You are on page 1of 21

Waldemar Villamayor Venialbo

Horacio Legal Ayala

Facultad Politcnica Universidad Nacional de Asuncin Paraguay

XXXV Conferencia Latinoamericana de Informtica (CLEI 2009) Pelotas, RS Brasil

Contextualization of the Problem Problem scenario Challenge State of the art Proposed method

Contextualization of the problem


F
Brand Logo Sketch Query

Binarization

Stroke Segmentation

Feature Extraction

SF

Exclusion Measuring
SG

Brand Classification

G
Brand Logo Image Database

Binarization

Stroke Segmentation

Feature Extraction

Descriptors Knowledge Base

F G
Yes

No

SF

Candidate is rejected

An scheme of our livestock brand registration prototype system

Problem scenario: Stroke segementation


The cattle brand registration system requires these brands to be recorded as sketches handwritten on a digital table. With this digitalization mechanism:

The width of the strokes is not a predominant line feature.


The only informative feature that characterizes a curve is its orientation.

Challenge
The techniques for the segmentation of isolated lines are trivial, but the methods of detection and separation of overlapping curves are not. We need to locally analyze the regions of the strokes and establish some criteria to determine the existence of X-like intersections and T-like junctions.

State of the art


Current techniques
There exist several previous studies related to junction analysis. Partial stroke grouping and line continuation also received a considerable number of studies. Some successful methods allowing junction analysis and primitive stroke reconstruction make use of the orientation space.

Nevertheless, there are certain conditions where these methods may fail to correctly segment a set of overlapping strokes, e.g., when junctions or intersections occur at very small angles or on a region with high curvature.

Our view of the problem Stroke segmentation Implementation Experimental result

Several kind of points

Figure: A terminal point, ( ) = 1 labeled ; a normal point, ( ) = 2 labeled ; a T-like junction point, ( ) = 3 labeled ; a X-like intersection point, ( ) = 4 labeled ; and an isolated dot, ( ) = 0 labeled .

Types of primary strokes


Figure: Several types of primary strokes. An isolated dot, ; an isolated stroke, ; untied primary strokes, , , , , , and ; and a tied primary stroke, .

Primitive strokes

A primitive stroke is the union of one or more primary strokes, provided that the resulting region determines a continuous 2-dimensional curve or a simple polygonal. This definition allows for two or more primary strokes joined by L-vertexes to be treated as primitive strokes. E.g., in Fig. 3.

Figure: Primitive strokes.

Stroke segmentation
Stage 1 Primary stroke separation
1) Junction and intersection detection
A successful extraction of a primary stroke relies, primarily, on the correct identication of singular points.

2) Primary stroke extraction


To extract a primary stroke we proceed isolating each regular region together with each delimiting singular region, if any.

Stroke segmentation
Stage 2 Primitive stroke reconstruction
Orientation space mapping To determine which primary strokes are part of the same primitive stroke and group them together, we proceed as follows:

1) Construct a 3-dimensional orientation space .


The - and - axes are the same axis determined by the image's reference frame, and the -axis is the line orientation parameter.

2) Map each point ( , ) into point ( , , ( )) .


Where ( ) represent the orientation of the tangent line at point , measured with respect to the positive semi-axis in the 2-dimensional space containing .

Stroke segmentation
Stage 2 Primitive stroke reconstruction
Orientation space mapping

Orientation space mapping


After mapping all points in B, into the orientation space, every primary stroke will be connected to other primary strokes determining the primitive strokes they conform.

Figure: Mapping a curve into the orientation space.

Stroke segmentation
Stage 2 Primitive stroke reconstruction
Primary stroke grouping

Orientation space projection


Projecting continuous curves from the orientation space back onto individual planes will produce separated images for each primitive stroke in the original image.

Figure: Each continuous curve in the orientation space determines a primitive stroke.

Implementation details
Primary stroke separation
Once a suitable skeleton is achieved, we take the degree of freedom ( ) of each pixel in the skeleton as the number of pixels in the 8-neighborhood of pixel .

Figure: Skeleton of an actual cattle brand.

Figure: Junction and intersection points.

Implementation details
Primary stroke grouping
Each primary stroke is extracted and labelled, then they are mapped into the orientation space to group them in the primitive strokes forming the cattle brand. We only map the gray colored regions from the Fig. 8i.e., the singular regionsand its surrounding pixels to detect if the neighboring primary strokes are connected in the orientation space.

Figure: Labelled primary strokes and two singular regions, in gray.

Implementation details
Primitive stroke segmentation
A set of connected primary strokes is treated as an unique primitive stroke. The primitive strokes so obtained still preserve their original spatial distribution in the 2-dimensional image plane.

Figure: Primitive strokes.

Experimental results
n=315 real brand images
Table: Segmentation error rates for two stroke segmentation approaches. One that performs junction analysis on thick lines using the Point to Boundary Orientation Distance (PBOD) technique and our approach using the skeleton of the brand. Junction analysis using Transformation Displacement PBOD 13.3% Skeleton 8.3%

Rotation
Scaling Stretching Shearing Mean Values

18.1%
13.7% 14.9% 16.2% 15.2%

11.4%
9.0% 9.5% 10.7% 9.8%

We presented a stroke segmentation algorithm suitable to perform similarity measuring of handwritten livestock brand images. The skeleton of the brand sketch is used to identify the strokes junction and intersection points, and to determine the primary strokes that compose the cattle brand. An orientation space mapping technique is used to group these primary strokes in primitive strokes that will allow the local shape feature extraction needed to perform the similarity measuring.

Slight improvement: Compared to another well known technique, the proposed method shows a slight improvement in the segmentation of the strokes in cattle brand images undergoing several kinds of transformations.

Gracias! Thanks! Obrigado!

You might also like