A Crowd Counting Method Based on Multi-Scale Attention Network

Abstract

The task of crowd counting in computer vision is quite difficult, and a lot of excellent work has emerged recently. Recent research has suggested network algorithms that utilize both shallow feature information and deep information that the network has gathered. However, how to efficiently use these different scales of information is a question worth considering. We propose a multi-scale attention network (MSANet), which consists of a convolution module and a spatial attention module. The feature information of different scales extracted from the middle and deep layers is fused to obtain richer semantic features for crowd counting task detection. At the same time, considering that the fusion of different scale information will introduce some additional background information, we filter the redundant background information by introducing the attention module. Experiments show that this method has achieved good performance improvement on public datasets JHU + + and UCF-QNRF.