《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (11): 3579-3586.DOI: 10.11772/j.issn.1001-9081.2022111660

• 多媒体计算与计算机仿真 • 上一篇    

基于多分支混合注意力的小目标检测算法

秦强强, 廖俊国(), 周弋荀   

  1. 湖南科技大学 计算机科学与工程学院,湖南 湘潭 411201
  • 收稿日期:2022-11-09 修回日期:2023-03-03 接受日期:2023-03-03 发布日期:2023-03-20 出版日期:2023-11-10
  • 通讯作者: 廖俊国
  • 作者简介:秦强强(1997—),男,安徽芜湖人,硕士研究生,CCF会员,主要研究方向:人工智能、目标检测
    廖俊国(1972—),男,湖南衡阳人,教授,博士,CCF会员,主要研究方向:网络安全、人工智能、模式识别 jgliao@hnust.edu.cn
    周弋荀(1998—),男,湖北黄石人,硕士研究生,CCF会员,主要研究方向:人工智能、目标检测。

Small object detection algorithm based on split mixed attention

Qiangqiang QIN, Junguo LIAO(), Yixun ZHOU   

  1. School of Computer Science and Engineering,Hunan University of Science and Technology,Xiangtan Hunan 411201,China
  • Received:2022-11-09 Revised:2023-03-03 Accepted:2023-03-03 Online:2023-03-20 Published:2023-11-10
  • Contact: Junguo LIAO
  • About author:QIN Qiangqiang, born in 1990, M. S. candidate. His research interests include artificial intelligence, object detection.
    LIAO Junguo, born in 1972, Ph. D., professor. Her research interests include cyber security, artificial intelligence, pattern recognition.
    ZHOU Yixun, born in 1998, M. S. candidate. His research interests include artificial intelligence, object detection.

摘要:

针对图像中的小目标特征信息少、占比低、易受环境影响等特点,提出一种基于多分支混合注意力的小目标检测算法SMAM-YOLO。首先,将通道注意力(CA)和空间注意力(SA)相结合,重新组合连接结构,提出一种混合注意力模块(MAM),增强模型对小目标特征在空间维度上的表达能力。其次,根据不同大小的感受野对目标影响的不同,基于混合注意力提出一种多分支混合注意力模块(SMAM);根据输入特征图的尺度自适应调整感受野大小,同时使用混合注意力增强不同分支下对小目标特征信息的捕获能力。最后,使用SMAM改进YOLOv5中的核心残差模块,提出一种基于CSPNet(Cross Stage Partial Network)和SMAM的特征提取模块CSMAM,而且CSMAM的额外计算开销可以忽略不计。在TinyPerson数据集上的实验结果表明,与基线算法YOLOv5s相比,当交并比(IoU)阈值为0.5时,SMAM-YOLO算法的平均检测精度(mAP50)提升了4.15个百分点,且检测速度达到74 frame/s;此外,与现有的一些主流小目标检测模型相比,SMAM-YOLO算法在mAP50上平均提升了1.46~6.84个百分点,且能满足实时性检测的需求。

关键词: 小目标检测, 多分支网络, 混合注意力, 特征融合, 实时检测

Abstract:

Focusing on the characteristics of small objects in images, such as less feature information, low percentage, and easy to be influenced by the environment, a small object detection algorithm based on split mixed attention was proposed, namely SMAM-YOLO. Firstly, by combining Channel Attention (CA) and Spatial Attention (SA), as well as recombining the connection structures, a Mixed Attention Module (MAM) was proposed to enhance the model’s representation of small object features in spatial dimension. Secondly, according to the different influence of receptive fields with different sizes on the object, a Split Mixed Attention Module (SMAM) was proposed to adaptively adjust the size of the receptive field according to the scale of the input feature map, and the mixed attention was used to enhance the ability to capture small object feature information in different branches. Finally, the core residual module in YOLOv5 was improved by using SMAM, and a feature extraction module CSMAM was proposed on the basis of CSPNet (Cross Stage Partial Network) and SMAM, and the additional computational overhead of CSMAM can be ignored. Experimental results on TinyPerson dataset show that compared with the baseline algorithm YOLOv5s, when the Intersection over Union (IoU) threshold is 0.5, the mean Average Precision (mAP50) of SMAM-YOLO algorithm is improved by 4.15 percentage points, and the detection speed reaches 74 frame/s. In addition, compared with some existing mainstream small object detection models, SMAM-YOLO algorithm improves the mAP50 by 1.46 - 6.84 percentage points on average, and it can meet the requirements of real-time detection.

Key words: small object detection, split network, mixed attention, feature fusion, real-time detection

中图分类号: