Incorporating Orthogonal Moments in CNNs

Sammanfattning: Convolutional neural networks (CNNs) can accurately classify objects in images using convolutional layers that extract features from images. Features from images can also be extracted using image moments, such as Gabor and Zernike moments. The aim with this project was to evaluate the impact on the accuracy of a CNN when the initial layer of the CNN is substituted with a layer of Zernike or Gabor filters. Three CNNs were trained on the dataset Dogs vs cats; a CNN with one hidden layer, GaborNet and Alexnet. Additionally, GaborNet was trained on the KidneyECCV dataset. After each epoch, the accuracy of the CNN was measured using a validation set. A Zernike layer increased the accuracy of the CNN with one hidden layer by7.84% after the first epoch and by 2.83% after 50 epochs when training on the Dogs vs Cats dataset. When training GaborNet and AlexNet on Dogs vs Cats, an initial Gabor layer improved accuracy, while a Zernike layer decreased accuracy. An initialZernike layer yielded the highest initial accuracy for GaborNet on the KidneyECCV;and an initial Zernike or Gabor layer yielded higher accuracy than the reference valuebut not as high accuracy as the GaborNet with a Gabor layer after 50 epochs. In conclusion, substituting a CNN's initial layer with an image moment layer can improve the accuracy of the network, with the effect, however, varying betweendatasets and networks.