基于卷积神经网络的视觉算法模型的压缩和加速.pdfVIP

Abstract

Inrecentyears,convolutionalneuralnetworktechnologyhasdevelopedrapidly,

anditsexcellentperformancehasbeenwidelystudiedandverifiedinmanyfieldsof

computervision,andhasachievedexcellentresults.However,convolutionalneural

networksalsohaveshortcomingssuchaslargeparametersandcalculations,andlarge

energyconsumption,whichlimittheirpopularizationandapplicationinactual

industries.Therefore,modelcompressionandaccelerationhavebecomeanimportant

researchdirectionnowadays.Thisarticlefocusesbothonhigh-levelvisiontaskand

low-levelvisiontaskincomputervision.Fortheclassificationproblem,wepropose

twostructuredpruningmethods(ACOandALI)tocompressthemodel.Forthesuper-

resolutionproblemoflow-levelvisiontask,weproposeanincrementallinear

networkquantizationmethodtoacceleratethemodel.Weconductexperimental

comparisonsonmultipledatasetsandnetworkstructurestoprovetheeffectivenessof

ourmethod.Specifically,theinnovationsandcontributionsofourthreemethods

proposedinthisarticleareasfollow:

1.Whenevaluatingtheimportanceofconvolutionkernels,thecurrentmainstream

pruningmethodsareusuallybasedonacertaindimensionfeatureoftheconvolution

kernel,suchasL-norm,Euclideandistance,Taylorexpansionseries,etc.Thesesingle-

1

dimensionalevaluationmethodstendtobringincompleteconsideration,largepruning

errorandotherissues.Tosolvetheseproblems,weproposeapruningmethodbasedon

antcolonyoptimizationalgorithm(ACO)whichconsidersthecorrelation,similarity

betweendifferentnodesandtherelativesizeoftheabsolutevalueoftheconvolution

kernelatthesametime.Ourmethodevaluatestheconvolutionkernelcomprehensively.

Atthesamet