The　convolution　neural　networks　(CNNs)　can　extract　the　rich　feature　of　the　image.　It　was　widely　used　in　the　field　of　computer　vision　(CV)　and　made　great　breakthroughs.　However,　most　of　the　existing　CNNs　models　only　utilize　the　features　out　put　by　last　layer,　the　representation　of　features　is　not　comprehensive　enough.　In　this　paper,　we　propose　a　multilevel　features　fusion　method,　in　order　to　make　full　use　of　the　intermediate　layer　features.　This　method　can　strengthen　feature　propagation　and　improve　the　accuracy　of　downstream　tasks.　We　evaluate　our　method　through　experiments　on　two　image　classification　benchmark　tasks:　CIFAR-10　and　CIFAR-100.　The　experimental　results　show　that　our　method　is　able　to　significantly　improve　the　accuracy　of　VGG-like　model.　The　improved　model　is　better　than　most　existing　models.　©　Springer　Nature　Switzerland　AG　2018.