融合全局注意力和特征金字塔的小目标检测研究  

Research on Improved Object Detection Based on Global Attention Mechanism and AFPN

作　　者：邓佳慧 杜春梅[1] DENG Jia-hui;DU Chun-mei(Hebei University of Architecture,Zhangjiakou 075000,China)

机构地区：[1]河北建筑工程学院,河北张家口075000

出　　处：《电脑与电信》2024年第11期24-28,共5页Computer & Telecommunication

基　　金：2024年度河北省科协决策咨询调研课题基金(06-24-004-011)。

摘　　要：目标检测是计算机视觉领域的核心问题之一,针对小目标检测目标特征提取难、语义信息处理不精准等问题,在原模型的基础上加入了深度可分离卷积和全局注意力机制来优化网络结构,通过数据增强来提升目标检测算法的性能,使用渐近特征金字塔网络(AFPN)改进原有YOLOv8n的检测头,通过自适应空间融合和底层特征对齐,AFPN能够提升目标检测性能,尤其适用于处理矛盾信息。通过在CoCo数据集上进行大量实验验证,得到实验结果表明,所提出的改进措施在检测精度和实时性能方面取得了显著提升。Object detection is one of the core problems in the field of computer vision,in view of the difficulty of object feature extraction and the inaccurate semantic information processing of small target detection,this paper adds deep separable convolution and global attention mechanism on the basis of the original model to optimize the network structure,and strengthens data augmentation to improve the performance of object detection algorithm.The Asymptotic Feature Pyramid Network(AFPN)is used to improve the detection head of the original YOLOv8n,and the AFPN can improve the object detection performance through adaptive spatial fusion and underlying feature alignment,especially when dealing with contradictory information.We have carried out a large number of experimental verifications on the CoCo dataset.Experimental results show that the proposed improvement measures have achieved significant improvement in improving the detection accuracy and real-time performance.

关 键 词：小目标检测 全局注意力机制 YOLOv8n AFPN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]