Angle-aware object classification

A computer-implemented method of classifying objects in an image comprises receiving image data, S100, receiving incidence angle data that indicates the angle from which a detector collects the image data, S102, and using a machine learning model, e.g. a convolutional neural network (CNN), to classi...

Full description

Saved in:
Bibliographic Details
Main Authors Tapio Friberg, Pekka Laurila, Shay Strong
Format Patent
LanguageEnglish
Published 17.04.2024
Subjects
CALCULATING
COMPUTING
COUNTING
ELECTRIC DIGITAL DATA PROCESSING
PHYSICS
Online AccessGet full text

Cover

More Information
Summary:A computer-implemented method of classifying objects in an image comprises receiving image data, S100, receiving incidence angle data that indicates the angle from which a detector collects the image data, S102, and using a machine learning model, e.g. a convolutional neural network (CNN), to classify objects within the image into categories, e.g. water, ice, arable land, forest, or man-made constructions. Classifying the objects within the image, S104, is based on the image data's respective incidence angle data and parameter value data, e.g. intensity, colour channel, or phase information. The model may be trained (Fig.4) by receiving the training image and incidence angle data and concatenating the two to generate a training data patch. Generating this patch may involve augmenting the angle data to generate angle range data that are concatenated to the image data or may involve projecting the incidence angle or angle range data to a two-dimensional patch with the same size as the image data. The model may be trained on an elevation model of terrain imaged by the detector. The detector may be on-board an orbital satellite. The image may be a synthetic-aperture radar (SAR) image.
Bibliography:Application Number: GB20210018441