Sub-pixel mapping of abundance maps by contour minimization ; application to hyperspectral imaging and to binary image interpolation. C. Louchet MAPMO, UMR CNRS 7349 & Fédération Denis Poisson FR CNRS 2964 Université d’Orléans cecile.louchet@univ-orleans.fr Résumé Nous proposons une méthode d’estimation d’une image d’étiquettes à partir de la proportion de chaque étiquette sur des blocs disjoints. Nous l’appliquons à l’interprétation de cartes d’abondance données par le démélange d’images hyperspectrales et au zoom d’images binaires. Mots Clef Cartographie sous-pixellique, cartes d’abondance, imagerie hyperspectrale, zoom d’image binaire, problème inverse régularisé. Abstract We propose a method that estimates a label image from the proportion of the labels on disjointed blocks. We apply it to the interpretation of abundance maps provided by hyperspectral image unmixing and to binary image zooming. Keywords Sub-pixel mapping, abundance maps, hyperspectral imaging, binary image zoom, regularized inverse problem. 1 Introduction In hyperspectral imaging, hundreds of images are acquired of the same scene, each corresponding to a narrow range of spectrum, sometimes at the price of a low spatial resolution. The composition analysis of the scene requires the knowledge of the chemical pure elements (any chemical material as soil, grass, concrete...) present in the scene along with their spectra. From this information, spectral unmixing [1] builds abundance maps of the scene : each pixel is assigned with the proportion of each pure element. These abundance maps are very informative but in the case where the pure elements are physically separated in the scene (no transparency) and where the spatial resolution is low, the precise location of each pure element inside a pixel remains unknown and is interesting to guess from the abundances. Furthermore, if the precise location of each element is well estimated, the visualization of the abundance can be done through a unique image, instead of many images (the number of pure elements) as it is usually done. In this paper, we propose a method that estimates a label image from its abundance maps on disjointed blocks. In Section 2, we show that the operator that computes the abundances of a label image on its blocks is highly noninvertible, and propose to regularize the associated inverse problem by selecting the most regular image (the image that minimizes the contours length) obeying to the data constraint. In Section 3, we propose a simple simulated annealing algorithm able to solve the minimization problem. In Section 4, we apply the method to spectral abundance interpretation and to binary image reconstruction from a zoomed-out version, before concluding. 2 The variational model We aim at recovering a label image from the proportion of its labels on each block. We focus on images defined on a finite set Ω divided into a partition of NB blocks (Bn )0≤n≤NB −1 , with values in a finite set of labels L = {0, . . . , L − 1}. For each image u ∈ LΩ and each 0 ≤ n ≤ n ) where each NB −1, we consider the L-uple (θ0n , . . . , θL−1 n θl = #{x ∈ Bn : u(x) = l}/#Bn is the proportion of label l in the block Bn of u. We denote by A : LΩ → (∆L−1 )NB the operator that maps an image u to these NB elements of the (L − 1)-dimensional simplex. The aim is to invert A, but A is highly non-injective because many label configurations on a block can have the same quantity of each label (for each block, this number of configurations is the multinomial coefficient #Bn !/(k0 ! . . . kL−1 !) with kl = θl #Bn ). Among all these configurations we want to favor those with short contours, measured by the Potts potential X ∀u ∈ LΩ , E(u) = 1u(x)6=u(y) , x∼y where ∼ is a neighborhood system, typically the 4-th or the 8-th order neighborhood system, with specific boundary conditions. Finally, given a set v of abundance maps, we propose to solve the constrained minimization problem minimize E(u) among all u such that Au = v. (P) The analysis of Problem (P) is difficult. Firstly, the existence of a solution is guaranteed, but the uniqueness is not. Indeed, if L = 2 (binary image) and Ω = {0, . . . , 2NB −1} with Bn = {2n, 2n + 1}, if v contains only (0.5, 0.5) proportions of the labels, then (P) has 2 solutions (the one beginning with 0, the other beginning with 1). Secondly, the solutions are very sensitive to the boundary conditions of the neighborhood system. For example, the previous problem with NB odd and periodic boundary conditions has many more than 2 solutions. Here all solutions to the problem will be considered acceptable. But the formulation of Problem (P) raises interesting theoretical questions. Indeed, it is straightforward to generalize it to a continuous domain Ω. In this case, the analysis of its solutions (e.g. the regularity of the contours, the angles at triple points, the number of connected components, symmetry breakings, ...) is far beyond the scope of this paper but is probably linked to the Mumford-Shah model [3]. 3 Simulated annealing algorithm Problem (P) is a combinatorial problem that we solve using a simulated annealing algorithm. The solution is initialized with any image u0 in the constraint, i.e. satisfying Au0 = v. For example, if vn = (θl )0≤l≤L−1 and if the pixels of Bn are numbered, u0 can be assigned label 0 on the k0 first pixels of Bn , label 1 on the k1 following pixels, ..., label L − 1 on the kL−1 last pixels, or any permutation of this order, where for each l, kl = #Bn θl . From the image uN at iteration N , we propose a new image ũ by choosing a random (uniform distribution) block in uN , and 2 random pixels inside this block, and exchanging their labels. This proposal image has the advantage of being easy to implement and of remaining in the constraint. Furthermore, the underlying graph structure is connected, as any permutation is a composition of transpositions. If ∆E := E(ũ) − E(uN ) < 0, the proposal image is accepted, i.e. uN +1 is set to ũ. Otherwise, it is accepted with probability exp(−∆E/T N ), and rejected otherwise (and uN +1 = uN ). Here the temperature cooling scheme (T N )N ∈N is slowly decreasing toward 0. The theory of simulated annealing states that taking (T N )N ∈N sufficiently slow allows the sequence uN to converge in distribution to a uniform distribution on the solutions of (P). 4 Experiments and discussion Application 1. Four small abundance maps were synthesized such that the sum of the superposed maps was a constant image equal to 1. Each pixel of abundance maps is considered as a z × z block (in our experiment z = 32) and Figure 1 shows how the method described above organizes each block by distributing the elements in the z × z pixels in the most grouped possible way. The initialization was a random image in the constraint, the temperature was T N = 10 × 0.999(N/#Ω) , and weighted 8-order neighborhoods were chosen along with periodic boundary conditions. Application 2. Suppose a binary (0-1) image has been zoomed out using a square block averaging procedure [2]. The obtained image is gray-valued and each gray level equals the proportion of white in the corresponding block. Our method is able to estimate the original image from its 4 abundance maps result F IGURE 1 – An interpretation of 4 abundance maps. The black in the right image represents locations of element 0 (top-left abundance map) ; dark gray represents element 1 (top-right abundance map) ; and so on. The result is regular and contains some isolated grains. It represents a most likely ground composition given the abundance maps. data initialization result F IGURE 2 – Reconstruction of a binary image from a zoomed-out version (16 × 16 zoom). zoomed-out version. Figure 2 illustrates this on a 16 × 16 zoom of a letter “a” with T N = 10 × 0.99(N/#Ω) . Discussion. If the algorithm is intrinsically slow and would deserve deeper insight, on the other hand the model appears to be interesting both from a practical and a theoretical viewpoint. A penalized version would be useful to be more robust to noise in abundance data. Acknowledgement The author wants to thank Laurent Delsol (Université d’Orléans) for their fruitful discussions. Références [1] J. M. Bioucas-Dias, A. Plaza, N. Dobigeon, M. Parente, Q. Du, P. Gader and J. Chanussot. Hyperspectral unmixing overview : Geometrical, statistical, and sparse regression-based approaches. IEEE J. Selected Topics in Applied Earth Observations and Remote Sensing, 5(2) : 354-379, 2012. [2] F. Guichard and F. Malgouyres, Total variation based interpolation. Proc. of Eusipco, 3 : 1741-1744, 1998. [3] D. Mumford and J. Shah, Optimal approximations by piecewise smooth functions and associated variational problems, Commun. Pure Appl. Math., 42(4), 1989.