摘要
针对网络中注意力通常处于多通道状态,更新却依赖单通道标签导致优化困难的问题,提出了一种间接—即时注意力优化(IIAO)模块。基于SoftMax-Attention策略,将多通道注意力在数学意义上转化为单通道密度图,同时自动为特征金字塔模组提供自适应多尺度融合服务。考虑到转化方式的特殊性,设计了区域相关性损失函数(RCLoss)来检索连续易错区域,平滑空间信息。实验结果表明,所提算法有效且性能更加稳定。
An indirect-instant attention optimization(IIAO)module was proposed for the problem that attention in the network is usually in a multi-channel state,while its update relies on a single channel label.It is based on softmax-attention strategy to transform multi-channel attention into single-channel density map in a mathematical sense,while automatically providing adaptive multi-scale fusion service for feature pyramid modules.Considering the specificity of the transformation method,a region correlation loss function(RCLoss)is designed to retrieve continuous error-prone regions and smooth spatial information.Experimental results show that the proposed algorithm is effective and more stable.
作者
韩素玉
王国栋
王永
刘瑞
HAN Su-yu;WANG Guo-dong;WANG Yong;LIU Rui(College of Computer Science and Technology,Qingdao University,Qingdao 266071,China;Songli Holding Group Co.,Ltd.,Qingdao 266073,China)
出处
《青岛大学学报（自然科学版）》
CAS
2023年第2期50-57,共8页
Journal of Qingdao University(Natural Science Edition)
基金
山东省自然科学基金(批准号:ZR2019MF050)资助
山东省高等学校优秀青年创新团队支持计划(批准号:2020KJN011)资助。
