通道-空间多尺度增强与双池化注意的表情识别网络
打开文本图片集
关键词：人脸表情识别；多尺度增强；双池化注意；通道-空间多尺度结构；高效通道-空间注意力机制中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2025）10-037-3182-10doi：10.19734/j. issn.1001-3695.2024.12.0524
Expression recognition network of channel-spatial multi-scale enhancement and dual-pooling attention
Liu Juan1†，Zhang Minyang1， HuMin2 ，Huang Zhong1，²，Jiang Julang1 （1.Schoolofon&elintcurinngalUsitn; KeyLaboratorffCuindcdelllfeecfaneUi of Technology，Hefei ，China）
Abstract：Aimingat theproblems that expressionfeature extractioninnatural scenesonlyfocusesonchannel-spatial singlescale information and average poling iseasytolose local saliencysemantics，this paper proposed an expresion recognition networkofchanel-spatialmulti-scaleenhancementanddual-polingattention.Firstly，tocapturethewholechannel-spatial multi-scale enhancement semantics，thispaperdesignedachannelsymmetriccascademulti-scale moduleandaspatialmultiscalefeatureextraction module，andconstructeda wholefeatureenhancementsubnetwork basedonthechannel-spatial multiscalestructure.Then，toepresentthechannel-spatialregiondual-poolingsalientsemantics，thispaperimprovedtheeficient localattentionmechanismitoaneficientchanel-spatialatetionmechanismandmbeddeditintotheregionfeatureaen tionsubnetwork.Finally，toobtainthepotentialcorelationbetweenthewholemulti-saleenancedsemanticsandtheregional dual-polingsalientsemantics，thispaperusedthecross-atentionmechanismtoperformthefeatureinteractionbetweenthe wholefeaturesandtheregionalfeatures，anddesignedthefeaturefusionsubnetworktocompletethemodel-levelfusionofthetwo typesoffeatures.TheexperimentalresultsshowthattheexpresionrecognitionratesonthefacialexpresiondatasetsRAF-DB and FERPlus reach 89.97% and 90.26% respectively，which are 13.54 and 10.95 percentage points higher than the baseline network.Compared withother networks，the proposed network hasbeter expressionrecognitionperformance innatural scenes.
KeyWords：facialexpressionrecognition；multi-scaleenhancement；dual-poolingatention；channel-spatialmulti-scale structure;efficientchannel-spatialattentionmechanism
0 引言
面部表情是传达人类情感最有效方法和手段之一，在远程教育、人机交互、医疗诊断和辅助驾驶等自然场景中受到广泛关注[1，2]。（剩余22523字）
