eduzhai > Applied Sciences > Engineering >

CyCNN A Rotation Invariant CNN using Polar Mapping and Cylindrical Convolution Layers

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Deep Convolutional Neural Networks (CNNs) are empirically known to beinvariant to moderate translation but not to rotation in image classification.This paper proposes a deep CNN model, called CyCNN, which exploits polarmapping of input images to convert rotation to translation. To deal with thecylindrical property of the polar coordinates, we replace convolution layers inconventional CNNs to cylindrical convolutional (CyConv) layers. A CyConv layerexploits the cylindrically sliding windows (CSW) mechanism that verticallyextends the input-image receptive fields of boundary units in a convolutionallayer. We evaluate CyCNN and conventional CNN models for classification taskson rotated MNIST, CIFAR-10, and SVHN datasets. We show that if there is no dataaugmentation during training, CyCNN significantly improves classificationaccuracies when compared to conventional CNN models. Our implementation ofCyCNN is publicly available on this https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×