计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (15): 211-220.DOI: 10.3778/j.issn.1002-8331.2305-0254

• 图形图像处理 • 上一篇    下一篇

采用动态样本分配的特征融合目标检测算法

牛文涛,王鹏,陈遵田,李晓艳,郜辉,孙梦宇   

  1. 1.西安工业大学 电子信息工程学院,西安 710021
    2.西安工业大学 发展规划处,西安 710021
    3.西安机电信息技术研究所,西安 710065
    4.西安工业大学 光电工程学院,西安 710021
  • 出版日期:2024-08-01 发布日期:2024-07-30

Feature Fusion Target Detection Algorithm Using Dynamic Sample Assignment

NIU Wentao, WANG Peng, CHEN Zuntian, LI Xiaoyan, GAO Hui, SUN Mengyu   

  1. 1.School of Electronic Information Engineering, Xi’an Technological University, Xi’an 710021, China
    2.Development Planning Service, Xi’an Technological University, Xi’an 710021, China
    3.Xi’an Institute of Mechanical and Electrical Information Technology, Xi’an 710065, China
    4.School of Optoelectronic Engineering, Xi’an Technological University, Xi’an 710021, China
  • Online:2024-08-01 Published:2024-07-30

摘要: 针对轻量级目标检测算法SSD-Lite检测精度低、对小目标预测能力差等问题,提出了一种采用动态样本分配策略的多尺度特征融合目标检测算法。在轻量级目标检测算法SSD-Lite的颈部网络引入特征金字塔结构(feature pyramid network,FPN),并对其进行轻量化设计,同时引入残差特征增强模块(residual feature augmentation,RFA),采用残差分支注入不同空间的上下文信息来改善高层特征的特征表达,以提升网络对小目标的检测能力;在特征金字塔结构中插入轻量级注意力机制ECA模块,提升网络对重要特征的关注能力;针对网络训练过程中采用的固定交并比(intersection-over-union,IOU)阈值的样本分配策略导致的正负样本分配适应性差、难以选出高质量正样本等问题,设计了一种动态样本分配策略,取消锚框的预设置,采用中心点采样的方式,同时结合样本均值、标准差作为筛选阈值,减少人工先验的影响,在不改变网络结构的情况下提升算法性能。算法在Pascal VOC数据集上测试,实验结果表明:该算法整体预测精度相较于基准算法提升1.9个百分点,对小目标检测能力提升3.3个百分点,算法推理时延仅增加2.32%;实验证明了该算法可以以较小的性能代价,显著提升算法的预测精度。

关键词: 特征金字塔结构, 残差特征增强模块, 轻量级注意力机制, 动态样本分配策略

Abstract: A multi-scale feature fusion target detection algorithm with dynamic sample allocation strategy is proposed to address the problems of low detection accuracy and poor prediction ability of small targets in the lightweight target detection algorithm SSD-Lite. Firstly, the feature pyramid network (FPN) is introduced in the neck network of the lightweight target detection algorithm SSD-Lite and designed to be lightweight, while the residual feature augmentation (RFA) module is introduced, which uses residual branches to inject different. Then, this paper inserts a lightweight attention mechanism ECA module into the feature pyramid structure to improve the ability of network to focus on important features. Finally, to address the problems of poor adaptability of positive and negative sample assignment and difficulty in selecting high-quality positive samples caused by the fixed Intersection-over-Union (IOU) threshold sample assignment strategy used in the network training process, this paper designs a dynamic sample assignment strategy, which eliminates the pre-setting of anchor frames and adopts the centroid sampling method, while combining the sample mean and standard deviation as screening thresholds to reduce the influence of artificial a priori and improve the algorithm performance without changing the network structure. The algorithm is tested on Pascal VOC dataset, and the experimental results show that the overall prediction accuracy of the algorithm is improved by 1.9 percentage points compared with the benchmark algorithm, the detection ability of small targets is improved by 3.3 percentage points, and the inference delay of the algorithm is increased by only 2.32%. The experiments demonstrate that the algorithm can significantly improve the prediction accuracy of the algorithm with a small performance cost.

Key words: feature pyramid structure, residual feature augmentatione module, lightweight attention mechanism, dynamic sample assignment strategy