面向分割的局部分块与全局多尺度注意力机制  

Segmentation-oriented patch block and global multi-scale attention mechanism

作　　者：谭荆彬 赵旭俊[1] 苏慧娟 TAN Jing-bin;ZHAO Xu-jun;SU Hui-juan(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1141-1148,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61572343);山西省应用基础研究计划基金项目(20210302123223)。

摘　　要：现有的注意力机制仅增强特征图的通道或空间维度,未能充分捕捉细微视觉元素和多尺度特征变化。为解决此问题,提出一种基于局部分块与全局多尺度特征融合的注意力机制(patch and global multiscale attention,PGMA)。将特征图分割成多个小块,分别计算这些小块的注意力得分,增强对局部信息的感知能力。使用一组空洞卷积计算整个特征图的得分,获得全局多尺度信息的权衡。实验中,将PGMA集成到U-Net、DeepLab、SegNet等语义分割网络中,有效提升了它们的分割性能。这表明PGMA在增强CNN性能方面优于当前主流方法。Existing attention mechanisms typically enhance feature maps along either the channel or spatial dimensions,failing to fully capture fine visual elements and multiscale feature variations.To address this issue,an attention mechanism based on local patch division and global multiscale feature fusion was proposed,named patch and global multiscale attention(PGMA).The feature map was divided into multiple patches and the attention scores for each patch were calculated separately,thereby enhancing the ability to perceive local information.A set of dilated convolutions was employed to compute the scores across the entire feature map,achieving a balance of global multiscale information.In experiments,integrating PGMA into semantic segmentation networks such as U-Net,DeepLab,and SegNet effectively improves their segmentation performance.This demonstrates that PGMA outperforms current mainstream methods in enhancing CNN performance.

关 键 词：卷积神经网络 注意力机制 局部信息 分块策略 细节感知 全局多尺度信息 语义分割 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]