基于全局渲染的特征金字塔网络目标检测算法
Object Detection Algorithm Based on Global Rendering Feature Pyramid Network
何志鹏1
作者信息
- 1. 安徽理工大学 人工智能学院,安徽 淮南 232001
- 折叠
摘要
为了提高目标检测的准确率,同时优化计算资源的使用,提出了一种名为全局渲染特征的金字塔网络目标检测算法(GRFPN).在GRFPN中,结合了传统卷积方法和Transformer注意力机制的优势,引入了基于等距离分组的自注意力机制和全局特征渲染模块.GRFPN能够在全局和局部信息之间进行有效交互,同时保持计算效率.在COCO2017 数据集上的实验结果显示:与特征金字塔网络(FPN)相比,GRFPN实现了1.4AP的提升;和Feature Pyramid Transformer(FPT)相比,GRFPN只使用了62%运算量和80.79%的参数量就达到99.7%的效果.
Abstract
To enhance the accuracy of object detection and optimize the utilization of computing resources,a global rendering feature pyramid network object detection algorithm(GRFPN)is proposed.In GRFPN,combi-ning the advantages of traditional convolution method and Transformer attention mechanism,a self-attention mech-anism based on equidistant grouping and a global feature rendering module are introduced.GRFPN enables effec-tive interaction between global and local information while maintaining computational efficiency.The experimental results on the COCO2017 dataset demonstrate that GRFPN achieves 1.4 improvement in Average Precision(AP)compared to the Feature Pyramid Network(FPN).Compared with Feature Pyramid Transformer(FPT),GRFPN only uses 62%of the computation and 80.79%of the parameters to achieve 99.7%of the effect.
关键词
深度学习/目标检测/特征金字塔/注意力机制/TransformerKey words
deep learning/object detection/feature pyramid/attention mechanism/Transformer引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据