基于注意力机制的多级监督人群计数算法

Multi-level Supervised Crowd Counting Network Based on Attention Mechanism

针对人群计数问题,提出了一种基于卷积神经网络的人群计数网络,该网络由主干网和多级监督分支结构组成,在主干网络的多个阶段引入注意力机制学习不同尺度的人群特征。算法采用VGG16模型的前13层作为主干网,并且加入膨胀卷积网络结构,融合图像中的多尺度人群特征,解决多尺度人群计数问题,从而生成高质量的密度图。同时,在3个不同尺度的分支结构中引入注意力机制,在损失函数中加入不同尺度的注意力损失,从而使整个网络聚焦图像中的人群区域。算法在4个主要的数据集上进行了测试,算法结果优于最近其他的方法。

In order to solve the crowd counting problem,a new crowd counting network based on convolutional neural network is proposed.The network is composed of the backbone network and the multi-level supervised branch structures.Attention mechanism is introduced in multiple stages of the backbone network to learn crowd features of different scales.In this method,the first thirteen layers of the VGG16 model are used as the backbone network,and the dilated convolution layer is added to fuse the multi-scale crowd features in the image to solve the problem of multi-scale crowd counting,so as to generate a high-quality density map.At the same time,attention mechanism is introduced into three branch structures of different scales,and attention loss with different scales is added into the loss function in the training process,so that the whole network can focus on the crowd area in the image.The method is tested on four main datasets,and the results are better than other recent methods.