Image Classification Based On Second-order Information Convolution Neural Network

In recent years,Convolution Neural Network(CNN)has been successfully applied to many computer vision tasks.The core of CNN is through convolution operation,pooling and nonlinear operation.It enables the network to fuse spatial and channel information in the local receptive field of each layer to establish high-dimensional feature representation.By performing linear combination and element by element nonlinear operation,traditional CNN can extract only first-order information from the input image.The second-order statistical information is the second-order correlation obtained by calculating the covariance matrix,Fisher information matrix or based on the vector outer product operation for the local feature group according to the channel.Some research has shown that using high-order representation can enhance the capability of nonlinear modeling.However,most current research generates much more parameters when considering high-order information,which leads to large storage and low efficiency of the model.In this thesis,the first-order and second-order information in the image is combined to further improve the generalization ability of Convolutional Neural Network.And study how to make full use of second-order information in the case of maximum dimension reduction of the information matrix.Notice that the second-order information matrix is symmetric and positive definite,a decomposition method is found to reduce the parameters to lose as little information as possible.In order to achieve the above goal,we design an efficient and convenient algorithm and apply it to the first-order information and the second-order information network model.Our main research contributions are divided into the following two parts.In the first part,we propose a new CNN structure based on second-order statistics,including information matrix calculation,parameter transformation and second-order statistical information decomposition.This layer structure has its own advantages: it can(I)extract the covariance matrix from convolution activation,and integrate the first-order information into the covariance matrix to complete the nonlinear operation of the second-order and first-order information;(II)decompose the information matrix with eigenvalues to measure the importance of the feature channel;(III)give the formula for setting the size of the bilinear parameter matrix,and(IV)analyzing the information matrix The Cholesky-DLT compression strategy is established for the second-order information matrix.The transformation from Riemannian manifold structure to Euclidean space is realized by the proposed matrix decomposition method,and the number of parameters is reduced.Experimental results show that by adding Cholesky-DLT compression strategy(Cholesky Decomposition Log Transformation),the model parameters are halved and the accuracy of image classification is improved.In the second part,we propose a new module which uses multi-layer one-dimensional convolution to model the second-order statistics,effectively extracts high-order statistics for feature representation,and can show the direct dependence between modeling features and remote features.By adding So Conv-block module(Second order Convolution block),the attention mechanism is used to learn the preference degree of the features,and the importance of the modeling features is displayed.The proposed So Conv-blocks can be well applied to any Network model.By adding a small number of parameters,the classification accuracy of images can be improved.