Print Email Facebook Twitter Rotation invariant filters in CNNs Title Rotation invariant filters in CNNs: applied to segmentation of aerial images for land-use classification Author Dhar, Aniket (TU Delft Electrical Engineering, Mathematics and Computer Science) Contributor van Gemert, J.C. (mentor) Reinders, M.J.T. (graduation committee) Scharenborg, O.E. (graduation committee) van der Maas, Daan (mentor) Degree granting institution Delft University of Technology Programme Electrical Engineering | Embedded Systems Date 2018-11-21 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 invariantconvolutional 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 formulti-class semantic segmentation on the DeepGlobe 2018 Satellite Image Understanding Challenge. Subject Computer VisionDeep LearningMachine LearningConvolutional Neural Networks To reference this document use: http://resolver.tudelft.nl/uuid:1624a31f-7976-425a-a2b6-d6937cc39895 Part of collection Student theses Document type master thesis Rights © 2018 Aniket Dhar Files PDF MSc_Thesis_Report_Aniket_Dhar.pdf 1.79 MB Close viewer /islandora/object/uuid:1624a31f-7976-425a-a2b6-d6937cc39895/datastream/OBJ/view