基于扩散方法的特征动态库  

Feature Dynamic Library Based on Diffusion Method

作　　者：黄山 范慧杰[2,3] 林森 曹镜涵[1,2,3] 唐延东 HUANG Shan;FAN Huijie;LIN Sen;CAO Jinghan;TANG Yandong(School of Automation and Electrical Engineering,Shenyang Ligong University,Shenyang 110168,China;State Key Laboratory of Robotics,Shenyang Institute ofAutomation ChineseAcademy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,沈阳110168 [2]中国科学院沈阳自动化研究所机器人学国家重点实验室,沈阳110016 [3]中国科学院机器人与智能制造创新研究院,沈阳110169

出　　处：《计算机工程与应用》2025年第5期241-249,共9页Computer Engineering and Applications

基　　金：国家自然科学基金(U20A20200,62273339);辽宁省自然科学基金(2022-YGJC-66);中国科学院青年创新促进会项目(2019203)。

摘　　要：弱监督语义分割(weakly-supervised semantic segmentation,WSSS)利用更简单易得的图像级标注(框、点、线和类别信息),以训练分类模型的方式获取物体分割区域,从而实现图像像素级预测。由于主流的分割算法仍然需要依赖大量的标注数据,而逐像素标注十分昂贵、费时且费力,为了减轻这一负担,人们越来越关注高效利用弱标注的信息生成标签这种方法。在弱监督方面关注两个问题:一是生成的伪标签图像不够完整,二是分类不够精准。现有的绝大部分方法都只考虑了第一点,而提出的基于扩散方法的特征动态存储方案既保证了标签的完整性又兼顾了类别特征,通过抑制最显著区域以扩散激活图的激活区域方式,并使用前景背景分离的细化方案剔除类激活图的冗余噪声;最后输入特征动态库在训练阶段引入语义标签使得每个类别实时更新。在Pascal VOC 2012数据集上,提出的方法对比现有大部分弱监督语义分割方法,均展现了优越性;对比以Transformer为骨干网络的新方法,在验证集和测试集上分别高2.5个百分点和2.9个百分点。Weakly supervised semantic segmentation(WSSS)is a technique that utilizes simpler and more easily obtainable image level annotations(boxes,points,lines,and category information)to train classification models to obtain object segmentation regions and achieve pixel level prediction of images.Since mainstream segmentation algorithms still rely on a large amount of annotated data,pixel-by-pixel annotation is expensive,time-consuming,and labor-intensive.In order to alleviate this burden,people are increasingly paying attention to the efficient use of weakly labeled information to generate labels.In terms of weak supervision,there are two issues to focus on:first,the generated pseudo label images are not complete enough;second,the classification is not accurate enough.The vast majority of existing methods only consider the first point,while the paper proposes a feature dynamic storage scheme based on diffusion method,which not only ensures the integrity of labels but also takes into account category features.Specifically,this paper suppresses the most prominent regions to diffuse the activation regions of the activation map,and uses a refinement scheme of foreground background separation to remove redundant noise from the class activation map.Finally,the input feature dynamic library introduces semantic labels during the training phase to update each category in real-time.Compared with most existing weakly supervised semantic segmentation methods on the Pascal VOC 2012 dataset,the method proposed in this paper shows superiority.Compared with the new method using Transformers as the backbone network,this method is 2.5 percentage points and 2.9 percentage points higher on the validation and test sets,respectively.

关 键 词：弱监督 特征提取 语义分割 TRANSFORMER 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]