多尺度局部聚类的Kmeans-DETR目标检测方法  

Multi-scale Local Clustering Method for Object Detection Based on Kmeans-DETR

作　　者：崔鹏 杨海峰[1] 蔡江辉[1] 王玉鹏 CUI Peng;YANG Haifeng;CAI Jianghui;WANG Yupeng(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,太原030024

出　　处：《小型微型计算机系统》2024年第5期1136-1142,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(U1931209)资助。

摘　　要：在利用DETR进行目标检测时,复杂的矩阵运算不仅对稀疏冗余特征产生了大量无效计算,还阻碍了对图像更多尺度信息的使用.针对上述问题,本文提出了多尺度局部聚类的Kmeans-DETR目标检测方法.首先构造了局部Kmeans聚类方法,通过在特征图的局部区域内聚类得到对应簇,并选取特征代表该簇以降低稀疏冗余特征的数量,进而减少矩阵计算量与模型复杂度;其次通过3种尺度的局部聚类,引入多尺度信息的同时通过不同尺度聚类区域重叠的方式解决局部信息不互通的问题;最后改进了位置编码方式用以记录局部聚类后特征的位置信息,并嵌入到簇的代表特征中,利用Transformer结构完成检测任务.本文提出的模型在COCO数据集上与主流的目标检测模型进行了对比,在多个指标上均有较好的表现.When using DETR for object detection,complex matrix operations not only generate a lot of invalid calculations for sparse redundant features,but also hinder the use of more scale information of the image.To solve the above problems,this paper proposes a multi-scale local clustering method for object detection based on Kmeans-DETR.Firstly,the local Kmeans clustering method is constructed.The corresponding cluster is obtained by clustering in the local area of the feature map,and the feature is selected to represent the cluster to reduce the number of sparse and redundant features,thereby reducing the matrix calculation and model complexity;Secondly,through three scales of local clustering,multi-scale information is introduced and the problem of local information not communicating is solved by overlapping clustering regions of different scales;Finally,the location coding method is improved to record the location information of the features after local clustering,which is embedded into the representative features of the cluster,and the Transformer structure is used to complete the detection task.The model proposed in this paper is compared with the mainstream object detection model on the COCO dataset,and has good performance on multiple indicators.

关 键 词：目标检测 Kmeans DETR 多尺度 TRANSFORMER 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]