Abstract: In order to make full use of the feature information hidden in the image, this paper proposes to fusion the low latitude feature in fully connected layer, then constructs a double-channel convolutional neural network with the high-low level latitude feature. First, we construct a traditional double-channel convolutional neural network, and then set different sizes of convolution kernel on each channel, connect the double-channel CNN pooling layer to the full connection layer, at the same time, the features extracted from the first layer of the two channel convolution neural network are also directly transmitted to the fully connected layer, and this allows the extracted primary and advanced feature maps to be fused on the full link layer. Finally, the fusion data is input to the Softmax classifier to classify.  Simulation results of different algorithms on fashion-mnist and CIFAR-10 databases show that this model obtains higher classification accuracy.

Key words: feature fusion, double-channel convolutional neural network, convolution kernel, pooling layer