数据与计算发展前沿 ›› 2020, Vol. 2 ›› Issue (2): 120-135.doi: 10.11871/jfdc.issn.2096-742X.2020.02.010

• 专刊: 数据分析技术与应用 • 上一篇    下一篇

基于深度学习的小目标检测与识别

冷佳旭1,2,刘莹1,2,*()   

  1. 1. 中国科学院大学计算机科学与技术学院,北京 100089
    2. 中国科学院大学数据挖掘与高性能计算实验室,北京 101400
  • 出版日期:2020-04-20 发布日期:2020-06-03
  • 通讯作者: 刘莹 E-mail:yingliu@ucas.ac.cn
  • 作者简介:冷佳旭,博士生,目前就读于中国科学院大学。主要研究方向包括:计算机视觉、深度学习、目标检测、目标跟踪和双目立体视觉。
    本文主要负责算法设计与实验验证部分。
    Leng Jiaxu is currently pursuing his Ph.D. degree in School of Computer Science and Tecnology in University of Chinese Academy of Sciences. His current research interests include computer vision, deep learning, object detection, object tracking, and stereo vision.
    In this paper, he is responsible for the design and experimental analysis of the proposed algorithms.
    E-mail: lengjiaxu17@mails.ucas.ac.cn|刘莹,中国科学院大学教授,中国科学院数据挖掘与高性能计算实验室负责人。主要研究方向包括数据挖掘、人工智能、并行计算等。
    本文中完成了论文的国内外现状分析、方法原理和结论展望。
    Liu Ying is currently a professor of School of Computer Science and Tecnology in University of Chinese Academy of Sciences, and the Dean of the Data Mining and High Performanle Computing Lab. Her research interests include data mining, artificial intelligence, parallel computing, etc.
    In this paper, she is responsible for the literature review, principles and conclusions.
  • 基金资助:
    国家自然科学基金(71671178);国家自然科学基金(91546201);中国科学院大学优秀青年教师科研能力提升重点项目

Small Object Detection and Recognition Based onDeep Learning

Leng Jiaxu1,2,Liu Ying1,2,*()   

  1. 1. School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing 100089, China
    2. Data Mining and High Performance Computing Lab, University of Chinese Academy of Sciences, Beijing 101400, China
  • Online:2020-04-20 Published:2020-06-03
  • Contact: Ying Liu E-mail:yingliu@ucas.ac.cn

摘要:

【目的】目前,现有的基于深度学习的检测算法针对小目标的检测效果较差。本文旨在通过充分考虑小目标的特点来提升小目标的检测与识别性能。【方法】本文从不同方面来提升小目标检测与识别,其中包括特征融合、上下文学习和注意力机制。针对小目标特征难以提取问题,提出一种双向特征融合的方法。另外,鉴于小目标特征不明显问题,提出一种利用上下文信息来提升检测性能的方法。更进一步,为了更好地识别小目标的类别,提出一种注意力转移的方法。【结果】实验结果表明,我们提出的方法在公共数据集上均显著地提高了小目标的检测和识别性能。【结论】研究特征融合、上下文利用和注意力机制的方法对于提升小目标检测与识别是非常有价值的。

关键词: 小目标检测, 特征融合, 上下文学习, 注意力机制

Abstract:

[Objective] In this paper, we aim to improve the detection performance for small objects by considering the characteristics of small objects under deep learning-based detection frameworks. [Methods] This paper improves small object detection and recognition performance from different aspects, including feature fusion, context learning and attention mechanism. Since the features of the small object are not evident, a bidirectional feature fusion method is proposed to improve the feature expression capability for small objects. In addition, a novel method is proposed to improve the detection performance by using the context information of small objects. Furthermore, to better identify the categories of small objects, an attention transfer method is proposed to improve the recognition rate. [Results] Experimental results show that the three proposed methods can significantly improve the detection and recognition performance for small objects on public datasets. [Conclusions] The research on feature fusion, context utilization and attention mechanism is very valuable for improving small object detection in complex scenes.

Key words: small object detection, feature fusion, context learning, attention mechanism