多层RepPoints表达特征的YOLOF改进模型研究  

Research on Enhanced YOLOF Model with Multilayer RepPoints Representation

作　　者：聂鹏 刘槟 罗凯 NIE Peng;LIU Bin;LUO Kai(School of Software&Internet of things Engineering,Jiangxi University of Finance and Economics,Nanchang 330013,China)

机构地区：[1]江西财经大学软件与物联网工程学院,江西南昌330013

出　　处：《电脑与信息技术》2025年第1期106-112,共7页Computer and Information Technology

基　　金：江西省高等学校教学改革研究项目(No.JXJG-19-4-18);江西财经大学教育教学改革研究项目(No.JG2023039)。

摘　　要：针对YOLOF模型存在的多尺度特征融合模糊与传统位置回归表达精度不足问题,提出了一种具备多层RepPoints表达特征的改进模型MLRP-YOLOF。该模型通过新的RepPoints分层解耦头将YOLOF单一多尺度融合特征图转换为分层多尺度RepPoints特征图,并在具备学习能力的RepPoints分层伪框转换器的作用下完成目标的回归任务与分类任务。实验结果表明,MLRP-YOLOF模型较YOLOF在mAP50-95、mAP50与mAP75指标上取得了2.18%~5.75%的相对性能优势。同时,对于结构更简单的Backbone,MLRP-YOLOF模型具有更显著的精度提升。Aiming at the problems of multi-scale feature fusion ambiguity and insufficient accuracy of traditional location representation in regression of YOLOF,an improved model named MLRP-YOLOF,which possesses the expressive characteristics of multi-layer RepPoints,was proposed.The model converted the singular multiscale fusion feature map of YOLOF into hierarchical multiscale RepPoints feature maps by a novel RepPoints hierarchical decoupled head,and performed the object classification task and regression task with a self-learning multilayer RepPoints pseudo box transformer.The experimental results show that MLRP-YOLOF achieves the outperformance from 2.18%to 5.75%in the metrics of mAP50-95,mAP50 and mAP75 compared with YOLOF.At the same time,for Backbone with simpler structure,MLRP-YOLOF model has more significant accuracy improvement.

关 键 词：目标检测 深度学习 YOLOF 多层RepPoints 可变形卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]