结合动态循环金字塔与任务解耦的无锚框检测  

Anchor-free detection model based on dynamic cyclic pyramid and task decoupling

作　　者：孙一杰 李晓明[1] SUN Yi-jie;LI Xiao-ming(College 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,山西太原030024

出　　处：《计算机工程与设计》2025年第4期1157-1166,共10页Computer Engineering and Design

基　　金：国家自然科学基金项目(61373099)。

摘　　要：为解决现有无锚框网络缺乏精确的特征融合引导、解耦头获取任务特征不足,以及无锚框本身存在的边界框漂移的问题,提出一种结合循环特征融合与任务解耦的无锚框检测模型。设计动态循环特征金字塔,动态对齐并融合多尺度特征,以循环机制增强特征表达;提出新的任务解耦头,设计双维任务感知器获取任务特征,采用提出的任务一致性参数和Dynamic Varifocal损失函数完成任务对齐;在标签分配过程中,结合box重组算法,重新选取高质量的正负样本。在COCO数据集上,所提模型使mAP在ResNet50主干网络相对于baseline提升3.1%,在ResNet101上达到45.2%,检测性能优于其它先进的无锚框网络模型。To solve the problems of lack of precise feature fusion guidance,insufficient decoupling head to obtain task features,and bounding box drift in existing anchor-free networks,an anchor-free detection model combining cyclic feature fusion and task decoupling was proposed.A dynamic cyclic feature pyramid was designed to dynamically align and fuse multi-scale features,and a cyclic mechanism was used to enhance feature expression.A new task decoupling head was proposed.A dual dimensional task perceptron was designed to obtain task features,and the proposed task consistency parameters and Dynamic Varifocal loss function were used to complete task alignment.During the label assignment process,high-quality positive and negative samples were selected using the box recombination algorithm.On the COCO dataset,the proposed model improves the mAP by 3.1%on the resnet50 backbone network compared to the baseline,and reaches 45.2%on resnet101,with better detection performance compared to that of other advanced anchor-free network models.

关 键 词：目标检测 无锚框 特征融合 任务特征 多尺度特征 任务对齐 标签分配 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]