Rotation invariant filters in CNNs
applied to segmentation of aerial images for land-use classification
A. Dhar (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Jan van Van Gemert – Mentor
Marcel JT Reinders – Graduation committee member
Odette Scharenborg – Graduation committee member
Daan van der Maas – Mentor
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
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.