A Convolutional Neural Network with Dynamic Correlation Pooling

Abstract

A dynamic correlation pooling method is proposed based on Mahalanobis distance to improve the accuracy of image recognition. The proposed correlation technique employs the correlation information between adjacent pixels of the image and is applied to Lenet-5 convolution neural network model, the performance of which is tested on data sets of MMIST, USPS and CIFAR-10, respectively. The empirical studies show that the proposed pooling method can improve the convergence rate and recognition accuracy in comparison with the max pooling, average pooling, stochastic pooling and mixed pooling.