distance sets for shape filters and shape recognition

33
Distance Sets for Shape Filters and Shape Recognition Cosmin Grigorescu, Student Member, IEEE, and Nicolai Petkov

Upload: alisa-frazier

Post on 31-Dec-2015

48 views

Category:

Documents


1 download

DESCRIPTION

Distance Sets for Shape Filters and Shape Recognition. Cosmin Grigorescu , Student Member, IEEE, and Nicolai Petkov. INTRODUCTION. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Distance Sets for Shape Filters and Shape Recognition

Distance Sets for Shape Filtersand Shape Recognition

Cosmin Grigorescu, Student Member, IEEE,

andNicolai Petkov

Page 2: Distance Sets for Shape Filters and Shape Recognition

INTRODUCTION

When only the contours or even parts of the contours of the objects are left, like in Fig. 1(b) and (c), we are still able to discriminate between the two objects.

Page 3: Distance Sets for Shape Filters and Shape Recognition

A. Distance Sets

Let be a set of perceptually significant points in an image, which we will call feature points.

Let be the distance between point and its-nearest neighbor from S, . We call the local descriptor :

the distance set, more precisely the N-distance set, of point to its first nearest neighbors within .

Page 4: Distance Sets for Shape Filters and Shape Recognition

A. Distance Sets

The feature points in an image can be of different types.

In the character “ C”, for instance, one can consider the points at the extremities of the contour to be of an “end-of-line” type and all other points to be of another, “line” or “contour” type.

As a matter of fact, similar, evidently perceptually significant features are extracted in the visual cortex by so-called simple and complex cells (for lines and edges) , and end stopped cells.

Page 5: Distance Sets for Shape Filters and Shape Recognition

A. Distance Sets

Given two points and from two images and their associated distance sets and , we define the relative difference between the i-neighbor and j-neighbor distances of and , respectively, as

Page 6: Distance Sets for Shape Filters and Shape Recognition

A. Distance Sets

Page 7: Distance Sets for Shape Filters and Shape Recognition

In this way, the quantity indicates how dissimilar two points and are with respect to their distances to feature points in their respective neighborhoods.

A. Distance Sets

Page 8: Distance Sets for Shape Filters and Shape Recognition

A. Distance Sets -Example 1

Page 9: Distance Sets for Shape Filters and Shape Recognition

A. Distance Sets -Example 2

A weaker requirement , can be imposed on the points of

Page 10: Distance Sets for Shape Filters and Shape Recognition

B. Labeled Distance Sets

Let be the set of possible feature labels and let be one such label.

We define the labeled distance subset of a point p to its first neighbor feature points of type as follows:

Page 11: Distance Sets for Shape Filters and Shape Recognition

B. Labeled Distance Sets

Let be the dissimilarity between two labeled distance subsets of type computed according to (3).

Page 12: Distance Sets for Shape Filters and Shape Recognition

B. Labeled Distance Sets

for each label type we can assign a different weight and define the dissimilarity between two labeled distance sets and associated with and as follows:

Page 13: Distance Sets for Shape Filters and Shape Recognition

B. Labeled Distance Sets – Example 1

We assign two types of labels: a regular point, associated with a pixel coming from the contours of the letters, and an end-of-line point at the extremities of the contours.

Page 14: Distance Sets for Shape Filters and Shape Recognition

SETS OF (LABELED) DISTANCE SETS (Definition)

Page 15: Distance Sets for Shape Filters and Shape Recognition

SETS OF (LABELED) DISTANCE SETS -Example

Page 16: Distance Sets for Shape Filters and Shape Recognition

Dissimilarity Between Sets of (Labeled) Distance Sets

Given two sets of points Let be a one-to-one mapping from

and let be the set of all such mappings. Let be a positive constant.

Page 17: Distance Sets for Shape Filters and Shape Recognition

Dissimilarity Between Sets of (Labeled) Distance Sets

Finally, we define the dissimilarity between the sets of distance sets S1 and S2 as follows:

Page 18: Distance Sets for Shape Filters and Shape Recognition

Dissimilarity Between Sets of (Labeled) Distance Sets

Similarly, in the case of sets of labeled distance sets,and the cost of a mapping , is defined as

Page 19: Distance Sets for Shape Filters and Shape Recognition

Dissimilarity Between Sets of (Labeled) Distance Sets

Page 20: Distance Sets for Shape Filters and Shape Recognition

Orientation and Scale Invariance

Unless orientation is involved in the definition and extraction of features, a set of distance sets does not depend in any way on the orientation of an object in the 2-D plane:only distances matter.

scale invariance is obtained by resizing all objects to the same bounding box.

Page 21: Distance Sets for Shape Filters and Shape Recognition

DISTANCE SET SHAPE FILTERS (Definition)

Let S1 be the set of all possible subsets of a reference object.

We define the operator associated with the set S1 and having as parameters the number of neighbors taken into account , and a threshold value as follows:

Page 22: Distance Sets for Shape Filters and Shape Recognition

Example of Printed Character Recognition

Page 23: Distance Sets for Shape Filters and Shape Recognition

DISTANCE SET SHAPE FILTERS (Definition)

If sets of labeled distance sets are used with being the set of possible labels, the operator is defined as follows:

Page 24: Distance Sets for Shape Filters and Shape Recognition

Example of Printed Character Recognition

Page 25: Distance Sets for Shape Filters and Shape Recognition

Application to Continuous Handwritten Text

Page 26: Distance Sets for Shape Filters and Shape Recognition

Application to Continuous Handwritten Text

Page 27: Distance Sets for Shape Filters and Shape Recognition

Finding Objects in Complex Scenes

Page 28: Distance Sets for Shape Filters and Shape Recognition

We assigned a different label to each of the four orientations, and a labeled distance set was built for each point of the combined edge map by computing the distances to the first nearest neighbor points occurring in each oriented edge map, k=1…4 .

Finding Objects in Complex Scenes

Page 29: Distance Sets for Shape Filters and Shape Recognition

SHAPE COMPARISON

Page 30: Distance Sets for Shape Filters and Shape Recognition

SHAPE COMPARISON

For computing the cost of mappings, the value of the constant in (11) was chosenas the average pair-wise dissimilarity of two points:

Page 31: Distance Sets for Shape Filters and Shape Recognition

SHAPE COMPARISON –Application to Handwritten Character Recognition

In a more elaborate example, a database of 286 characters, 11 characters for each of the 26 letters, handwritten by one person was used, Fig. 13.

Page 32: Distance Sets for Shape Filters and Shape Recognition

SHAPE COMPARISON –Application to Object Recognition Based on 2-D Views

In the following, we evaluate the shape comparison procedure for recognition of objects based on their 2-D appearances.

Each object from this database is photographed from 72 different views around the object, corresponding to equally spaced 5 angular offsets

The number of neighbors taken into account within a distance set is , where are the numbers of feature points in two contour maps that are compared.

Page 33: Distance Sets for Shape Filters and Shape Recognition

SHAPE COMPARISON –Application to MPEG-7 Shape Database Retrieval

The time needed for computing the dissimilarity between two objects, each being described by approximately 250 points and having associated distance sets of 100 distances, is about 0.7 s on a regular Pentium III/667 MHz computer.

With this experimental setup, we achieved an overall retrieval rate of 78.38%. Previous studies report 78.17% [21], 76.51% [20], 76.45% [62] and 75.44% [19].