As　the　important　research　achievement,　deep　convolutional　neural　networks　have　been　widely　applied　to　various　fields　such　as　computer　vision,　natural　language　processing,　information　retrieval,　speech　recognition,　semantic　understanding,　and　have　attracted　a　wave　of　neural　networks　research　from　both　academia　and　industry　and　have　contributed　to　the　development　of　artificial　intelligence.　The　convolutional　neural　networks　directly　treat　the　original　data　as　input,　automatically　learn　the　feature　representations　from　a　large　number　of　training　data.　The　convolutional　neural　networks　have　the　characteristics　of　local　connection,　weight　sharing　and　pooling　operation,　which　can　effectively　decrease　the　network　complexity　and　reduce　the　number　of　training　parameters,　so　that　the　model　has　some　certain　invariance　to　translation,　distortion　and　scale.　Currently,　many　approaches　of　deep　neural　networks,　including　the　increase　of　size　and　complexity　of　neural　networks,　the　use　of　larger　sets　of　training　data,　the　improvement　of　neural　network　architecture　and　training　methods,　etc.,　have　been　proposed　to　simulate　the　complex　hierarchical　cognitive　attributes　of　human　brain　and　pull　close　the　gap　between　the　human　brain　and　visual　system,　so　that　the　machine　has　the　capability　to　capture　＇abstraction　concepts＇.　The　deep　convolutional　neural　networks　have　been　a　great　success　in　many　computer　vision　tasks,　such　as　image　classification,　object　detection,　face　recognition,　and　person　re-identification.　In　this　paper,　we　first　review　the　history　of　the　development　of　convolutional　neural　networks,　and　briefly　introduce　M-P　neuron　model,　Hubel-Wiesel　model,　Neocognitron,　LeNet　for　handwriting　recognition,　and　deep　convolutional　neural　network　for　image　classification　in　the　ImageNet　competition.　Then　we　have　a　detailed　analysis　of　the　fundamental　principle　of　deep　convolutional　neural　networks,　and　introduce　the　mathematical　representation　and　the　respective　functions　of　the　convolution　layer,　the　pooling　layer　and　the　fully　connected　layer.　Besides,　this　paper　focuses　on　the　representative　works　of　convolutional　neural　networks　on　the　following　three　aspects,　and　demonstrates　various　technical　methods　in　improving　the　accuracy　of　image　classification　using　examples.　In　the　aspect　of　increasing　the　number　of　neural　networks＇　layers,　the　architectures　of　classical　convolutional　neural　networks　such　as　AlexNet,　ZF-Net,　VGG,　GoogLeNet　and　ResNet　are　discussed　and　analyzed.　In　the　aspect　of　increasing　the　amount　of　data,　we　introduce　the　difficulties　of　increasing　the　number　of　annotated　samples　by　manual　way,　and　the　effect　in　improving　the　performance　of　convolutional　neural　networks　by　data　augmentation.　In　the　aspect　of　improving　training　methods,　we　introduce　the　generalized　regularization　techniques　such　as　the　L2　regularization,　Dropout,　DropConnect　and　Maxout,　several　frequently-used　neuron　activation　functions　such　as　the　sigmoid　function,　the　tanh　function,　the　ReLU　function,　the　LReLU　function,　and　the　PReLU　function,　several　different　loss　functions　such　as　the　softmax　loss,　the　hinge　loss,　the　contrastive　loss　and　the　triplet　loss,　and　the　basic　idea　of　the　batch　normalization　technique.　In　the　field　of　computer　vision,　this　paper　focuses　on　the　more　recent　research　progress　of　convolutional　neural　networks　in　image　classification,　object　detection,　face　recognition,　pedestrian　recognition,　image　semantic　segmentation,　image　captioning,　image　super　resolution,　human　action　recognition　and　image　retrieval.　From　the　prospective　of　the　human　visual　cognitive　mechanism,　we　analyze　the　relevant　theoretical　achievements　of　hierarchical　processing　in　the　visual　system　and　＇global　first＇　visual　and　cognitive　process,　and　some　theoretical　implications　for　the　current　computational　models.　Finally,　some　remained　problems　and　challenges　of　the　brain-like　intelligence　research　based　on　deep　convolutional　neural　networks　are　concluded.　©　2019,　Science　Press.　All　right　reserved.