多尺度注意力机制的双路人群计数网络

石祥滨; 吕浩杰

doi:10.3969/j.issn.2095-1248.2023.03.003

沈阳航空航天大学学报 ›› 2023, Vol. 40 ›› Issue (3) : 16 -27. DOI: 10.3969/j.issn.2095-1248.2023.03.003

信息科学与工程

多尺度注意力机制的双路人群计数网络

石祥滨 ,
吕浩杰

作者信息 +

Two-way crowd counting network with amulti-scale attention mechanism

Author information +

文章历史 +

PDF (2442K)

摘要

针对密集人群计数任务中人群尺度变化大、背景干扰以及特征融合导致的语义失调问题，提出了一种多尺度注意力机制的双路人群计数网络。网络主要由骨干网络、尺度增强模块、多尺度模块、上下文注意模块、注意力掩膜分支网络组成。首先，尺度增强模块通过捕捉不同尺度下的人群特征，并学习图像上每个特征的重要性，从而增强对尺度快速变化的适应性。其次，多尺度模块通过使用不同膨胀率的空洞卷积在保持原有特征图大小的前提下，对特征图进行多尺度变换，使得网络能够适应不同密度的人群场景。再次，上下文注意模块通过自适应地加权局部和全局上下文信息，实现了特征的融合与优化，以缓解不同级别特征存在的语义失调问题。最后，注意力掩膜分支网络通过生成与输入图像尺度相关的掩膜，降低背景干扰对网络性能的影响。通过这4个模块的相互配合，有效地提高了密集人群计数任务的准确性和稳定性，在多个数据集上的实验结果表明，该方法取得了较好的效果。

Abstract

A two-way crowd-counting network with a multiscale attention mechanism was proposed for the problem of semantic dissonance caused by large-scale variation of the crowd， background interference， and feature fusion in dense crowd-counting tasks.The network consisted of a backbone network， a scale enhancement module， a multi-scale module， a context attention module， and an attention mask branch network. Firstly， the scale enhancement module captured the crowd features at different scales and learns the importance of each feature on the image，the refore，adaptation to rapid scale changes arreenhanced. Secondly， the multi-scale module transformed the feature map at multiple scales while maintaining the original size of the feature map， so that different crowd densities can be adapted by network. Thirdly， the context attention module adaptively weighted local and global context information to optimize feature fusion and mitigate semantic inconsistency caused by features at different levels. Finally， the attention mask branch network generates masks related to the input image scale，the influence of background interference on network performance was reduced. The coordinated use of these four modules effectively，the accuracy and stability of dense crowd counting tasks are improved. Experimental results on multiple datasets demonstrate the effectiveness and feasibility of the proposed method.