Rotation invariant filters in CNNs

Rotation invariant filters in CNNs: applied to segmentation of aerial images for land-use classification

Dhar, Aniket (TU Delft Electrical Engineering, Mathematics and Computer Science)

van Gemert, J.C. (mentor)
Reinders, M.J.T. (graduation committee)
Scharenborg, O.E. (graduation committee)
van der Maas, Daan (mentor)

Delft University of Technology

Electrical Engineering | Embedded Systems

2018-11-21

Convolutional neural networks are showing incredible performance in image classification, segmentation, object detection and other computer vision applications in recent years. But they lack understanding of affine transformations to input data. In this work, we introduce rotational invariant
convolutional neural networks that learn rotational invariance by design, and not from data. We build rotation invariant filters through parametric learning of linear combination of a basis set of filters, rather than modelling the filters ourselves. Our approach combines the learning capability of CNNs with custom filter selection. We show stability in performance under rotations in input images. We first validate our findings for classification on MNIST and then for
multi-class semantic segmentation on the DeepGlobe 2018 Satellite Image Understanding Challenge.

Computer Vision
Deep Learning
Machine Learning
Convolutional Neural Networks

http://resolver.tudelft.nl/uuid:1624a31f-7976-425a-a2b6-d6937cc39895

Student theses

master thesis

© 2018 Aniket Dhar