计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (18): 172-179.DOI: 10.3778/j.issn.1002-8331.2203-0577

• 模式识别与人工智能 • 上一篇    下一篇

面向小目标检测的并行高分辨率网络设计

牛润,曲毅,郑乐辉,魏建国   

  1. 1.武警工程大学 研究生大队,西安 710086
    2.武警工程大学 信息工程学院,西安 710086
  • 出版日期:2022-09-15 发布日期:2022-09-15

Parallel High-Resolution Network Design for Small Object Detection

NIU Run, QU Yi, ZHENG Lehui, WEI Jianguo   

  1. 1.Postgraduate Brigade, Engineering University of PAP, Xi’an 710086, China
    2.School of Information Engineering, Engineering University of PAP, Xi’an 710086, China
  • Online:2022-09-15 Published:2022-09-15

摘要: 当前目标检测算法对小目标检测存在特征信息易丢失的问题,利用网络处理高分辨率特征图数据可以缓解,但存在语义信息不足和计算负担大的缺点。为弥补这些缺点,提出一种有效处理高分辨率特征图、多深度子网并行连接的特征提取网络。构建输入图像金字塔,搭建多深度分支子网并行连接的结构,使用浅层网络处理图像金字塔中高分辨率特征图,深层网络处理低分辨率特征图,多分支同时运行并在中间位置进行两次特征融合,充分结合高分辨率特征信息和低分辨率语义信息;使用融合因子构建对小目标针对性强的多尺度特征融合结构,增强对小目标检测能力;使用注意力机制进一步提高特征提取能力。在公开数据集AI-TOD上进行实验表明,所设计的特征提取网络相较于其他常用特征提取网络对小目标的检测能力更强,在two-stage经典模型Faster-RCNN、one-stage经典模型SSD、YOLOv3以及anchor-free经典模型CenterNet上替换上原主干网络,检测平均精度mAP与原来相比分别提升了2.7、3.4、3.3、1.7个百分点,证明了所提网络结构的适用性和有效性。

关键词: 小目标检测, 特征提取, 多尺度, 上下文信息, 注意力机制

Abstract: The current object detection algorithm has the problem of easy loss of feature information for small object detection, which can be alleviated by using the network to process high-resolution feature map data, but it has the shortcomings of insufficient semantic information and large computational burden. To remedy these shortcomings, this paper proposes a feature extraction network that effectively handles high-resolution feature maps and parallel connections of multiple depth subnetworks. This paper constructs an input image pyramid, builds a parallel connection structure of multi-depth branch subnets, uses a shallow network to process high-resolution feature maps in the image pyramid, and uses a deep network to process low-resolution feature maps, multi-branch runs at the same time and performs two feature fusions in the middle position, fully combining high-resolution feature information and low-resolution semantic information. Using the fusion factor to build a multi-scale feature fusion structure that is highly targeted to small targets to enhance the detection ability of small targets. Using attention mechanism to further improve feature extraction ability. Experiments on the public dataset AI-TOD show that the designed feature extraction network has stronger detection ability for small targets than other commonly used feature extraction networks. Replacing the original backbone network on the two-stage classic model Faster-RCNN, the one-stage classic model SSD, YOLOv3 and the anchor-free classic model CenterNet. Compared with the original, the average detection accuracy of mAP is increased by 2.7, 3.4, 3.3 and 1.7 percentage points respectively, which proves the applicability and effectiveness of the proposed network structure.

Key words: small object detection, feature extraction, multi-scale, contextual information, attention mechanism