基于Transformer的多尺度物体检测  

Multi-Scale Object Detection Based on Transformer

作　　者：侯越千 张丽红[1] HOU Yueqian;ZHANG Lihong(College of Physical and Electronic Engineering,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学物理电子工程学院,山西太原030006

出　　处：《测试技术学报》2023年第4期342-347,共6页Journal of Test and Measurement Technology

基　　金：山西省研究生创新资助项目(2021Y154);山西省高等学校教学改革创新资助项目(J2021086)。

摘　　要：目前,Transformer基本模型对同一场景内不同尺寸物体的检测能力不足,其主要原因为各层等尺度的输入嵌入无法提取跨尺度特征,导致网络不具备在不同尺度的特征之间建立交互的能力。基于此,提出一种基于Transformer的多尺度物体检测网络,该网络采用跨尺度嵌入层初步对图像特征进行嵌入处理;利用多分支空洞卷积对输入进行下采样,通过调整并行分支的膨胀率使该结构具有多样的感受野;然后,由残差自注意力模块对输出嵌入结果进行处理,为特征图的局部和全局信息构建联系,使注意力计算融入有效的多尺度语义信息,最终实现多尺度物体检测。模型在COCO等数据集上进行训练,实验结果表明该方法与其他物体检测方法相比具有显著优势。The current Transformer basic model is inadequate for detecting objects of different sizes within the same scene.The main reason for this is that the equal-scale input embedding of each layer cannot extract cross-scale features,resulting in a network that does not have the ability to establish interactions between features of different scales.In this paper,we propose a Transformer-based multiscale object detection network,which uses cross-scale embedding layers to initially embed image features,in which the input is downsampled using multi-branch null convolution,and the structure is made to have diverse sensory fields by adjusting the expansion rate of parallel branches.The output embedding results are then processed by the residual self-attention module to construct links for local and global information of the feature map,so that the attention calculation incorporates effective multi-scale semantic information and finally achieves multi-scale object detection.The models are trained on datasets such as COCO,and the experimental results show that the method has significant advantages over other object detection methods.

关 键 词：物体检测 多尺度 TRANSFORMER 注意力机制 空洞卷积 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]