Mosaicing documents for translation using video streams

摘要

Aspects of the present disclosure propose techniques for reconstructing a document mosaic using video streams of the document. The streams provide information identifying a layout that relates sequential frames of a video to each other. Once the streams are captured using a mobile device, it is then possible to reconstruct a virtual view of the entire document as though it were taken with a single camera shot. The reconstructed virtual view of the document will be suitable as input to an optical character recognition engine, which can be used for translating the document.

主权项

1. A computer-implemented method comprising:
for each of a first image of a first view of an object and a second image of a second view of the object, generating a respective feature point descriptor for each of multiple feature points included in the image; determining that a quantity of one or more of the feature point descriptors from the first image that are (i) indicated as similar to one or more feature point descriptors associated within a predefined sub-region of the second image, and (ii) associated within a corresponding predefined sub-region of the first image, satisfies a quantity threshold; and based on determining that a quantity of one or more of the feature point descriptors from the first image that are (i) indicated as similar to one or more feature point descriptors associated within a same predefined sub-region of the second image, and (ii) associated within a same predefined sub-region of the first image, satisfies a quantity threshold, creating a mapping between the predefined sub-region of the first image and the predefined sub-region of the second image.