RPPM-Net：基于多尺度特征并行融合的图像识别方法

打开文本图片集

中图分类号：TN911.73-34；TP391.4 文献标识码：A 文章编号：1004-373X（2025）15-0071-07

RPPM-Net： Image recognition method based on multi-scale featureparallel fusion

LIUSiyuan，ZHANGWendong，LURun，NIUSen，MAMengnan （School ofSoftware，XinjiangUniversity，Urumqi83oooo，China）

Abstract：The traditional deep neural networkshavesome deficiencies intheimagerecognitiontasks，forinstance，failure tofocusonkygiosndatusielyfatiogadatiouetallssoyeldndisuntalio abilityof themodel，soanimagerecognitionnetworknamedresidualparalelpyramidmulti-net（RPPM-Net）whichisbasedon multi-scaleparalelfusionisproposed.Aparallattentionnetworkisdesignedtocapturediversefeatureinformationbyahybrd atentionmechanismwithparalelmultipleheads，whichenhancestheetwork'sabilitytofocusonimportantregionsandkeyfeatures intherecognitiontaskswhilemaintainingthemodeleficiencysoastostrengthenitsatentiontokeyfeatures.Amultiscalepyramidalconvolutionalatentionmoduleisdesignedtoavoidinformationdegradationduetothesmallsensoryfieldby effectivelyintegratingulti-scalefeaturesatdiferentlayers.Theenhancdfullyconnectedlayeandregularizationtechiqueare adoptedtoeffctivelyllviatetheoverfitingndimprovethegeneralizationabilityofthemodelondiferentdatasetsTheexperimental results show that the accuracy rate of RPPM-Net on CIFAR-10，CIFAR-100，and Caltech-256 datasets reaches 97.02% 85.04% and 89.69% ，respectively，which can combine structural regularization and feature fusion fully and enhance the network performance effectivelywithout significantlyincreasing the computational cost.

Keywords：imagerecognition;convolutionalneural network；multi-scalefeature;parallelatention;pyramidalconvolution; feature extraction

0 引言

图像识别是计算机视觉领域中的重要研究方向，其主要任务是从原始图像数据中提取特征，并基于这些特征对图像内容进行识别和分析。（剩余10273字）