Microsoft Technology Licensing, LLC

Boris Epshtein

Eyal Ofek

Yonatan Wexler

Detecting text using stroke width based text detection. As a part of the text detection, a representation of an image is generated that includes pixels that are associated with the stroke widths of components of the image. Connected components of the image are identified by filtering out portions of the pixels using metrics related to stroke width. Text is detected in the image based on the identified connected components.

Detecting text using stroke width based text detection

Images may be sorted and categorized by defining a frustum for each image and overlaying the frustums in two, three, or four dimensions to create a density map and identify points of interest. Images that contain a point of interest may be grouped, sorted, and categorized to determine representative images of the point. By including many images from different sources, common points of interest may be defined. Points of interest may be defined in two or three Euclidian dimensions, or may include a dimension of time.

Geo-relevance for images

A video that advertises a particular web site may be a form of video spam. For example, pornographers often advertise their web sites by displaying a link to their web sites in videos, and then placing the videos on video-sharing services. This type of video spam may be detected by analyzing the video for the presence of text and then determining whether the text is a URL. If the text is a URL, the URL may be checked to determine whether it points to an objectionable web site. The determination of whether a URL points to an objectionable web site may be made by comparing the URL with a blacklist and/or whitelist, or by retrieving the URL and analyzing the retrieved content. If a video is found to be an advertisement for an objectionable web site, action may be taken, such as removing the video from a content database.

Detection of objectionable videos

Visual images may be used to detect, or verify the existence of, landmarks. In one example, the landmarks may be used to provide driving or walking directions. A set of records may be examined to identify the street addresses of businesses or other entities. Additionally, street-level images may be captured with a camera, and the locations at which the images were captured may be recorded. The images may be evaluated with an Optical Character Recognition (OCR) process to determine what words appear in the images. The words in the image are compared with the names of entities whose addresses are near where the image was captured. If the words match the entity name, then a sign identifying the entity is presumed to be visible along a particular route, and the entity may be used as a landmark in a set of walking or driving directions.

Visual assessment of landmarks

Businesses and other entities may be assigned geographic locations based on various types of data. In one example, entities are associated with segments in a street grid. Additionally, data points (e.g., geotagged blog entries or other online information) are associated with segments in the street grid. For a given segment, data points are compared with entities to identify possible matches between the data points and the entities. Each match is assigned a score, and each match represents a possible physical location for a given entity. The locations of entities on the street are assigned, from among the matches, based on which assignment of entities optimizes the aggregate score for the segment, while satisfying one or more physical constraints.

Identifying physical locations of entities

In one example, video may be analyzed and divided into segments. Character recognition may be performed on the segments to determine what text appears in the segments. The text may be used to assign tags to the video and/or to the segments. Segments that appear visually similar to each other (e.g., segments that appear to be different views of the same person) may be grouped together, and a tag that is assigned to one segment may be propagated to another segment. The tags may be used to perform various types of tasks with respect to the video. One example of such a task is to perform a search on the video.

Tagging video using character recognition and propagation

A system described herein includes a text extractor component that extracts text from a digital image and a determiner component that automatically determines whether or not the digital image is a map of a geographic region based at least in part upon the extracted text. The system additionally includes a correlator component that generates correlation data that causes the digital image to be correlated with a portion of a reference map that pertains to the geographic region if the determiner component determines that the digital image is a map of the geographic region.

Map aggregation

Guided photography with image capturing device rendered user recommendations. Data is accessed that is associated with an intended photograph. A knowledge base is accessed to obtain data that is related to the data that is associated with the intended photograph. Recommendations are determined for the intended photograph based on the knowledge base data. The recommendations are provided to a rendering system of the device before the intended photograph is taken.

Growing community of inventors

Bothell, WA, United States of America

Boris Epshtein