Convolutional Correlation Branch Network for Image Classification

Abstract

ResNets are currently the mainstream networks used in engineering and research. However, cellular neural networks have inherent flaws, including typical problems such as structural defects and loss of convolutional information. In view of this, we propose a multi-branch deep network based on the correlation features of feature maps. Convolution-correlated features are derived from covariance matrices of feature maps within network layers. Because the covariance matrix belongs to a Riemannian manifold, in the network branch we use a self-attention network to map it to a linear space and combine it with advanced deep features to achieve image recognition. Using the covariance information of feature maps enables the network to effectively compensate for important information that may be lost in ResNet, and the multi-branch structure of the network effectively solves the problem of gradient disappearance or explosion during deep network training.The experimental results of image classification on four data sets: Cifar-10, Cifar-100,Flower-102 and Car-196 show that our proposed network significantly improves the performance of existing networks. It is worth noting that the multi-branch network model based on cooperative prevention difference we proposed can also be extended to the Transformer network, so it has good application prospects.