Frequency Domain Image Classification with Convolutional Neural Networks

Abstract

The purpose of this thesis was to explore image classification in the frequency domain using convolutional neural networks. Image classification is a common application of machine learning where image samples are categorized into classes. This application is valuable in many fields, and can be performed using various machine learning methods.

Neural networks are layered structures of computational units. They can be trained to perform classification. Convolutional neural networks add convolution operations to this layered structure. Traditionally, samples used for image classification have red, green, and blue color channels which display information in the spatial domain. The frequency domain representation of these images captures crucial information about geometric features, which may improve the performance of neural networks.

To conduct this experiment, convolutional neural network models were built following a selected baseline architecture. The models were built for both color images and images transformed into the frequency domain using a Fourier transform operation. These models were then trained and evaluated to compare their performance in relation to accuracy.

While the models required similar amounts of training, the ones using frequency domain images were only able to achieve half the prediction accuracy of their color image counterparts. The results demonstrate that the benefits of convolutional neural networks are lost when classifying images that are transformed into the frequency domain.