1 June 2018 in IEEE Signal Processing Letters
Affine Moment Invariant (AMI) is a kind of hand-crafted image feature, which is invariant to affine transformations. This property is precisely what the standard convolution neural network (CNN) is difficult to achieve. In this paper, we present a kind of network architecture to introduce AMI into CNN, which called AMI-Net. We achieved this by calculating AMI on the feature maps of the hidden layers. These AMI's will be concatenated with the output of fully connected(FC) layers of standard CNN and become the final output of the AMI-Net. By calculating AMI on the feature maps, we can not only extend the dimension of AMIs, but also introduce affine transformation invariant into CNN. Two network architectures and training strategies of AMI-Net are illuminated, one is two-stages, the other is end-to-end. To prove the effective of the AMI-Net, several experiments have been conducted on common image datasets, MNIST, MNIST-rot, affNIST, SVHN and CIFAR-10. By comparing with the corresponding standard CNN respectively, we verify the validity of AMI-net.