METHOD AND APPARATUS FOR OBTAINING A SYMMETRY INVARIANT DESCRIPTOR FROM A VISUAL PATCH OF AN IMAGE

• US 20130301932A1
• Filed: 07/06/2010
• Published: 11/14/2013
• Est. Priority Date: 07/06/2010
• Status: Active Grant

• Current Assignee: Jastec Co.
• Original Assignee: LTU Technologies S.A.S. (Jastec Co.)
• Inventors: Dolle, Simon, Jahard, Frederic, Massoudi, Ayoub, Winter, Alexandre, Moulon, Pierre, Wengert, Christian
• Granted Patent: US 9,251,432 B2
• Time in Patent Office: Days
• US Class Current: 382/203
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 10/757 Matching configurations of points or features

Abstract

The present invention concerns a method for deriving from an arbitrary local image descriptor a descriptor that is invariant under arbitrary mirror symmetries. For a point of interest, pi, for which the descriptor is to be determined a direction, ODi is determined. The local patch from which the descriptor is to be extracted is mirrored along ODi if a “smaller than” relation does not hold a feature extracted from the left half of the local image patch, in regard of ODi, compared to the right half of the local image patch. Thereby, the local patch is brought into a normalized intrinsic orientation. Thereafter, the descriptor is extracted. Examples include a mirror symmetry invariant version of Lowe'"'"'s SIFT.

Claims

1. ) A method for obtaining a descriptor that is invariant to arbitrary axial symmetries from a visual patch (PAi, nPAi) around a detected point of interest (pi) of an image (I), said descriptor being extracted by scanning the visual patch (PAi, nPAi) according to a scanning path (SPi), characterized in that the method

obtains an orientation direction (ODi) of the point of interest (pi),

computes a polarity (Poi) of the point of interest which defines the location of the scanning path starting point (sSPi) from the orientation direction (ODi), and

if the computed polarity (Poi) does not conform to a predefined polarity (pPoi), mirrors the extracted visual patch (PAi, nPAi) with respect to the orientation direction (ODi).

2. ) A method as claimed in claim 1, wherein a polarity of the point of interest is computed by dividing the visual patch into a plurality of parts and by computing statistical features from these parts, said statistical features defining the location of the scanning path starting point (sSPI) from the orientation direction (ODi).

3. ) A method as claimed in claim 2, wherein the visual patch is divided into two non-overlapping parts (A, B) separated by the orientation direction (ODi), a couple of features ((λ

_A,λ

_B)) is computed from these two non-overlapping parts (A, B), a hierarchical relation between the two computed features (λ

_A and λ

_B) of said couple is defined, and the computed polarity (Poi) does not conform to the predefined polarity (pPoi) when the two computed features which defined said computed polarity do not check said hierarchical relation.

4. ) Method as claimed in claim 2, wherein the visual patch is extracted from the image by a rotation invariant detector, the orientation direction is then obtained from an orientation angle resulting from the rotation invariant detector, and the features are computed once the extracted patch has been aligned with the orientation direction.

5. ) Method of claim 2, wherein the visual patch is extracted from the image by a rotation invariant detector, the orientation direction is then obtained from a process which is independent of the detector, and the features are computed once the extracted patch has been aligned with the orientation angle resulting of the rotation invariant detector.

6. ) Method of claim 1, which the image is extracted from a video.

7. ) An image indexing method in which at least one point of interest (pi) of an image is detected and a descriptor is obtained for each detected point of interest by scanning an extracted patch (PAi, nPAi) around said detected point of interest according to a scanning path (SPi), characterized in that the extracted patch (PAi, nPAi) is normalized according to the method of claim 1 to obtain an arbitrary axial symmetry invariant descriptor.

8. ) Image indexing method as claimed in claim 7, wherein the entire image may be taken as a visual patch and normalized against arbitrary axial symmetry, the orientation direction (ODi) being then computed on the entire visual patch.

9. ) Image indexing method as claimed in claim 7 wherein multiple points of interest are extracted from the image.

10. ) An image retrieval method in which a signature (S) of a query image and/or a previously stored signature (S) of an image I is/are formed by at least one descriptor which has been normalized according to the method of claim 1 to obtain an arbitrary axial symmetry invariant descriptor.

11-12. -12. (canceled)