基于多尺度及双注意力机制的小尺寸人群计数

Small size crowd counting based on multi-scale and dual attention mechanism

本文针对背景干扰、特征信息不足以及尺度剧烈变化等问题,提出了一种基于多尺度及双注意力机制(Multi-Scale and Dual Attention,MSDA)的小尺寸人群计数网络。MSDA网络主要由空间-通道双注意力(Spatial Channel-dual Attention,SCA)模块和多尺度特征融合(Multi-scale Feature Fusion,MFF)模块构成。MFF模块将特征送入三列拥有不同卷积核的膨胀卷积来扩大小目标的空间尺度,再通过特征级联及卷积操作进行多尺度特征融合;SCA模块把特征送入通道注意力网络,使用空间注意力中的池化操作及逐像素相乘操作加强细节信息;最后将处理好的特征送入密度图生成模块,通过1×1卷积获得密度图。在Mall数据集和Shanghaitech数据集上进行了测试,取得了较好的准确率与鲁棒性。

Aiming at the problems of background interference,insufficient feature information,and dramatic changes in scale,This paper proposed a small-scale crowd counting network based on multi-scale and dual attention mechanism(Multi-Scale and Dual Attention,MSDA). The MSDA network was mainly composed of a Spatial Channel-dual Attention(SCA)module and a Multi-scale Feature Fusion(MFF)module. The MFF module sent the features into three columns of dilated convolutions with different convolution kernels to expand the spatial scale of small targets,and then performs multi-scale feature fusion through feature cascade and convolution operations;the SCA module input the features into the channel attention network,Then use the pooling operation in the spatial attention,and use the pixel-by-pixel multiplication operation to enhance the detailed information;Finally,send the processed features to the density map generation module,and obtain the density map through1×1 convolution.This paper tested the proposed model on Shanghai Tech and Mall datasets,and the results show that the model achieve good accuracy and robustness.