Research On Data Feature Fusion Classification Technology Based On Deep Learning

In recent years,deep learning has got violent development and made outstanding achievements in various fields.The advantage of deep learning lies in the fact that deep neural networks can fit large amounts of data.In other word,with the help of multi-layer structures,deep learning models can abstract the data from different scales.However,training deep learning models requires substantial computation,large amount of data and skills.In view of traits of deep learning,it is necessary to study the classification of features on different layers and saving computation.Firstly,the basic structures and learning methods of neural networks are introduced.In addition,the common structres of the convolutional neural networks are summarized,and several deep learning frameworks are recommended.Afterwards,analysis and research in regard to CNN and Support Vector Machines(SVM)hybrid classifier are conducted.With the analysis of training CNN,a CNN enhanced model is developed from utilizing the immediate feature of the CNN,which add another enhanced CNN on the original CNN to save computational cost while retaining the precision.CIFAR-10 and Scene15 datasets are used to verify the CNN enhanced model.In the study of classification,special attention is paid to decision fusion and feature fusion,concretely,majority voting and parallel and serial feature fusion strategies are implemented on GoogLeNet.Besides,MIT indoor scene and Stanford dogs datasets are harnessed to prove the effectiveness and the influence of feature compression on serial feature fusion.Finally,the application of cascade CNN for face detection is under researching.Part faces are applied to generate face candidate boxes and the fully connected layers are removed to streamline the CNN on the second stage,then FDDB and AFW datasets are used to verify the model.