Small-scale Convolutional Neural Networks with Learnable Gabor Filter for Image Classifications

Recently, small-scale convolutional neural networks (CNNs) have become a hot research topic due to the fewer training parameters and shorter training time, while their accuracy is generally degraded compared with large-scale CNNs. To this end, in this paper, we propose a small-scale CNN with sparse representation (a sparse small-scale CNN) for image classification. First, the Gabor filter is a representative sparse representation, so we integrate it into the standard AlexNet to establish a sparse Gabor-AlexNet architecture to better extract the feature of images. Second, to improve further the performance of the above Gabor-AlexNet., a learnable Gabor-AlexNet with only the increase of a few learning parameters is constructed for image classification. Finally, we test the Gabor-AlexNet on small-scale datasets including Extended Yale B, CASIA-FaceV5., Original Brodatz and, DTD. The proposed method outperforms the standard AlexNet in the precision and speed as classification benchmark. Then, our learnable Gabor-AlexNet is also superior to the Gabor-AlexNet for classification performance on the CIFAR-10 dataset.