Fast global SA(2,R) shape registration based on invertible invariant descriptor
Graphical abstract
Introduction
It is well known that it is difficult to compare images when they are acquired at different times, sensors and poses. The obtained images need to be aligned in such a way that differences can be detected. A similar problem appears when looking for a prototype or a model in an image. These problems and several variations can be solved by image registration techniques. The image registration algorithm consists on the determination of the geometric transformation in order that the different points in one image find their correspondence in another image. An optimal transformation describing such alignment is determined according to a given criterion. The most registration methods proposed in literature is related to a rigid deformation which is modeled by a rotation followed by a translation. Such transformation is often referred to as Euclidean. However, in the case of three-dimensional object reconstruction from two of its projections, the transformation linking these two projections is often modeled by planar homography. The matching of curves up to a planar homography, needs planar equi-projective re-parametrization curves which requires at least five numerical derivations [1]. This induces errors of approximation equivalent to the value order of the quantities to estimate. For this reason, it is often proposed to replace the homography by the associated affine transformation. In addition to 3D reconstruction, the need for image registration occurs in several practical problems with various computer vision and image analysis applications such as robot navigation [2], medical image matching [3] and face recognition [4], [5]...
In this work, we intend to introduce a fast global planar curve registration algorithm relatively to planar special affine group SA(2,R) and independent to the original parametrization. This result becomes possible thanks to the existence of an invertible and stable invariant set of descriptors which allows us to reconstruct the target curve from the set of invariant descriptors of a reference contour. The proposed algorithm consists on three steps. First, an affine curve re-parametrization is applied [6]. The second step consists on the computation of the invertible set of descriptors for the reference and target objects. Using these two sets of descriptors, pose parameters are estimated by applying geometric shift theorem relative to SA(2,R). The invariant of the reference curve and the pose parameters, allows us the alignment.
The remainder of the paper is organized as follows: Section 2 introduces the related work to our approach. Then, we recall the affine invariant descriptor and evaluate the shape reconstruction by solving many numeric problems in Section 3. In the next section, we give a description of the proposed method for shape Registration. In Section 5, we study the accuracy of the proposed approach using MPEG-7, MCD Kimia-99, Kimia216, ETH-80 and Swedish leaf database in the task of shape retrieval. Finally, the last section submits the conclusion.
Section snippets
Related work
In this section, we focus on works that serve to place this paper as relative to literature. The motion estimation was pioneered by the work of Persoon and Fu [7] in the case of 2D shapes. The shape registration was formulated as the minimization of the quadratic error between shape contours. In [8], the authors propose an optimal estimation of the transformation parameter based on Nyquist–Shannon theorem and B-spline approximation. In [9], an approach for shape registration was proposed for
Recall of the affine shape reconstruction description
This work aims to use the special planar affine transformation group instead of the projective group. This approximation is often valid if the depth variation of the object is negligible compared to the observation distance. In the following, we recall the affine shape reconstruction algorithm based on the complete and stable Invariant Descriptors defined in [53]. These invariants are computed on a normalized affine arc length obtained by re-parametrization procedure recalled by Spivak [6], and
The global affine shape registration based on invertible invariant descriptors
Toward the solution of this challenging problem of shape registration, in this section, we propose a global shape registration based on the invertible propriety of the invariant descriptors. Consider a normalized affine arc length parametrization of two planar and closed contour and correspond to the same special affine shape. We proof that the alignment of and is obtained by applying the construction formula to the input parameters and (invariant descriptor of ) and the
Experiments and results
In this section, we provide the recognition rates of the proposed algorithm and compare it with the state-of-the-art shape registration methods. The experiments on these datasets are carried on six popular benchmarks : (MPEG-7 dataset, Multiview Curve Dataset (MCD), ETH-80, Swedich leaf, Kimia-99 and Kimia-216 datasets). The results of the different methods for these datasets are gathered from their respective articles. As well as the comparisons are made under the same conditions. First, we
Analysis of complexity
Shapes descriptors computation and shape matching are two independent stages of the proposed method. The complexity of the proposed approach will be evaluated separately here. Assume having uniform sampled points (in the sense of the special affine arc length) for every shape’s contour. Firstly, we should start by the affine arc length re-parametrization for each contour, which needs the computation of first and second derivatives. This step is not specific to the introduced algorithm. Thus
Conclusion
In this paper, we have introduced a new fast global shape registration method relatively to the planar special affine transformation from the invertible of the complete and stable invariant set under 2D special affine transformations and original parametrization. Also, we evaluate and compare our algorithm to other existing methods under different shape variations... Experiments are carried out on several popular shapes benchmarks, including MPEG-7, MCD Kimia-99, Kimia216, ETH-80 and Swedish
CRediT authorship contribution statement
Sinda Elghoul: Conception, Ideas, Methodology study design, Software programming, Implementation and testing code, Validation analysis, Interpretation of data, Writing - original draft, Writing - review & editing. Faouzi Ghorbel: Ideas, Formal analysis, Mathematical... critical revision of paper, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (87)
- et al.
Multi-scale contour flexibility shape signature for fourier descriptor
J. Vis. Commun. Image Represent.
(2015) - et al.
A novel method for 2D nonrigid partial shape matching
Neurocomputing
(2018) - et al.
Invariant multi-scale descriptor for shape representation, matching and retrieva
Comput. Vis. Image Underst.
(2016) - et al.
Distance interior ratio: A new shape signature for 2D shape retrieval
Pattern Recognit. Lett.
(2016) - et al.
Multiscale fourier descriptor based on triangular features for shape retrieval
Signal Process., Image Commun.
(2019) - et al.
An almost complete curvature scale space representation: Euclidean case
Signal Process., Image Commun.
(2019) - et al.
Synthesized affine invariant function for 2D shape recognition
Pattern Recognit.
(2007) - et al.
An application of wavelet-based affine-invariant representation
Pattern Recognit. Lett.
(1995) - et al.
Non-rigid registration using distance functions
Comput. Vis. Image Underst.
(2003) - et al.
An affine invariant curve matching method for photo-identification of marine mammals
Pattern Recognit.
(2005)