A Compact Spectral Model for Convolutional Neural Network

The convolutional neural network (CNN) has gained widespread adoption in computer vision (CV) applications in recent years. However, the high computational complexity of spatial (conventional) CNNs makes real-time deployment in CV applications difficult. Spectral representation (frequency domain) is one of the most effective ways to reduce the large computational workload in CNN models, and thus beneficial for any processing platform. By reducing the size of feature maps, a compact spectral CNN model is proposed and developed in this paper by utilizing just the lower frequency components of the feature maps. When compared to similar models in the spatial domain, the proposed compact spectral CNN model achieves at least 24.11 $$\times $$ and 4.96 $$\times $$ faster classification speed on AT &T face recognition and MNIST digit/fashion classification datasets, respectively.