A Robust Approach for Local Interest Point Detection in Line-Drawing Images

1
A Robust Approach for Local Interest Point
Detection in Line-Drawing Images
The Anh Pham, Mathieu Delalandre, Sabine Barrat
and Jean-Yves Ramel RFAI group- PolytechTour,
France. CIL Talk Wednesday 7th March
2012 Athens, Greece
2
Overview

• Introduction
• Junction detection in line-drawing images
• Experiments and results
• Conclusion and future works

3
Introduction (1)

• This work is interested with graphic documents,
  especially the line drawings, some examples

4
Introduction (2)

• Interest points are a kind of local features
  (i.e. an image pattern which differs from its
  immediate neighborhood).
• Popular interest points include edges, blobs,
  regions, salient points, etc.
• In graphics documents, interest points are
  end-points, corners and junctions

Comparison of the approaches for corner and
junction detection
Approach Corner Junction Robustness
High curvature detection The-Chin89 ?
Intensity-based methods Harris89 ?
Model-based methods Chul05 ? ?
Segmentation-based methods Burge98 ?
Contour matching methods Ramel00 ? ?
Tracking methods Song02 ?
Local interest points
5
Introduction (3)

• High curvature detection is the task of
  segmenting a curve at distinguished points of
  high local curvature (e.g. corners, bends,
  joints).
• High curvature detection methods often includes
  include polygonal and B-splines approximation,
  wavelet analysis, etc.

• Key idea of the work is to drive high curvature
  detection methods to achieve junction detection.
• Two problems
• (1) How to extract the curves
• (2) How to merge the multiple detections

6
Overview

• Introduction
• Junction detection in line-drawing images
• Experiments and results
• Conclusion and future works

7
Junction detection in line-drawing images (1)
Flow-work of our approach
8
Junction detection in line-drawing images (2)

• (1) Skeletonization based on Di Baja
  (3,4)-chamfer distance DiBaja94
• (2) Branch linking and Skeleton Connective Graph
  Construction (SCG)
• based on Popel02

Skeletonization, branch linking
Skeleton graph
Path extraction
2D paths
Path representation
1D signals
High curvature detection

• Skeleton Connective Graph (SCG)
• node ended and crossing points
• edge skeleton branch

Candidates
Refining Correcting
9
Junction detection in line-drawing images (3)

• Path definition a sequence of edges of SCG that
  describes a complete stroke or a circuit.
• Three types of paths Stroke path, Circuit path
  and Hybrid path.
• Paths are extracted using anticlockwise direction
  between the nodes of graph SCG

Skeletonization, branch linking
Skeleton graph
Path extraction
2D paths
Path representation
1D signals
High curvature detection
Candidates
are branch pixels
d0





are branch extremities
Refining Correcting
is a crossing pixel
d0 is the extremity-crossing direction
10
Junction detection in line-drawing images (4)

• A 2D path P consists in N points (x1y1),
  (x2y2),,(xNyN)
• To represent a 2D path in 1D signal, we selected
  the Rosenfeld-Johnston method

Skeletonization, branch linking
Skeleton graph
Path extraction
2D paths
Path representation
1D signals
High curvature detection
? f(t)
straight-line ? -1
high curvature ?/2 0
Candidates
Refining Correcting
11
Junction detection in line-drawing images (5)

• Due to the q parameter, we must make the method
  shift invariant.
• To do so, we select starting point of lowest
  curvature i.e. f(t)?-1

Skeletonization, branch linking
Skeleton graph
Path extraction
A good starting point here (shift-invariant).
Not good starting point.
2D paths
Path representation
1D signals
High curvature detection
Candidates
Refining Correcting
12
Junction detection in line-drawing images (6)

• Using multi-resolution wavelet analysis because
  of its robustness and scale invariance (i.e.
  multi-resolution)Gao06.

Skeletonization, branch linking
Skeleton graph
Image (I)
2D curcuit path
1D representation
Path extraction
2D paths
Path representation
Multi-resolution wavelet analysis
1D signals
High curvature detection
Candidates
Refining Correcting
13
Junction detection in line-drawing images (7)

• (1) Single path level Remove the unreliable
  segments (i.e. length less than line thickness)
  and Connect the reliable segments togethers.
• (2) Inter-path level (using voting scheme)
• merging close junctions together based on line
  thickness.

Skeletonization, branch linking
Skeleton graph
Path extraction
2D paths
Path representation
a path with high curvature points
a SCG with high curvature points
result after removing short segments
1D signals
High curvature detection
Candidates
Refining Correcting
14
Overview

• Introduction
• Junction detection in line-drawing images
• Experiments and results
• Conclusion and future works

15
Experiments and Results (1)

• Evaluation protocol

Evaluation Criteria is the repeatability score
Schmid00
p
q
p is a model point
q is a detected point
Detection of p is positive if d(p,q)lt? with
d(p,q) the Euclidean distance
16
Experiments and Results (2)

• Datasets

Logos-UMD ISRC2011
Models 106 150
Degradation Rotation Scaling Kanungo noise Rotation Scaling Kanungo noise
Test images 1272 3600
17
Experiments and Results (3)

• Some results

Liu99 Identification of Fork point on the
Skeletons of Handwritten Chinese Characters,
PAMI (1999). Haris detector A combined
corner and edge detector. Alvey Vision
Conference, (1988).
18
Experiments and Results (4)

• Some visual results

19
Overview

• Introduction
• Junction detection in line-drawing images
• Experiments and results
• Conclusion and future works

20
Conclusions and future works

• Conclusions
• A junction detector is proposed for line-drawing
  images
• The obtained results are rather promising
• Future works
• The method is threshold dependent, we are looking
  for threshold adaptation
• (e.g. region of support
• Improve the robustness of the merging step using
  topological analysis
• (e.g. line bending energy minimization)
• More experiments with more interest points
  detector and datasets
• Applications of recognition of spotting (logos,
  symbols) and image indexing

21
Thank you for your attention!