融合多尺度注意力的太赫兹图像目标检测研究

Research on Target Detection in Terahertz Images with Multi-scale Attention

基于太赫兹技术的安检设备具有独特的无损伤检测特性,在安防领域具有良好的发展前景.太赫兹图像分辨率低,特征信息有限,且高亮区域多,目标与背景亮度相近,为太赫兹图像中的目标检测带来了一定的困难.针对此,本文提出一种融合多尺度注意力的目标检测框架MSAD-SSD,在SSD算法的基础上,采用ResNet101作为特征提取网络,并设计一种多尺度注意力模块.该模块采用多支路膨胀卷积进行多尺度特征提取,扩展网络的感受野,充分获取目标的上下文信息;通过注意力机制对特征图的通道特征进行权重分配,增强有用特征,抑制无关信息的干扰;最后对多个分支进行特征融合.实验结果表明,所提MASD-SSD在太赫兹图像数据集中达到82.63%的平均检测精度,能够有效提高目标的检测精度.

Security inspection equipment based on terahertz technology has unique non-destructive detection characteristics and has a good development prospect in the security field.Terahertz images have low resolution,limited feature information,and many highlight areas.The brightness of the target and the background are similar,which brings certain difficulties to target detection in the terahertz image.In response to this,this paper proposes a target detection framework MSAD-SSD that integrates multi-scale attention.Based on the SSD algorithm,ResNet101 is used as the feature extraction network,and a multi-scale attention module is designed.This module uses multi-branch dilated convolution for multi-scale feature extraction,expands the receptive field of the network,and fully obtains the context information of the target;then uses the attention mechanism to weight the channel features of the feature map to enhance useful features and suppress irrelevant information The interference;finally feature fusion is performed for multiple branches.Experimental results show that the proposed MASD-SSD achieves an average detection accuracy of 82.63%in the terahertz image data set,which can effectively improve the detection accuracy of the target.