eduzhai > Applied Sciences > Engineering >

Image classification in frequency domain with 2SReLU a second harmonics superposition activation function

  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Deep Convolutional Neural Networks are able to identify complex patterns andperform tasks with super-human capabilities. However, besides the exceptionalresults, they are not completely understood and it is still impractical tohand-engineer similar solutions. In this work, an image classificationConvolutional Neural Network and its building blocks are described from afrequency domain perspective. Some network layers have established counterpartsin the frequency domain like the convolutional and pooling layers. We proposethe 2SReLU layer, a novel non-linear activation function that preserves highfrequency components in deep networks. It is demonstrated that in the frequencydomain it is possible to achieve competitive results without using thecomputationally costly convolution operation. A source code implementation inPyTorch is provided at: this https URL

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×