Although complex-valued convolutional neural networks (CVCNNs) have shown an improvement over their real-valued counterparts (RVCNNs) when trained on real-valued images, in order to harness the full potential of CVCNNs, they should be used with fully complex-valued inputs. Because every image has one and only one Fourier transform, the problem of classifying real-valued images is equivalent to the problem of classifying their complex-valued Fourier transforms. Experiments done using the MNIST, SVHN, and CIFAR-10 datasets show an improved performance of CVCNNs trained on Fourier-transformed images over CVCNNs trained on real-valued images (for which the imaginary part of the input is considered to be zero), and over RVCNNs.
CITATION STYLE
Popa, C. A., & Cernăzanu-Glăvan, C. (2018). Fourier transform-based image classification using complex-valued convolutional neural networks. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 10878 LNCS, pp. 300–309). Springer Verlag. https://doi.org/10.1007/978-3-319-92537-0_35
Mendeley helps you to discover research relevant for your work.