Fourier Neural Operator For Image Classification

The present work seeks to analyze the performance of the Fourier Neural Operator (symbolized by FNO) as a convolution method for an image classification and how is its performance when compared to ResNet20. The possible advantage of this technique is the success in pattern recognition through the wave-forms properties of Fourier analysis and due the power of synthesis that the Fourier Transform has. ResNet20 took 21 minutes and 51 seconds for training, while the FNO took 4:11:14 hours to complete a hundred of epochs. The convolution occurs normally in the FNO, being perfectly possible to use in the image recognition processes, with accuracy, recall, precision and F-Score slightly better than ResNet20 and quite similar to other neural networks available in the literature. The overall Recall for FNO is 0.9934 for the training set and 0.660 for testing set. The overall Precision for FNO is 0.9935 for training and 0.662 for testing. The overall Recall for ResNet20 is 0.8480 for training and 0.596 for testing. The overall Precision for ResNet20 is 0.85 for training and 0.601 for testing. The overall accuracy for FNO is 0.9935 for training and 0.660 for testing. The overall accuracy for ResNet20 is 0.8480 for training and 0.596 for testing.