Auto-filtrage des entrées par le complément spectral et des normalisations fréquentielles et l’utilisation de filtres en loi de puissance pour les réseaux de neurones convolutifs

Abstract

This thesis studies the use of image preprocessing methods using the Fourier domain, in order to improve the accuracy rates of the CNNs (convolutional neural networks) for image classification. For this purpose, three new methods are proposed. The first is the SFSC (Self-Filtering by the Spectral Complement) which consists of the use of a self-filtering of the inputs combined with combinations of normalization and frequency averaging, in order to simplify the images and make them easier to analyze by CNNs. The second and third method consist in filtering the input images or the output of the SFSC using power-law filters which evolve during learning. The efficiency of our new processes is verified by experiments on a total of 14 unique architectures, with random or pre-trained weights, for three types of image classification tasks : the general classification of natural images (CIFAR), the classification of textures (DTD) and the fine-grained classification of facial emotions (KDEF). The high versatility and adaptability of the SFSC allowed it to improve all models used except DenseNet and Wide ResNet on CIFAR which exhibited slight overall losses of less than 1 % accuracy. As for the power-law filters, they are not suitable for pre-trained models and generally perform less well than the SFSC, although they have exceeded it in some cases. Future research may be based on our methods and conclusions in order to improve them, to use them differently with CNNs or to develop new methods using the frequency domain with CNNs.

Date	18 May 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Matthew Toews (Supervisor)