Second-Order Convolutional Neural Network Based on Cholesky Compression Strategy

In the past few years, Convolution Neural Network (CNN) has been successfully applied to many computer vision tasks. Most of these networks can only extract first-order information from input images. The second-order statistical information refers to the second-order correlation obtained by calculating the covariance matrix, the fisher information matrix, or the vector outer product operation on the local feature group according to the channels. It has been shown that using second-order information on facial expression datasets can better capture the distortion of facial area features, while at the same time generate more parameters which may cause much more computational cost. In this article we propose a new CNN structure including layers which can (i) incorporate first-order information into the covariance matrix; (ii) use eigenvalue vectors to measure the importance of feature channels; (iii) reduce the bilinear dimensionality of the parameter matrix; and (iv) perform Cholesky decomposition on the positive definite matrix to complete the compression of the second-order information matrix. Due to the incorporation of both first-order and second-order information and the Cholesky compression strategy, our proposed method reduces the number of parameters by half of the SPDNet model, and simultaneously achieves better results in facial expression classification tasks than the corresponding first-order model and the reference second-order model.