面向可重构结构的CNN模型混合压缩方法

打开文本图片集

DOI：10.16652/j.issn.1004-373x.2026.01.026 引用格式：，，，等.面向可重构结构的CNN模型混合压缩方法[J].现代电子技术，2026，49（1）：167-173.

Hybrid compression method for CNNtargeting reconfigurablearchitectures

LIUPengfei¹，JIANGLin²，LIYuancheng²，WUHai³ （1.CollegeofElectricalandControlEnginering，Xi'an UniversityofScienceand Technology，Xi’an7o6oo，China; 2.CollgeofScienceandTechnology，Xi'anUniversityofScienceandTechnology，Xi'an71O6oo，China; 3.CollegeofCommunicationandInformationTechnology，Xi'anUnversityfSienceandTechnology，Xi’an6oChina）

Abstract：Asthescaleofconvolutionalneural network（CNN）continues toexpand，thenumberof parametersand computationburdenhaveincreasedsignificantly，whichleadstoseverememoryaccessbotlenecksandlimitscomputational effciency.Inviewofthis，anovelhyridcompressionmethodforCNNtargetingreconfigurablerchitecturesisproposedInthis method，apruning-tquantizationsategymployigfilterpruningbasedonfistorderayloexpansion，fulloedlar （FClayer）weightpruning basedonthresholdvalues，andmixed-precisionadaptivequantizationaproach，isproposed.These techniqusstrive toreducetheparameterquantityandcomputationalcomplexityof themodel，andthemethodisdeployedona self-developedreconfigurableprocessor.Experimentalresultsdemonstratethattheproposedmethodachievescompressonratios of 31.4 timesand7.9 times ontheVGG16 modeland ResNet18 model，respectively，withaccuracy lossesof only 1.20% and （20 0.74% ，respectively.On thereconfigurable arrayprocessorbuiltonthe VirtexUltraScale VU440 FPGA，theexecutioncycleof the compressed VGG16 model is reduced by up to 62.7% .It is proved that the proposed method is suitable for edge computing devices with limited resources.

Keywords：CNN；model compresion；structured pruning；adaptivequantization;；paralel computing;；reconfigurable architecture

0 引言

近年来，卷积神经网络（ConvolutionalNeuralNetwork，CNN）在各个领域取得重大进展，包括图像分类、目标检测2与自然语言处理3等。（剩余10204字）