As　the　convolutional　neural　network　is　widely　used　in　digital　image　processing　,　natural　language　processing,　etc,　it　is　facing　more　and　more　challenges.　The　convolutional　neural　network　is　designed　to　be　more　complex,　and　the　amount　of　data　required　to　train　convolutional　neural　network　has　also　grown　substantially,　which　leads　to　higher　time　cost　and　hardware　requirements　for　training　convolutional　neural　network.　This　paper　puts　forward　a　new　solution　in　order　to　solve　this　problem　.　In　this　paper,　the　whole　image　classification　task　is　regarded　as　a　competition,　each　feature　in　the　image　is　regarded　as　an　individual　contestant,　the　image　dimensionality　reduction　technology　used　in　the　model　is　regarded　as　a　preliminary　contest,　and　the　convolutional　neural　network　in　the　model　is　regarded　as　a　final.　The　image　processed　by　dimensionality　reduction　technology　is　input　to　CNN(Convolutional　Neural　Network).　In　order　to　verify　the　idea　of　this　paper,　we　design　two　convolutional　neural　network　architectures　:　model　1　and　model　2.　From　the　experimental　results　in　this　paper,　it　can　be　seen　that　the　accuracy　of　the　model　1　training　dataset　decreases　with　the　increase　in　dimensionality　reduction　and　approaches　to　86.508%　of　the　accuracy　of　training　dataset　without　dimensionality　reduction;　the　accuracy　of　model　2　training　dataset　is　not　sensitive　to　the　selection　of　dimensionality　reduction　range　and　dimensionality　reduction　method　and　fluctuates　within　7.126%　as　a　whole;　the　accuracy　and　time　cost　of　the　test　dataset　of　model　1　and　model　2　are　greatly　affected　by　the　dimensionality　reduction　method　and　dimensionality　reduction　range.The　final　experimental　results　show　that　the　loss　of　accuracy　can　be　reduced　to　an　acceptable　range,　and　the　dimensionality　reduction　operation　is　very　meaningful　for　the　reduction　of　time　cost.　©　2021　ACM.