Methods for improving image search in large-scale databases
AssigneeFuji Xerox Co., Ltd.
InventorsQiong Liu, Chunyuan Liao, Xin Yang
Life of this patent
Estimating the term…
Abstract
Systems and methods for addressing the problem of specific planar object search from a large-scale image database. Given a query image with a particular object and a planar surface, the systems and methods attempt to find, from a large image corpus, a set of representative images in which that object appears. According to the complexity of geometric transformations between a query image and its target images, the whole problem can be categorized into two classes: Rotation-Scale-Translation (RST)-transformed image search, and affine/homography-transformed image search.
Claims19
A method of searching a data storage for an image, comprising:
- grouping features from a query image as a segment or a polygon;
- utilizing a processor to compute a transform of said segment or said polygon between the query image and at least one stored image stored in the data storage to obtain transform parameters;
- utilizing a grouping pattern on said transform parameters by identifying a dense region of said transformation parameters in Hough space;
- wherein the identifying the dense region of transformed features in Hough space comprises identifying a peak of total segment or polygon transformation parameters within a grid of the Hough space;
- utilizing a voting scheme to vote on inlier features for the query image;
- and
- utilizing the inlier features votes from the query image to search for a best matched stored image from said at least one stored image stored in the data storage.
A system, comprising:
- an apparatus for receiving a query image;
- a processor grouping features from the query image as segments or a polygons and grouping features from each stored image stored in a data storage as segments or polygons, wherein the processor further conducts segment/polygon transformation to obtain transformation parameters for the query image and each stored image, wherein the processor selects query image inlier features from transformation parameter dense regions in Hough Space;
- wherein said selecting query image inlier features from transformation parameter dense regions in Hough Space comprises identifying a dense region of transformed features in Hough space;
- wherein identifying the dense region of transformation parameters in Hough space further comprises identifying a peak of total segment or polygon transformations between the query image and a stored image within a grid of the Hough space;
- and
- a geometric verification unit checking the matching between the inlier features of the query image to corresponding inlier features of said each stored image stored in the data storage to search for a best matched stored image.
A non-transitory computer readable medium executing a process, the process comprising:
- grouping features from a query image as segments or polygons;
- performing a segment or polygon transformation check in Hough Space to determine inlier features for the query image;
- and
- utilizing the inlier features of the query image to search for a best matched stored image stored in a data storage
- wherein the determining the inlier features for the query image comprises identifying a dense region of transformation parameters in Hough space;
- and
- wherein identifying the dense region of transformed features in Hough space comprises identifying a peak of total segment or polygon transformation parameter sets within a grid of the Hough space.
Dependencies are inferred from the claim text.