SLKT:一种高精度图像语义预分割方法  

SLKT:A High Precision Image Semantic Pre-segmentation Method

作　　者：程柳 祁云嵩 刘杰 CHENG Liu;QI Yunsong;LIU Jie(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212100;Zhenjiang China-Australia Artificial Intelligence Research Institute,Zhenjiang 212002)

机构地区：[1]江苏科技大学计算机学院,镇江212100 [2]镇江中澳人工智能研究院,镇江212002

出　　处：《计算机与数字工程》2025年第2期523-527,550,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61471182)资助。

摘　　要：超像素聚类分割方法常作为许多深度学习模型的预处理方法,然而该方法结构的过简单化不仅使得图像预分割的精度降低,也会导致后续处理语义分割任务时模型的精度降低。为解决这一问题,论文在一般的超像素分割方法 SLIC的基础上设计了新的区域相似度度量及合并准,同时将融合了DPC算法的样本密度和三支聚类延迟决策的思想的K均值算法作为初分割后的再聚类方法。实验结果表明,新的相似性度量及合并准则能够充分考虑图像的空间特征,对SLIC处分割后的像素再聚类进一步提高了预分割的精度和实际分割效果。The superpixel cluster segmentation method is often used as a preprocessing method for many deep learning mod⁃els,but the oversimplification of the algorithm structure not only reduces the accuracy of image pre-segmentation but also reduces the accuracy of the model in subsequent semantic segmentation tasks.To deal with this problem,a new region similarity measure and merging criterion are designed based on the general superpixel segmentation method SLIC.At the same time,the K-means algo⁃rithm,which integrates the sample density of the DPC algorithm and the idea of a three-way cluster delay decision,is used as the re-clustering method after the initial segmentation.The experimental results show that the new similarity measure and merging crite⁃rion can fully consider the spatial characteristics of the image,and the reclustering of the segmented pixels at the SLIC can further improve the accuracy of pre-segmentation and the actual segmentation score.

关 键 词：超像素分割 样本密度 三支聚类 语义分割 空间特征 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]