引用本文
  • 李呓瑾,李少龙,贺彦,刘炜.基于特征融合注意力的小样本语义分割算法[J].广西科学,2023,30(5):951-960.    [点击复制]
  • LI Yijin,LI Shaolong,HE Yan,LIU Wei.Few-shot Semantic Segmentation Based on Feature Fusion Attention Mechanism[J].Guangxi Sciences,2023,30(5):951-960.   [点击复制]
【打印本页】 【在线阅读全文】【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 49次   下载 98 本文二维码信息
码上扫一扫!
基于特征融合注意力的小样本语义分割算法
李呓瑾1, 李少龙1, 贺彦2, 刘炜3
0
(1.云南电网有限责任公司信息中心, 云南昆明 650200;2.北京国科恒通科技股份有限公司, 北京 100085;3.清华大学电机工程与应用电子技术系, 北京 100084)
摘要:
针对小样本语义分割任务中对查询图片的信息利用不充分的问题,提出一种基于特征融合注意力的小样本语义分割算法。首先,利用共享主干网络编码支持图片和查询图片,从而获取图片的深度特征;然后,利用注意力机制获取支持特征和查询特征的强关联语义信息,从而构造任务注意力特征图;最后,提出一种多特征注意力融合模块,它能够自适应融合多种特征的深层语义信息并进行特征解码,从而获取目标物体的分割掩码。在PASCAL-5i和COCO-20i公开数据集进行了实验,结果表明,所提出模型比当前主流的小样本语义分割模型在1-way 1-shot和1-way 5-shot任务中分割得更加精准,尤其是在更具有挑战性的COCO-20i数据集上,所提出模型在1-shot的设定下达到了28.8%的mIoU和62.1%的FB-IoU,在5-shot设定下达到了36.9%的mIoU和64.8%的FB-IoU。
关键词:  小样本语义分割  多特征融合  注意力机制  深层语义信息  分割掩码
DOI:10.13656/j.cnki.gxkx.20231121.014
投稿时间:2022-10-15修订日期:2022-12-07
基金项目:国家自然科学基金项目(62222209)资助。
Few-shot Semantic Segmentation Based on Feature Fusion Attention Mechanism
LI Yijin1, LI Shaolong1, HE Yan2, LIU Wei3
(1.Information Center of Yunnan Power Grid Co., Ltd., Kunming, Yunnan, 650200, China;2.Beijing THPower Technology Co., Ltd., Beijing, 100085, China;3.Department of Electrical Engineering and Applied Electronic Technology, Tsinghua University, Beijing, 100084, China)
Abstract:
Aiming at the problem of insufficient information utilization in query images for small sample semantic segmentation tasks,a few-shot semantic segmentation algorithm based on feature fusion attention is proposed.Firstly,it utilizes shared backbone networks to obtain deep features of both image and query images.Secondly,attention mechanisms are employed to capture strong semantic correlation information between support features and query features,constructing task attention feature maps.Finally,a multi-feature attention fusion module is proposed,which can adaptively fuse multiple features' deep semantic information and perform feature decoding,thereby obtaining target object segmentation masks.The proposed model is evaluated on PASCAL-5i and COCO-20i datasets,and experimental results show that the proposed model outperforms current mainstream small sample semantic segmentation models in terms of more precise segmentation in both 1-way 1-shot and 1-way 5-shot tasks.Especially on the more challenging COCO-20i dataset,the proposed model achieves 28.8% mIoU and 62.1% FB-IoU under the setting of 1-shot,and 36.9% mIoU and 64.8% FB-IoU under the setting of 5-shot.
Key words:  few-shot semantic segmentation  multi-feature fusion  attention mechanism  deep semantic information  segmentation mask

用微信扫一扫

用微信扫一扫