Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (12): 91-100.DOI: 10.3778/j.issn.1002-8331.2307-0206

• Theory, Research and Development • Previous Articles     Next Articles

Double Path Feature Pyramid Network Detection Algorithm Based on Integrating Non-Dimensionality-Reducing Locally Cross-Channel Interaction Strategy

JIAO Bowen, WANG Yulin, WANG Peng, WANG Hongchang, YU Yixuan, SHEN Zhengkun   

  1. 1.College of Mechanical and Electrical Engineering, Qingdao University, Qingdao, Shandong 266071, China
    2.Nation Engineering Research Center for Intelligent Electrical Vehical Power System (Qingdao), Qingdao, Shandong 266071, China
  • Online:2024-06-15 Published:2024-06-14

融合不降维局部跨通道交互策略的双路径金字塔检测算法

焦博文,王玉林,王鹏,王洪昌,于奕轩,沈正坤   

  1. 1.青岛大学 机电工程学院,山东 青岛 266071
    2.电动汽车智能化动力集成技术国家地方联合工程研究中心(青岛),山东 青岛 266071

Abstract: In the object detection task for road obstacles, the detection accuracy of multi-scale objects is low, and the detection robustness and generalization ability are poor in different scenes. Based on the traditional feature pyramid network, the improved algorithm proposes a double path feature pyramid network that combines top-down and bottom-up features, and retains more shallow and deep semantic information in the prediction feature layer through feature concat and fusion methods. On this basis, a mechanism attention module with spatial and channel mechanisms in series is proposed, which further improves the network model detection performance by adopting a local cross-channel interaction strategy without dimension reduction. The experimental results show that the accuracy of object detection of the improved algorithm is 4.6?percentage points higher than that of the original algorithm. The accuracy of small object detection increased by 11.76 percentage points. The accuracy of medium object detection is increased by 5.78?percentage points. The accuracy rate of large object detection is increased by 3.7?percentage points.

Key words: computer vision, object detection, deep learning, feature pyramid network, attention mechanism

摘要: 针对道路障碍物目标检测任务中多尺度目标检测精度低,以及在不同场景下检测鲁棒性和泛化能力差等问题。改进算法基于传统特征金字塔网络,提出一种自上而下和自下而上结合的双路径特征金字塔网络模块,通过特征拼接和融合操作保留预测特征层中更多的浅层和深层次语义信息。在此基础上,提出一种空间和通道机制串联的注意力网络模块,通过采用不降维的局部跨通道交互策略,进一步提升网络模型检测性能。经实验验证,改进算法相较于原始算法目标检测准确率提升4.6个百分点;小目标检测准确率提升11.76个百分点;中目标检测准确率提升5.78个百分点;大目标检测准确率提升3.7个百分点。

关键词: 计算机视觉, 目标检测, 深度学习, 特征金字塔网络, 注意力机制