Adaptive Gabor convolutional networks

Despite the great breakthroughs that deep convolutional neural networks (DCNNs) have achieved on image representation learning in recent years, they lack the ability to extract invariant information from images. On the other hand, several traditional feature extractors like Gabor filters are widely used for invariant information learning from images. In this paper, we propose a new class of DCNNs named adaptive Gabor convolutional networks (AGCNs). In the AGCNs, the convolutional kernels are adaptively multiplied by Gabor filters to construct the Gabor convolutional filters (GCFs), while the parameters in the Gabor functions (i.e., scale and orientation) are learned alongside those in the convolutional kernels. In addition, the GCFs can be regenerated after updating the Gabor filters and convolutional kernels. We evaluate the performance of the proposed AGCNs on image classification using five benchmark image datasets, i.e., MNIST and its rotated version, SVHN, CIFAR-10, CINIC-10, and DogsVSCats. Experimental results show that the AGCNs are robust to spatial transformations and have achieved higher accuracy compared with the DCNNs and other state-of-the-art deep networks. Moreover, the GCFs can be easily embedded into any classical DCNN models (e.g., ResNet) and require fewer parameters than the corresponding DCNNs.