Accurate Image Recognition in Convolutional Neural Networks Based on Two-dimensional Discrete Fourier Transform
Fo Zhi Zhou, Guo Chun Wan, Mei Song Tong · 2019
A two-dimensional discrete Fourier transform (DFT) can transform an image from a spatial domain to a frequency domain. This conversion can more intuitively observe and process the image from the perspective of the frequency domain, and is more advantageous for operations such as frequency domain filtering. In recent years, convolutional neural networks (CNNs) have become a hot topic in the field of image recognition. Relevant algorithms are generally preprocessed using two-dimensional DFT before inputting the network. However, in practice, most common CNNs are processed in the real number domain, and the frequency domain information and phase information cannot be fully utilized. This paper aims to exploring the application of two-dimensional DFT in CNNs for image recognition. Due to the globality of Fourier transform, we use the grayscale image of human face that has been aligned as the training data to design the experiment. The experiment includes the following steps: the feature is identified in the frequency domain after the image undergoes a two-dimensional DFT; the frequency domain image after the two-dimensional DFT is input together with the original image to expand the input layer of the CNNs and is then identified. The experiment shows that the image can also recognize the features in the frequency domain after the two-dimensional DFT when the frequency domain features are good and the recognition accuracy can be greatly improved.