基于改进的Swin Transformer的目标检测算法  

Target Detection Algorithm Based on Improved Swin Transformer

作　　者：辛晓明 杜春梅[1] 张振亚 XIN Xiaoming;DU Chunmei;ZHANG Zhenya(Hebei University of Architecture,Zhangjiakou,Hebei 075000,China)

机构地区：[1]河北建筑工程学院,河北张家口075000

出　　处：《长江信息通信》2024年第12期87-89,共3页Changjiang Information & Communications

基　　金：张家口市重点研发计划项目科技文化融合专项(2121015B);校级科研基金项目(Y201412)。

摘　　要：在目标检测任务中,基于Detection Transformer(DETR)的无锚框方法由于不需要依赖复杂的后处理步骤如非极大值抑制从而受到了广泛的关注。针对DETR使用的残差骨干网络ResNet(Residual Network)在提取全局信息能力上的不足,本文章提出一种基于改进的Swin Transformer的目标检测算法。该模型的骨干网络基于Swin Transformer改进,在这里使用了一种新的规范化方式,称为“后规范化”,这种新的方式会在整个网络产生更温和的激活值,然后将骨干网络与特征金字塔结合,获得不同尺度上的特征表示,从而能够更好地适应不同尺度的目标或图像变化。In the task of target detection,Detection Transformer(DETR)-based anchorless frame methods have received widespread attention because they do not rely on complex post-processing steps such as non-maximal value suppression.Aiming at the shortcomings of ResNet(Residual Network),the residual backbone network used in DETR,in terms of its ability to extract global information,this paper proposes a target detection algorithm based on the improved Swin Transformer.The backbone network of the model is improved based on Swin Transformer,where a new normalization method called“post-normalization”is used,which generates milder activation values throughout the network,and then the backbone network is combined with the feature pyramid to obtain feature representations at different scales,thus better adapt to target or image variations at different scales.

关 键 词：深度学习 目标检测 Swin Transformer 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]