深度学习图像标注与用户标注比较研究

Image Annotation Tags by Deep Learning and Real Users: A Comparative Study

【目的】利用用户对图像标注的标签提出用户标签框架,并通过用户标签框架总结深度学习自动标注图像的不足。【方法】统计分析从Flickr上下载的大约100万张图像数据集中的用户标签,抽取高频词进行用户标签框架匹配。将用户标签与Image Net数据库标签进行对比总结。对含有高频词的图像使用MXNet深度学习算法进行标注,分析标注结果。【结果】当前深度学习自动标注,在图像背景知识、总体描述以及人类感官描述等方面还存在缺陷。【局限】数据集的范围需要扩大,深度学习算法的种类需要增加。【结论】自动标注图像的发展,需要建立图像信息与背景知识、描述等的联系;并且深度学习未来发展还需要赋予计算机逻辑推理以及情境感知的能力。

[Objective] This paper proposes a user tagging framework and examines the limitations of tagging image with deep learning techniques, aiming to improve the performance of automatic annotation services. [Methods] We analyzed the user-added tags from one million images on flickr.com to extract the high frequency ones. Then, we mapped these tags with the proposed framework, and compared them with tags from the Image Net database. Finally, we analyzed images with high frequency tags with the deep learning algorithm-MXNet. [Results] The automatic image annotation techniques based on deep learning could not effectively understand the image＇s background knowledge, as well as the image＇s descriptions from the human perceptive. [Limitations] Our dataset needs to be expanded and analyzed with other deep learning algorithms. [Conclusions] The development of automatic image annotation, requires us to establish the association between image information, background knowledge, and description, as well as cultivate deductive reasoning and context-aware abilities.