Input Self-Filtering by The Spectral Complement and Frequency Normalizations for CNN

Résumé / Summary:

The use of spectral domains with convolutional neural networks (CNN) has increased in the literature in recent years. While some efforts have targeted input preprocessing or processing within specialized models, few have reported Fourier-domain methods capable of improving the performance of general existing CNN architectures. We present a novel method called SFSC (Self-Filtering by the Spectral Complement), the first method to improve upon general CNN performance by modifying only the input in the Fourier domain. SFSC operates by projecting the input magnitude spectrum onto the signal axis under the assumption of an additive white noise model, and is implemented using an input block with only 20 additional parameters. To demonstrate the general effectiveness of the method, experiments are performed on 14 unique architectures, with random or pretrained weights, using three distinct image classification tasks (categories: CIFAR, textures: DTD, faces: KDEF). Ultimately, the versatility and adaptability of the SFSC allows it to significantly improve in virtually all cases with an overall accuracy up to +3.73%.

Find out more information related to our research at the LIVIA website: https://liviamtl.ca/