation publishes the results of research and development in the underlying theory, novel techniques and applications of nondestructive testing and evaluation in the form of letters, original papers and review articles.
Articles concerning both the investigation of physical processes and the development of mechanical processes and techniques are welcomed. Studies of conventional techniques, including radiography, ultrasound, eddy currents, magnetic properties and magnetic particle inspection, thermal imaging and dye penetrant, will be considered in addition to more advanced approaches using, for example, lasers, squid magnetometers, interferometers, synchrotron and neutron beams and Compton scattering.
Work on the development of conventional and novel transducers is particularly welcomed. In addition, articles are invited on general aspects of nondestructive testing and evaluation in education, training, validation and links with engineering.