ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (9): 1951-1974.doi: 10.7544/issn1000-1239.2021.20200281

• 人工智能 • 上一篇    下一篇



  1. 1(广西多源信息挖掘与安全重点实验室(广西师范大学) 广西桂林 541004);2(西北师范大学计算机科学与工程学院 兰州 730070);3(中国科学院智能信息处理重点实验室(中国科学院计算技术研究所) 北京 100190) (
  • 出版日期: 2021-09-01
  • 基金资助: 

Research Progress on Image Captioning

Li Zhixin1, Wei Haiyang1, Zhang Canlong1, Ma Huifang2, Shi Zhongzhi3   

  1. 1(Guangxi Key Laboratory of Multi-Source Information Mining and Security (Guangxi Normal University), Guilin, Guangxi 541004);2(College of Computer Science and Engineering, Northwest Normal University, Lanzhou 730070);3(Key Laboratory of Intelligent Information Processing (Institute of Computing Technology, Chinese Academy of Sciences), Chinese Academy of Sciences, Beijing 100190)
  • Online: 2021-09-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61966004, 61663004, 61866004, 61762078) and the Guangxi Natural Science Foundation (2019GXNSFDA245018, 2018GXNSFDA281009, 2017GXNSFAA198365).

摘要: 图像描述生成结合了计算机视觉和自然语言处理2个研究领域,不仅要求完备的图像语义理解,还要求复杂的自然语言表达,是进一步研究符合人类感知的视觉智能的关键任务.对图像描述生成的研究进展做了回顾.首先,归纳分析了当前基于深度学习的图像描述生成方法涉及的5个关键技术,包括整体架构、学习策略、特征映射、语言模型和注意机制.然后,按照发展进程将现有的图像描述生成方法分为四大类,即基于模板的方法、基于检索的方法、基于编码器-解码器架构的方法和基于复合架构的方法,并阐述了各类方法的基本概念、代表性方法和研究现状,重点讨论了基于编码器-解码器架构的各种方法及其创新思路,如多模态空间、视觉空间、语义空间、注意机制、模型优化等.接着,从实验的角度给出图像描述生成的常用数据集和评估措施,并在2个基准数据集上比较了一些典型方法的性能.最后,以提升图像描述的准确性、完整性、新颖性、多样性为依据,展示了图像描述生成的未来发展趋势.

关键词: 图像描述生成, 编码器-解码器架构, 复合架构, 注意机制, 卷积神经网络, 循环神经网络, 长短期记忆网络

Abstract: Image captioning combines the two research fields of computer vision and natural language processing. It requires not only complete image semantic understanding, but also complex natural language expression. It is a crucial task for further research on visual intelligence in line with human perception. This paper reviews the research progress on image captioning. Firstly, five key technologies involved in current deep learning based image captioning methods are summarized and analyzed, including overall architecture, learning strategy, feature mapping, language model and attention mechanism. Then, according to the development process, the existing image captioning methods are divided into four categories, i.e. template based methods, retrieval based methods, encoder-decoder architecture based methods and compositional architecture based methods. We describe the basic concepts, representative methods and research status of each category. Furthermore, we emphatically discuss the various methods based on encoder-decoder architecture and their innovative ideas, such as multimodal space, visual space, semantic space, attention mechanism, model optimization, and so on. Subsequently, from the experimental point of view, we show the common benchmark datasets and evaluation measures in the field of image captioning. In addition, we compare the performance of some typical methods on two benchmark datasets. Finally, based on improving the accuracy, integrity, novelty and diversity of image caption, several future development trends of image captioning are presented.

Key words: image captioning, encoder-decoder architecture, compositional architecture, attention mechanism, convolutional neural network, recurrent neural network, long short-term memory