Design of Convolutional Neural Network Based on Tree Fork Module

Abstract

The convolution kernel in the convolutional layer of the traditional convolutional neural network is single, and the feature extraction of the feature map is insufficient and not deep enough. This paper proposes a network built by a tree fork module. The tree-fork module uses multi-convolution kernels for cross-convolution, which not only increases the complexity of the network, but also facilitates feature selection, improves the generalization ability of the network, and extracts more recessive features in the feature map. The basic architecture of the network is unchanged. The traditional convolution module in the network intermediate convolutional layer is replaced by a tree fork module. By training in several public datasets, the performance of tree-fork convolution network is compared with that of traditional convolution network. The accuracy of the network of the tree-fork module in 101_food, caltech256, GTSRB and cifar10 is 4.1 percentage points, 4.7 percentage points, 1.8 percentage points and 1.9 percentage points higher than the traditional CNN. It can be seen from the experimental results that the network has improved in recognition accuracy performance.