基于注意力机制的多任务目标计数系统设计
Design of Multi-Task Object Counting System Based on Attention Mechanism
李永慧1
作者信息
- 1. 山西科技学院 光机电工程学院,山西 晋城 048000
- 折叠
摘要
提出基于注意力机制的深层神经网络用于目标计数,其任务是为输入图像的目标进行精确数目统计.该网络模型同时引进多任务学习方法,多尺度融合得到密度特征图和注意力特征图进行目标计数.首先,使用交叉特征金字塔网络进行特征提取;其次,将提取的特征分别用于密度特征图及注意力特征图进行交叉融合;最后,通过多任务学习将两个输出特征图逐元素运算,得到精确的密度特征图.提出的网络模型在行人检测数据集(ShanghaiTech)与多类别的行为识别数据集(UCF_CC_50)上进行了训练与测试,实验结果表明,通过在各个分支引入注意力机制,可以有效提高整个模型预测结果的准确率.
Abstract
This article proposes a deep neural network based on the attention mechanism for object counting,whose task is to accurately count the number of targets in the input image.The network model simultaneously introduces a multi-task learning approach,fusing multi-scale features to obtain density and attention feature maps for object counting.Firstly,a cross-feature pyramid network is used for feature extraction.Then,the extracted features are used for density and attention feature maps,which are cross-fused.Finally,through multi-task learning,the two output feature maps are element-wise operated to obtain an accurate density feature map.The network model of this paper is trained and tested on pedestrian detection datasets(ShanghaiTech)and multi-category action recognition datasets(UCF_CC_50).The experimental results show that introducing the attention mechanism in each branch can effectively improve the accuracy of the entire model's predictive results.
关键词
目标计数/注意力机制/多任务学习/交叉特征金字塔网络Key words
object counting/attention mechanism/multi-task learning/spatial pyramid pooling fast cross stage partial connection引用本文复制引用
出版年
2024
NETL
NSTL
万方数据