基于改进UNet的图像边缘细节分割算法

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2096-4706（2026）04-0099-07

An Image Edge Detail Segmentation Algorithm Based on Improved UNet

LYU Feiyu， YAN Jianhong （College ofComputer Scienceand Technology，TaiyuanNormal University，JinzhongO3o619，China）

Abstract：Aiming atthe problem that the traditional UNetloses edge information via gradual poling during downsampling andfails toaccuratelysegment imageedge details，asegmentationalgorithm integrated withanedge detail enhancement mechanism is proposed.This algorithm improves the originalUNet structure by introducing aFeature Extraction Module in the encoderstage，whichconsistsofamuli-scaleedgepixel difference moduleandaconvolutionblock.Themulti-scale edge pixel diferencemoduleextractsedgeinformationatdiferentsalestoreducepooling-inducedinfomationlossandenhanceitsedge segmentationacuracy.Meanwhile，theconvolutionblockextractsglobalinformationthroughtwoconvolutionallayers，fusesit with edge information，and thus improves the segmentationaccuracyof the network.Experimentsareconductedonthe public datasets MSD Spleen and SLIVER07.The results show that the proposed model's Dice coefficients reach 94.2% and 96.4% ，and Intersection over Union （IoU） values reach 89.2% and 92.7% respectively， proving its effectiveness.

Keywords： image segmentation; multi-scale; edge detail; Convolutional Neural Network

0 引言

传统图像分割方法例如区域增长法、基于聚类的方法[1和基于边缘的方法[2在分割目标区域时，不仅易受噪声、光照变化、低对比度等图像质量问题影响，降低分割精度，而且在处理目标形态不规则、边界模糊等复杂场景时，仅依赖像素级特征（如灰度、颜色、局部纹理）进行分割，未考虑像素间的全局空间关联或高层语义信息，无法对图像局部细节与全局上下文信息进行有效的提取，难以精准地分割出目标区域的轮廓，导致分割精度下降。（剩余9087字）