Deep Low-dimensional Discriminative Learning Framework Based On CNN

Convolutional neural network(CNN)can extract deep-level features such as color,texture,shape,and topology of an image.It has good robustness and computational efficiency in applications that recognize displacement,scaling,and other forms of distortion invariance.However,when applying CNN to the transfer learning task of image data,there is a problem of low computational efficiency caused by the high feature dimension of the fully connected layer and the huge amount of parameters.Therefore,this paper proposes a more efficient CNN-based deep low-dimensional discriminative learning framework for image processing tasks in terms of transfer learning.First,the traditional CNN is used to extract high-dimensional redundant features(before the fully connected layer)of the image data to solve the impact of transfer learning on the difference between the distribution of the source and target domains and the model caused by the fine-tuning stage overfitting.Then,linear discriminant analysis(LDA)is used to map the highdimensional features to the low-dimensional space to solve the problem of low computational efficiency caused by a large amount of redundant information when the high-dimensional features are directly accessed into the dictionary for learning.Finally,combine dictionary learning with low-dimensional discriminative representation learning,use the sparse representation of dictionary learning to mine non-linear potential discriminative information hidden in high-dimensional data(features),and use label consistency to complete classification while dictionary learning.The experimental results show that the optimal classification accuracy on public data sets such as The Extended Yale B Faces can reach 99.92%,which significantly improves the image recognition rate compared with traditional CNN.And it effectively reduces the number of training parameters and reduces the consumption of computing resources during training.Compared with the original deep convolutional network,the framework of this paper discards the fully connected layers with huge parameters,which not only ensures the diversity of the original high-dimensional image data,but also enhances the robustness of the effective information.At the same time,the superior performance of the framework for image classification shows that it is generally feasible and useful to migrate the high-dimensional information of the image to a more sparse architecture.The paper has 33 pictures,15 tables,and 59 references.