Convolution with Logarithmic Filter Groups for Efficient Shallow CNN

In this paper, we propose a new logarithmic filter grouping which can capture the nonlinearity of filter distribution in shallow CNNs. Residual identity shortcut is incorporated with the filter grouping to enhance the performance of shallow networks. The proposed logarithmic filter grouping is evaluated using a compact CNN structure with logarithmic group convolution modules which are composed of logarithmic filter groups and a residual identity shortcut. Our classification results on Multi-PIE dataset for facial expression recognition (FER) and CIFAR-10 dataset for object classification revealed that the compact CNN with proposed logarithmic filter grouping scheme outperforms the same network with the uniform filter grouping in terms of accuracy and parameter efficiency. Our results indicate that the efficiency of shallow CNNs can be improved by the proposed method.