Gabor Binary Layer in Convolutional Neural Networks

Convolutional neural networks(CNNs) have achieved overwhelming success in image recognition. For all the architectures of CNNs, the low-level features extracted from the input is essential for the whole model because it determines the upper bound of accuracy. However, due to the random initialization and end-to-end training mechanism, there is no guarantee that the learned convolutional layer is an excellent low-level feature extractor. Thus, we propose Gabor binary layer(GBL) as a more efficient alternative to the first convolutional layer in CNNs. The design principle of GBL is motivated by local Gabor binary patterns. The GBL is comprised of a module of pre-defined Gabor convolutional filters in different orientations and shapes, and a module of fixed randomly generated binary convolutional filters to encode the Gabor features. By replacing the first layer of Resnet-56, Resnet-110, and LBCNN with the GBL, we achieve better performances on SVHN, CIFAR-10 and CIFAR-100 than that of the original models. Results show that the proposed GBL outperforms the standard convolutional layer for extracting low-level features, and thus the GBL can easily improve the performances of CNNs on image recognition.