«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年06期

页码:

10-17

栏目:

媒体计算

出版日期:

2025-06-10

文章信息/Info

Title:

A Weakly Supervised Semantic Segmentation Method Based on Non-salient Region Enhancement

文章编号:

1673-629X(2025)06-0010-08

作者:

鲍凯辰1; 刘宁钟1; 张婧颖2

1. 南京航空航天大学 计算机科学与技术学院,江苏 南京 211106; 2. 江苏省青少年科技中心,江苏 南京 210019

Author(s):

BAO Kai-chen1; LIU Ning-zhong1; ZHANG Jing-ying2

1. School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China; 2. Jiangsu Youth Science and Technology Center,Nanjing 210019,China

关键词:

弱监督学习; 非显著区域; Transformer; 语义分割; 单阶段; 对比学习

Keywords:

weakly supervised learning; non-salient regions; Transformer; semantic segmentation; single-stage; contrastive learning

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0012

摘要:

弱监督语义分割技术能够使用分类标签生成像素级伪标签,再利用像素级伪标签训练语义分割网络,最终获得分割结果。 该技术在很大程度上缓解了因语义分割标签难以大量获取而带来的算法瓶颈。 目前的弱监督语义分割模型一般是基于分类网络的,而分类网络往往只关注图像的显著区域。 这种现象会使目标边界等非显著区域无法被充分激活,进而导致最终的语义分割结果不够准确。 为了解决这一问题,该文提出了一种基于非显著区域增强的端到端弱监督语义分割模型。 该模型引入了两个重要模块:针对非显著区域的对比学习模块和基于比例的全局池化模块。 前者通过选择不确定区域、分配正负标签以及对比学习的策略,对于不确定区域进行筛选和鉴别,充分激活了目标的非显著区域,获得了更为完整的像素级伪标签。 而后者则确保了模型的注意力更加平衡,而不是过度集中在少数显著的离群值上。 在 PASCAL VOC2012 数据集上所做的实验显示,该模型有效改善了非显著区域的欠激活问题,提高了语义分割的准确性。

Abstract:

Weakly supervised semantic segmentation can generate pixel-level pseudo labels using image-level labels,and then use these pseudo labels to train a semantic segmentation network,ultimately obtaining segmentation results. This approach significantly alleviates the algorithmic bottleneck caused by the difficulty of acquiring pixel - level semantic segmentation labels. Current weakly supervised semantic segmentation models are generally based on classification networks,while classification networks tend to only focus on the salient regions of an image. This phenomenon may cause the non-salient regions such as object boundaries to not be fully activated, thereby resulting in inaccurate final semantic segmentation results. To address this problem,we propose an end-to-end weakly supervised semantic segmentation model based on enhanced attention. The model introduces two important modules:a contrastive learning module for non-salient regions and a global pooling module based on ratio. The former adopts a strategy of selecting uncertain regions,assigning labels,and leveraging contrastive learning to screen and identify uncertain regions. This approach fully activates the non-salient regions of the target,resulting in more comprehensive pixel-level pseudo labels. The latter ensures that the model's attention is more balanced,rather than being overly concentrated on a few salient outliers. Experiments conducted on the PASCAL VOC2012 dataset demonstrate that the proposed model effectively improves such issue,leading to increased accuracy in semantic segmentation results.

常用功能

更新日期/Last Update: 2025-06-10