Fourier Transform Layer: A proof of work in different training scenarios

In this work two established methods were merged: Fourier Transform and Convolutional Neural Net-work to classify images in several datasets. Fourier Transform is commonly used in signal processing as well as in image processing. Convolutional Neural Networks are well-suited to image analysis tasks, yet require a lot of processing power and/or time during training process. Fourier Transform Layer is introduced to increase the processing speed without sacrificing accuracy. The motivation is to present and compare an alternative approach to Convolutional Neural Networks, which could reduce the need for GPU training. Models containing only the novel layer were trained to classify images from widely accepted datasets and compared to classification results of simple models containing one convolutional layer. The comparison was performed in terms of test accuracy, Area-Under-Curve and training times. The results showed that, for images of size 128 x 128 and larger, models with the proposed layer reached test accuracy comparable to that reached by convolutional models (accuracy: 96% and 98% respectively), with at least 27% decrease in training time per one epoch on Central Processing Unit.& COPY; 2023 Nalecz Institute of Biocybernetics and Biomedical Engineering PAS. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Automated summary

In this study, the combination of the Fourier Transform and Convolutional Neural Network methods was used to classify images in multiple datasets. By incorporating Fourier Transform Layer, the processing speed was increased without sacrificing accuracy, providing an alternative approach to Convolutional Neural Networks that reduces the need for GPU training. Experimental results showed that models with the proposed layer achieved comparable test accuracy to convolutional models for images of size 128 x 128 and larger, with a minimum of 27% reduction in training time per one epoch on Central Processing Unit.