基于轻量化编码器与特征增强的人脸毛孔分割方法

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）03-0107-05

Facial Pore Segmentation Method Based on Lightweight Encoder and FeatureEnhancement

YANG Zhou， WANG Jianchao （School ofComputerand InformationEngineering，Shanghai Polytechnic University，Shanghai ，China）

Abstract：Toaddress challengesof structuralambiguity（similarto spots），sebum highlight interference，and multi-scale variations existing in facial pore segmentation，this paper proposes a method integrating a lightweight encoderand feature enhancement.This method employsaDepthwise Separable Convolution toconstruct alightweight backbone network，thereby reducing complexity.The SEchannel atention module is embedded to enhancepore feature response and suppressinterference. It introduces depth informationpredictionasanauxiliarytask，andenhances discriminativepowerbyleveragingpore depresion geometric features.Experimentsonaself-built high-resolution facial pore datasetdemonstrate thatthemethodachieves an IoUof0.872，Diceof0.921，precisionof0.896，andrecallof0.913，sigificantlyoutperforming CNNandUNetbaselines.It efectivelyreducesfalsepositivesofspotsandfalsenegativesofpores，providinganefcientandacuratesegmentatiotolfor skin analysis and cosmetic evaluation.

Keywords： pore segmentation; Depthwise Separable Convolution; SEchannel atention; depth information prediction; UNet

0 引言

目前在美容行业，图像分割、目标检测已被广泛用于研究面部特征和评估护肤品的使用功效，所使用的领域包括皮肤老化及皱纹测量[1、皮肤损伤[2]、皮肤病变检测分析[等。（剩余8792字）