Convolutional　neural　networks　(CNN)　have　made　significant　breakthroughs　in　image　feature　extraction.　A　variety　of　CNN　architectures　have　been　proposed　and　then　continued　to　be　improved.　However,　the　representation　of　image　features　by　a　single　CNN　structure　is　not　comprehensive　enough.　For　example,　shallow　CNN　extraction　is　more　general　and　less　sensitive,　vice　versa.　In　this　paper,　we　propose　an　ensemble　method　using　LSTM　to　obtain　image　features　that　represent　the　image　more　comprehensive.　We　use　the　output　of　a　single　CNN　model　as　an　input　for　LSTM　for　a　moment　to　ensemble.　We　evaluated　our　approach　using　VGG　and　other　models　on　the　Cifar10　and　Cifar100　datasets.　The　accuracy　of　classification　using　ensemble　features　is　significantly　higher　than　that　of　a　single　model.