Convolutional Neural Networks based on Random Kernels in the Frequency Domain
Yuna Han, Bilel Derbel, Byung‐Woo Hong · 2021
Image classification in Fourier domain has been researched for many years via deep learning process, especially for spectral pooling methods and visualization. Point-wise multiplication of Fourier transformed image and kernel has solved high computational cost which is required for convolution operation through Convolutional Neural Network (CNNs) in spatial domain. However, there is still an open problem to deal with kernel method in the Fourier domain because larger images need bigger amount of computational cost by using the same size of kernel. In this work, we propose an efficient discrete Fourier transform-based CNNs using sparse random kernel. we expect the sparse random kernel contains critical low frequency and high frequency contents, but many zeros in kernel affect to the lower cost computation in the Frequency domain. The evaluation was conducted using the benchmark MNIST datasets using LeNet-5 (LeCun) which showed the possibility of this work, so we can expect that the implementation can be expanded for the future work.