航空学报 > 2022, Vol. 43 Issue (1): 24889-024889   doi: 10.7527/S1000-6893.2021.24889

面向目标分类识别的多任务学习算法综述

李红光1, 王菲2, 丁文锐1   

  1. 1. 北京航空航天大学 无人系统研究院, 北京 100191;
    2. 北京航空航天大学 电子信息工程学院, 北京 100191
  • 收稿日期:2020-10-19 修回日期:2021-04-28 出版日期:2022-01-15 发布日期:2021-03-26
  • 通讯作者: 李红光 E-mail:lihongguang@buaa.edu.cn
  • 基金资助:
    国家自然科学基金面上项目(62076019)

Survey on multi-task learning for object classification and recognition

LI Hongguang1, WANG Fei2, DING Wenrui1   

  1. 1. Research Institute of Unmanned System, Beihang University, Beijing 100191, China;
    2. School of Electronics and Information Engineering, Beihang University, Beijing 100191, China
  • Received:2020-10-19 Revised:2021-04-28 Online:2022-01-15 Published:2021-03-26
  • Supported by:
    Surface Project of National Natural Science Foundation of China(62076019)

摘要: 多任务学习(MTL)可以在训练中联合利用多个任务的监督信号,并通过共享多个相关任务之间的有用信息来提升模型性能。本文从目标分类识别应用角度,全面梳理和分析了多任务学习的机制及其主流方法。首先,对多任务学习的定义、原理和方法进行阐述。其次,以应用较为广泛、具有代表性且具有共性特点的细粒度分类和目标重识别为例,重点介绍多任务学习机制在目标分类和识别任务应用的2类方法:基于任务层的多任务学习和基于特征层的多任务学习,并针对每种类型进一步分类分析不同的多任务学习算法的设计思想和优缺点。接着,对本文综述的各种多任务学习算法在通用数据集上开展性能对比。最后,对面向目标分类和识别任务的多任务学习方法的未来趋势进行展望。

关键词: 多任务学习, 深度学习, 目标分类, 细粒度分类, 目标重识别

Abstract: Multi-Task Learning(MTL) aims to enhance the model performance by jointly leveraging supervisory signals and sharing useful information among multiple related tasks. This paper comprehensively summarizes and analyzes the mechanism and mainstream methods of multi-task learning for object classification and recognition applications. First, we review the definitions, principles and methods of MTL. Second, taking the representative and widely used fine-grained classification and object re-identification as examples, we emphatically introduce two types of multi-task learning for object classification and recognition: task-based multi-task learning and feature-based multi-task learning, and further categorize each type and analyze the design ideas, and advantages and disadvantages of different MTL algorithms. Third, we compare the performance of various MTL algorithms reviewed in this paper on common datasets. Finally, prospects on development trends of MTL algorithms for object classification and recognition are discussed.

Key words: multi-task learning, deep learning, object classification, fine-grained classification, object re-identification

中图分类号: