PROJECTION BASED APPROACH FOR REFLECTION SYMMETRY DETECTION Thanh Phuong Nguyen To cite this version: Thanh Phuong Nguyen. PROJECTION BASED APPROACH FOR REFLECTION SYMMETRY DETECTION. ICIP, Sep 2019, Taipei, Taiwan. hal-02133544 HAL Id: hal-02133544 https://hal.archives-ouvertes.fr/hal-02133544 Submitted on 18 May 2019 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. PROJECTION BASED APPROACH FOR REFLECTION SYMMETRY DETECTION Thanh Phuong Nguyen Universite´ de Toulon, CNRS, LIS, UMR 7020, 83957 La Garde, France Aix-Marseille Universite,´ CNRS, ENSAM, LIS, UMR 7020, 13397 Marseille, France ABSTRACT tries. They also introduced an another work [13] for detecting approximated reflection symmetry in a set of points using op- A novel method for reflection symmetry detection is ad- timization on manifold. Kazhdan [14] used Fourier methods dressed using a projection-based approach that allows to deal to detect and compute reflective symmetries. In [15], Der- effectively with additional noise, non-linear deformations, rode and Ghorbel applied Fourrier-Melin transform for rota- and composed shapes that are not evident for classic contour- tion and reflection symmetry estimation. Kiryati and Gofman based approaches. A new symmetry measure is also proposed [16] converted reflection symmetry detection into a global op- to measure how good the detected symmetry is. Experiments timization problem. Cornelius et al. [17] used local affine validate the interest of our proposed method. frames (LAFs) constructed on maximally stable extremal re- Index Terms— Reflection symmetry, Radon, R-transform gions to improve the detection of symmetric objects under perspective distortion. 1. INTRODUCTION We address in this paper a new method for reflection sym- metry detection using R-transform. Based on the projec- Symmetry is very popular in both artificial and natural scenes tion based approach, this can deal naturally with composed because most mand-made and biological objects have sym- shapes, additional noise, and non-linear deformations. metric properties. In addition, symmetric structures are im- portant visual features for human attention, therefore symme- 2. BASIC MATERIALS try detection plays an important role in computer vision. There are two main problems in symmetry detection. The Let us recall some basic materials of Radon transform and first one aims at detecting and measuring the rotational sym- R-transform [18]. Those will be used in the next section to metries in a shape. Lin et al. [1] proposed fold-invariant propose a new method for reflection symmetry detection. shape-specific points for detecting the orientations of rota- tionally symmetric shapes. Cornelius and Loy [2] detected 2.1. Radon transform planar rotational symmetry under affine projection. Prasad 2 2 and Davis [3] localized multiple rotational symmetries in nat- Let f 2 R be a 2D function and L(θ; ρ) = fx 2 R j 2 ural images using gradient magnitude field. Loy and Eklunhd x · n(θ) = ρg be a straight line in R , where θ is the angle [4] grouped symmetric pairs of feature points and character- L makes with the y axis, n(θ) = (cos θ; sin θ), and ρ is the izing the symmetries presented in an image. Flusser and Suk radial distance from the origin to L. The Radon transform [5] introduced a new set of invariant moments for recognition [19] of f, denoted as Rf , is a functional defined on the space of objects having n-fold rotation symmetry. Yip introduced of lines L(θ; ρ) by calculating the line integral along each line different methods using Hough transform [6] or Fourier de- as follows. scriptor [7] for the detection of rotational symmetry. Z The second one groups the methods for reflection symme- Rf (θ; ρ) = f(x) δ(ρ − x · n(θ)) dx (1) try detection. Ogawa [8] used Hough transform to detect axis of symmetry in shapes of line drawing. Yip [9] then devel- In shape analysis, the function f is constrained to take value oped this approach to deal with both reflection symmetry and 1 if x 2 D and 0 otherwise, where D is the domain of the skew-symmetry. Lei and Wong [10] also used Hough trans- binary shape represented by f (see Figure 1). form for detecting and recovering the pose of a reflection and ( rotational symmetry from a single weak perspective image. 1 if (x) 2 D f(x) = (2) Cornelius and Loy [11] detected bilateral symmetry in im- 0 otherwise ages under perspective projection by matching pairs of sym- metric features. Nagar and Raman [12] proposed an energy Radon transform is robust to additive noise and has some minimization approach to detect multiple reflection symme- interesting geometric properties [19] which are the base to 3. REFLECTION SYMMETRY DETECTION 3.1. R-transform and reflection symmetry Let us consider an arbitrary shape D. It has reflection sym- metry if there is at least one line which splits the shape in half so that one side is the mirror image of the other. For simplic- ity, D is called reflectionally symmetric in direction θ if it is reflectionally symmetric and contains an axis of symmetry in Fig. 1. Radon transform of a function f(x; y) that direction. As we have pointed out in Section 2.2, R-transform has #10 7 Original shape been proven to be robust against additive noise, nonlinear de- 3.5 Rotated shape formations [18]. It is invariant against similarity transforms: 3 translation, and rotation. In addition, it is also invariant ) 3 ( 2 2.5 against scaling if the transform is normalized by a scaling f R R 2 factor. Those beneficial properties suggest that -transform can be served as an useful tool for shape analysis. 1.5 We propose in this section an another interesting prop- 1 0 20 40 60 80 100 120 140 160 180 erty of R-transform for reflection symmetry detection. The 3 main idea is to convert the problem of detecting and measur- (a) Input (b) Rotated (c) R-transforms ing reflection symmetry of an arbitrary shape D into measur- ing the reflection symmetry in its R-transform. Due to [18], Fig. 2. Illustration of R-transforms [18]. Rf2(θ) is periodical of period π with respect to θ, it is suffi- cient to consider R-transform only on the set of projections developpe an effective shape signature, namely R-transform Θ = [0; π) or Θ = 00; 10;:::; 1790 for relection symmetry [18], presented in Section 2.2. detection. We also introduce the following notion that will be used Figure 2 shows the R-transforms of two reflection- θ latter. For each projection direction θ, the radial distances ρ1 ally symmetric shapes. Similarly, Figures 3.b presents R- θ θ and ρ2 are respectively defined as ρ1 = inf fρ j RD(θ; ρ) > transform of a synthetic shape which is an isosceles triangle θ 0g and ρ2 = sup fρ j RD(θ; ρ) > 0g. The “profile” (or in Figure 3.a. It should be noted that the studied shape con- θ Radon projection) of D in the direction θ, denoted as CD, tains evidently one axis of reflection symmetry in direction θ θ θ θ 0 is defined as RD(θ; ρ1:ρ2). More precisely, CD(ρ − ρ1) = O while its R-transform has two reflection symmetries at θ θ π RD(θ; ρ); 8ρ 2 [ρ ; ρ ]. 1 2 directions 0 and 2 . In addition, Figure 3.d shows the Radon projection of D in two above directions. We could make some 2.2. R-transform important following remarks from those Figures. Tabbone et al. [18] introduced a transform, called R- •R-transform contains rich information about rotational transform, for an effective shape representation as follows. symmetric properties of shapes. If D contains reflec- Z +1 tion symmetry, its R-transform is also reflectionally 2 Rf2(θ) = Rf (θ; ρ)dρ (3) symmetric. −∞ They have shown the following properties of this transform. • Each detected axis of reflection symmetry of D in di- rection θ0 leads to 2 reflection symmetries which sepa- • Periodicity: Rf2 is periodical with period of π. π rate by an interval of 2 . Those correspond generally to π • Rotation: A rotation of the image by an angle θ0 leads two orthogonal directions: θ0 and θ0 + 2 . This comes to a circular shift of Rf2 of θ0. from the fact that R-transform treats equally all projec- tion values in each direction. • Translation: Rf2 is invariant against translation • Between two above detected directions, the profile of • Scaling: A scaling of f implies only a scaling in the Radon projection (Rf (θ; ρ)) is also reflectionally sym- amplitude of Rf2 π metric in the direction θ0. For the direction θ0 + 2 , it is Please refer to [18] for more illustrations about the ro- reflectionally symmetric if and only if D contains also π bustness of R-transform against similarity transforms (trans- reflection symmetry in direction θ0 + 2 . Indeed, con- lation, rotation, scaling), and non-linear deformations. These trariwise to R-transform, Radon projection (Rf (θ; ρ)) properties make R-transform useful for shape analysis. It is consider the distribution of projection values for each also a base for different shape descriptors [20, 21, 22]. direction. Rf (θ; ρ) presents the distribution of two symmetric parts of D in direction θ0 of symmetric axis, Algorithm 1 Reflection symmetry detection of a shape D.