Invariant Generalised Hough Transform

27
Invariant GHT Adham Beyki Khoshk (abk2g11), MS Kapoor (msk1g11), Shashank Dhariwal (sd8g11) 1

Transcript of Invariant Generalised Hough Transform

Page 1: Invariant Generalised Hough Transform

Invariant GHTAdham Beyki Khoshk (abk2g11), MS Kapoor (msk1g11), Shashank Dhariwal (sd8g11)

1

Page 2: Invariant Generalised Hough Transform

Introduction

• HOUGH TRANSFORM • IMPORTANT TERMS & DEFINITIONS• GENERAL HOUGH TRANSFORM (GHT)• INVARIANT GHT• IMPLEMENTATION• RESULTS (DEMO)• ADVANTAGES AND DISADVANTAGES• APPLICATIONS

2

Page 3: Invariant Generalised Hough Transform

Hough Transform

• General Definition HT is a technique to extract required model features from Images and its purpose is primarily to find imperfect instances of objects within a certain class of shapes by a voting procedure. • Invented by R Duda and P Hart in year 1972 for

identifying position of arbitrary shapes [1]• Later in year 1981 , Dana H Ballard was responsible for

popularising GHT in Computer Vision community [1]• Example : HT of a line [ 1];

main idea – characterisation of line not as image points , but as certain parameters which define it.

3

Page 4: Invariant Generalised Hough Transform

Various Definitions

• Hough Space/Accumulator space the space where the votes of an image point is defined in the terms of the selected parameters

• Image Space The 2D space that represents the image or (edge detected image)

• Parameterisation Selection of certain parameters that map a model point into the Hough space

• Mapping Process of defining a rule(using parameters) that binds the Hough space and model

4

Page 5: Invariant Generalised Hough Transform

Various Definitions

• Accumulator array an array/matrix where result of HT is stored, its dimension is equal to the no of unknown parameters defining the transform. And the bin having the maxima is considered for extraction of the shape

• R table [2] Constructed by examining the boundary points of a shape. And later acts as look up table for obtaining parameter values of image points.

5

Page 6: Invariant Generalised Hough Transform

• Finds application where shapes are arbitrary and the voting is carried out for the particular position, orientation and scaling, based on pre defined look-up tables

• general equation :

ω (θ,b,λ,ρ) = b(x,y) + λR(ρ)v(θ) [ 3]

• Mapping function : b = ωi - λR(ρ)v(θ) [3]

i.e mapping from the orientation of an edge point (ωi)to the reference point (b)

6

Translation vector

Rotation MatrixScale factor

Model Shape

Image shape

General Hough Transform

Page 7: Invariant Generalised Hough Transform

• Evidence gathering – using 4D accumulator space

• Additional constraint – gradient direction information as we do not know v(θ)

φ΄i = φ΄ (θ) – ρ• For using the mapping function we need to know v(θ) and this

is determined by solving equation above, i.e we know φ΄i and find point v(θ) for which the gradient direction = 0.

• v(θ) is then used to compute λR(ρ)v(θ) using R- Table (where, rows = value of φ΄i )

• Therefore extraction = building R table + evidence gathering from the accumulator space.

7

General Hough Transform

Page 8: Invariant Generalised Hough Transform

• Previously GHT improved evidence gathering by considering a point and its gradient direction (which was not invariant).

• This leads to good noise handling but high computational complexity.

• Therefore we need to replace gradient direction by a feature that is invariant. (e.g ‘angle’ : rotation & scale invariance)

• The feature can be colour of the point or any other point that does not change in model and image space.

• Now search for a point with same ‘gradient direction’ in GHT is replaced with search with same ‘invariant feature’ in IGHT

• Search is in 2D space rather than 4D space. (reduced computation) 8

Invariant GHT

Page 9: Invariant Generalised Hough Transform

• Q (ωi) = Q (v(θ)) – constraint equationwhere, Q – invariant function that computes a feature at

the point• Given ‘ωi ‘and a fixed angle ‘α’ we can determine ‘ωj’ (second

point)• third point ‘ωT’ – defined by intersection of ωi & ω j

9

Invariant GHT

Page 10: Invariant Generalised Hough Transform

• The angle ‘ β’ obtained replaces the gradient angle in R table and is invariant to rotation and scaling• Now we have two parameters β and k (distance vector)• k changes with rotation and scale • therefore, k is redefined as invariant using the relation

k = φ΄i - φ΄’i (in terms of gradient angles)• This value of k is pre computed and stored in R table as

function of β.• Therefore now the solution is , given ωi we find v(θ) and φ΄’i

and β and using the look up table obtain k and carryout evidence- gathering procedure.

10

Invariant GHT

Page 11: Invariant Generalised Hough Transform

(𝑥 𝑖 , 𝑦 𝑖 )𝑀𝑎𝑝⇒

(𝑐 ,𝑚 )

Hough Transform for a Line in Cartesian System

11

Page 12: Invariant Generalised Hough Transform

(𝑥 𝑖 , 𝑦 𝑖 )𝑀𝑎𝑝⇒

(𝜌 ,𝜃 )

Hough Transform for a Line in Polar System

12

Page 13: Invariant Generalised Hough Transform

Hough Transform for a Circle

13

Page 14: Invariant Generalised Hough Transform

A model shape can be defined by a curve

Parameter Meaning

ω image shape

b translation vector (x0, y0)

λ scale factor

R(ρ) rotation matrix

v(θ) model shape

Generalised Hough Transform

14

Page 15: Invariant Generalised Hough Transform

• Using extra constraint

• Representation in polar coordinates

Generalised Hough Transform

15

Page 16: Invariant Generalised Hough Transform

Invariant Feature / New Constraint

Invariant Generalised Hough Transform

16

Page 17: Invariant Generalised Hough Transform

Invariant Generalised Hough Transform

17

Page 18: Invariant Generalised Hough Transform

• Efficient only if a high number of votes fall in the bin.

• Ambiguity when there are a large number of parameters.

• Becomes computationally expensive when number of parameters are more than three.

• Complexity increases at a rate of O(Am-2).

• Efficiency dependent on quality of input data.18

Limitations of Hough Transform

Page 19: Invariant Generalised Hough Transform

• Rotation and scale invariant.

• Computationally inexpensive.

19

Advantages of Invariant GHT

Page 20: Invariant Generalised Hough Transform

• Produces noise.

• More sensitive to occlusion and noise.

• Reduced accuracy.

• Lack of longitudinal localization of lines in parameter space.

20

Disadvantages of Invariant GHT

Page 21: Invariant Generalised Hough Transform

• Corner Detection.

• Detect 3D objects in laser range data.

• Artefact Matching.

• Inshore ships detection.

21

Applications of Invariant GHT

Page 22: Invariant Generalised Hough Transform

Since it is better to see the output space in a 3D view, it is better to run .m files and use the Rotate 3D tool in MATLAB.

22

Results

An image from a wrenchInput

HT Output – Accumulator Spaceusing Cartesian coordinates

Page 23: Invariant Generalised Hough Transform

23

Results

An image from a wrenchInput

HT Output – Accumulator SpaceUsing polar coordinates

Page 24: Invariant Generalised Hough Transform

24

ResultsThe Model Shape

The model shape in the image

GHT Output – Accumulator Space

Page 25: Invariant Generalised Hough Transform

25

ResultsThe Model Shape

The model shape in the image

IGHT Output – Accumulator Space

Page 26: Invariant Generalised Hough Transform

1. Wikipedia, Hough Transform, http://en.wikipedia.org/wiki/Hough_transform2. D H Ballard , Generalising The Hough Transform to Detect Arbitrary Shapes, Pattern Recognition Vol 13, 2.pp .111-122,Pergamon Press Ltd, 1981.3. Mark Nixon & Alberto Aguado, Feature Extraction & Image Processing, Second Edition, Elsevier Ltd, 2008.4. Jian Xu, Kun Fu, Xian Sun, An Invariant Generalized Hough Transform Based Method of Inshore Ships Detection. International Symposium on Image and Data Fusion (ISIDF), 2011.5. E.R. Davies, Application of the generalised Hough transform to corner detection. IEE PROCEEDINGS, Vol. 135, Pt. E, No. 1, JANUARY 1988.

26

References

Page 27: Invariant Generalised Hough Transform

6. Kourosh Khoshelham, EXTENDING GENERALIZED HOUGH TRANSFORM TO DETECT 3D OBJECTS IN LASER RANGE DATA. ISPRS Workshop on Laser Scanning 2007 and SilviLaser 2007, Espoo, September 12-14, 2007, Finland.7. Paul Lewis, Peter Durham, Stephen Shennan, Artefact Matching and Retrieval using Generalized Hough Transform. Proceedings of Computer Applications in Archaeology. 25—308. Jose A.R. Artolazabal and John Illingworth, LIGHT: Local Invariant Generalized Hough Transform. 18th International Conference on Pattern Recognition, 2006. ICPR 2006.

27

References