Affine invariant descriptors using Fourier series
Introduction
The determination of invariant characteristics has been an important problem in pattern recognition since the origin of the field in the 1960s. Invariant descriptors are properties of geometric configurations, which remain unchanged under an appropriate class of transformations. In many cases, one is only interested in those properties that are invariant under a particular class of transformations. The fundamental difficulty in recognizing an object from its image is that the appearance of shape depends on the viewpoint. Invariant descriptions can be measured directly from objects without detailed prior knowledge of the actual affine, projective, or other transformations (Rothe et al., 1996; Weiss, 1993).
Recently, several authors have proposed the use of invariant descriptors for tracking contours in a sequence of images (Huang, 1981; Blake et al., 1993; Daoudi et al., 1999). Some of them have proposed to use Fourier Descriptors for motion estimation and invariant description (Daoudi et al., 1999; Wang and Clarke, 1990; Pauwels et al., 1995; Zahn and Roskies, 1972; Persoon and Fu, 1977) in the case of similarity group (rotation, scaling). Abter et al. (1990) has proposed invariant descriptors based on Fourier Descriptors for the case of Affine group.
A great amount of work in invariant descriptors field deals with differential ones. For example, the affine curvature proposed by Spivac (1970) and Pauwels et al. (1995) is invariant under affine transformations except for its starting point.
In this paper we combine Fourier series (FS) and differential invariant descriptors to produce normalized ones that remain unchanged under any affine transformation. In the same time, a motion estimation algorithm using FS is proposed which allows estimating general two-dimensional affine motions (i.e. affinity matrix and translation vector).
Deriving invariant descriptors from FS is motivated by the fact that this method has some very important advantages over known ones:
- •
The main advantage of the proposed method is that, contrarily to the Fourier descriptors, no knowledge on the relation between starting points on a shape and its transformed is supposed.
- •
The method works very well on scenes with multiple objects.
- •
The method is easily extended to three-dimensional space.
The remainder of this paper is organized as follows. In Section 2 we will present the FS decomposition and the FS coefficients. The investigated affine invariant descriptors and motion estimation algorithm are described in Section 3. Section 4 gives a set of affine invariant descriptors for scene with multiple objects. In Section 5 the complexity of the FS method is discussed.
Section snippets
Fourier series decomposition
Given a periodic function f, with a period T=2π/w, the trigonometric serieswhereare called FS of the function f.
Another equivalent decomposition into FS iswhere cn=(1/2)(an−jbn).
In the following the coefficients and cn are called FS coefficients.
The proposed method is called FS method because it uses these FS coefficients as we will see in the remainder of the paper.
Representation
Any set of points can be described in the real space or in the complex plane. Using these two representations, the set of points X̃ will be said to be similar to the set of points X, if it can be mapped into X by a composition of an affine transformation A, and a translation B:where X and , A is a 2×2 matrix, and B is a 2-vector representing translation.
Parameterization
It is well known that there are different parameterizations to represent a given set of points. The familiar arc length one
Multiple objects
The invariant descriptors given in (3) cannot handle the possibility that some curves constituting the whole shape remain unchanged or are subject to different affine transformations.
Solution, to that problem, is presented through the use of a new set of affine invariants given bywhere
In fact, if a part of X undergoes an affine transform A1 and another part undergoes another affine transform
Complexity of the FS method
The algorithm contains at least two steps:
- 1.
The parameterization which needs the computation of first and second derivatives that are approximated by convolution with a Gaussian kernel of width σ=1 and then the computation ofand X(τ) for all t. This step is not specific to the algorithm.
- 2.
The computation of the needed Fourier coefficients:
- ◦
We use a normal formula to compute Fourier transform: We have to compute first the cosine and sine functions for each sample if the sampling
- ◦
Conclusion and future work
In this paper, we have proposed the use of one-dimensional Fourier Series coefficients to derive affine invariant descriptors. We have also discussed a motion estimation algorithm based on the minimization of a distance defined between FS coefficients. The obtained experimental results from synthetic and real shapes demonstrate the validity of our approach and its applicability.
In future, and since the vector form for the coefficients an,bn could easily be extended to more than two dimensions,
References (11)
- et al.
Shape distance for contour tracking and motion estimation
Pattern Recognition
(1999) - et al.
Application of affine-invariant Fourier descriptors to recognition of 3-D objects
IEEE, PAMI
(1990) - et al.
Affine invariant contour tracking with automatic control of spatio temporal scale
Image sequence analysis
(1981)Recognition of planar shapes under affine distortion
Internat. J. Comput. Vision
(1995)
Cited by (23)
Multi-scale contour flexibility shape signature for Fourier descriptor
2015, Journal of Visual Communication and Image RepresentationCitation Excerpt :In addition, these approaches are more robust to noise since the dominant features are those that persist across scales. There are many contour-based multi-scale description techniques such as curvature scale space (CSS) [14], multi-scale convexity concavity (MCC) [15], triangle area representation (TAR) [16]. The CSS image consists of several arch-shape contours representing the inflection points of the shape as it is smoothed.
Invariant curvature-based Fourier shape descriptors
2012, Journal of Visual Communication and Image RepresentationCitation Excerpt :Arbter et al. [43,44] have used a complex mathematical analysis and proposed a set of normalized descriptors which are invariant under any affine transformation. Oirrak et al. [45] have also used one-dimensional Fourier series coefficients to derive affine invariant descriptors. Zhang and Lu [41] have shown that although the affine Fourier descriptor [43] was proposed as a way to target affined shape distortion, it does not perform well on the standard affine invariance retrieval set of the MPEG-7 database because the affine Fourier descriptors are designed to work on a polygonal shape under affine transformation and are not designed for a non-rigid shape [41].
A 3-D Search engine based on Fourier series
2010, Computer Vision and Image UnderstandingFarthest point distance: A new shape signature for Fourier descriptors
2009, Signal Processing: Image CommunicationA new scheme for extraction of affine invariant descriptor and affine motion estimation based on independent component analysis
2005, Pattern Recognition LettersAffine invariant descriptors for color images using Fourier series
2003, Pattern Recognition LettersCitation Excerpt :Although some expression seems to be same in the two papers especially in the definition of FS decomposition and motion estimation, however the complex coefficients cn are different in the computation manner used, especially in the parameterization step. In (El Oirrak et al., 2002) we dispose only of shape represented by their contours or exterior profile i.e. bi-dimensional curve (2D), while in this case the whole color matrix image is considered. Besides the (x,y) coordinates in (El Oirrak et al., 2002) are the position of a point contour while in this case are the pixel position of an image color.