AUTOMATIC FLAMINGO DETECTION USING A MULTIPLE BIRTH AND DEATH PROCESS Stig Descamps1 , Xavier Descombes1 , Arnaud B chet2

, Josiane Zerubia1 e Equipe-Projet Ariana, INRIA/I3S 2004 route des Lucioles, BP 93, 06902 Sophia Antipolis Cedex (France) E-mail : [email protected] 2 Tour du Valat Le Sambuc, 13200 Arles (France)
ABSTRACT Here we present a new approach to automatically detect and count breeding Greater Flamingos (Phoenicopterus Roseus) on aerial photographs of their colonies. We consider a stochastic approach based on object processes also called marked point processes. The objects represent amingos which are dened as ellipses. We formulate a Gibbs density, associated with the marked point process of ellipses, which is dened w.r.t a Poisson measure. Thus, the issue is reduced to an energy minimization, where the energy is composed of a regularization term (prior density), which introduces some constraints on the objects and their interactions, and a data term, which links the objects to the features to be extracted in the image. Then, we sample the process to extract the conguration of objects minimizing the energy by a new and fast birth-and-death dynamics, leading to the total number of birds. This approach gives counts with good precision compared to manual counts. Additionally, this approach does not need image pre-processing or supervision of the extraction by an operator thus considerably reducing the overall processing time required to get the estimate. Index Terms bird colony, Object extraction, marked point processes, stochastic modeling, birth/death dynamics. 1. INTRODUCTION During breeding seasons, amingos gather in a large colony. Thus, specialists take advantage of it to asset the number of amingos in the colony. Since the 60s, several techniques have been developed to estimate the number of amingos from aerial images. Most of them are highly supervised. They are based on an expert counting on some predened small areas. The total number of amingos is then estimated from an interpolation procedure. This methodology is therefore time consuming for the experts and lacks precision. General sofware for object detection based on classical image processing tools such as mathematical morphology [1] or template matching [2] appear to be unsatisfactory. A more dedicated approach embedding a geometric model of the amingos and some constraints on their spatial repartition is required. We propose in this paper, a new method for automatically estimating the size of amingo populations based on object processes. We consider an ellipse as our reference object to model amingos. Indeed, on aerial images, amingos look like ellipses. Greater amingos are mainly covered with white plumage. This fact gives a feature to evaluate, with the Bhattacharrya distance, the contrast between the background and amingos themselves. The density associated with the marked point process of ellipses is dened with respect to the Poisson measure. Moreover, by formulating the model as a Gibbs density, we reduce the problem to an energy minimization. This energy is decomposed into a data term to locate amingo on the image, and a prior term which introduces constraints between the objects of the conguration. 2. A MARKED POINT PROCESS MODEL 2.1. Denition and notation We model aerial images as composed of amingos whose positions and attributes are some realization of a marked point process X, see [3] for more details. X is also a random variable whose realizations are random congurations of objects belonging to a set space = P M, where P is the position space, and M the space of the marks. We note the space of all congurations of a nite number of objects. The probability distribution PX (.) of the stochastic process is uniformly continuous with respect to the Poisson measure (.) of intensity (.) on . Then, by using the Gibbs energy formulation of the process density, we dene an energy U (x) as : PX (dx) = 1 exp(U (x))(dx) Z (1)
1

where Z is a normalizing constant. This energy will be minimized on by the amingo extraction. It takes into account

the interactions between the geometric objects (the prior energy Up (x)), and the way they t to the data (the data energy Ud (x)) : U (x) = Up (x) + Ud (x) 2.2. Objects of interest The 2D model, used to extract amingos, consists of a marked point process of ellipses. The associated set space is : = P M = [0, XM ][0, YM ][am , aM ][bm , bM ][0, [ where XM and YM are respectively the width and the length of the image I, (am ,aM ) and (bm ,bM ) respectively the minimum and the maximum semimajor axis and semiminor axis, and [0, [ the orientation of the objects. 2.3. Prior energy As we aim at detecting individuals in dense populations, we model amingos as possibly slightly overlapping ellipses xi r xj . Then, the prior energy Up (x) that introduces interactions, penalizes congurations according to the overlapping objects area, see [4] for more details: Up (x) = p
xi x

the background. To evaluate the contrast between the ellipses and the background, we calculate the Bhattacharya distance dB (x, F(x)) between the reectance distributions of the object and its boundary as follows: dB (x, F(x)) = 1 h1 (x)h2 (x)dx, (4)

(2)

where h1 (x) (resp. h2 (x)) is the empirical distribution of the pixels belonging to x (resp. F(x)). The data energy Ud (x) associated with the object x is then given by: Ud (x) = Qd (dB (x, F(x))) (5)

where Qd (dB ) [1, 1] is a quality function which gives some positive value to small Bhattacharya distance (weakly contrasted objects) and negative value (well located objects) otherwise. (1 dB ) d0 d exp( dBD 0 ) 1 si dB < d0 si dB d0

Qd (dB ) =

(6)

where D is a scale parameter calibrated to 100.

maxxj r xi A(xi , xj )

(3)

where A(xi , xj ) [0, 1] is an overlapping coefcient and p is a weight which ponders the repulsion between the objects of the process. Each object is penalized depending on the maximal overlapping it exhibits with neighboring ellipses. Fig. 2. An ellipse and its boundary

2.5. Parameter estimation The data term involves a parameter d0 which can be interprated as the mean value of the contrast between a amingo and its neighborhood. We propose to estimate it by considering a weighted histogram. For each pixel s of the image I, we compute the data energy s Ud (c) corresponding to a disc whose radius is equal to the minimal size of amingos to be extracted in the image. This map will be used for estimating the value of d0 . It requires an initial value which has been calibrated to d0 = 10. We derive a predetection map by computing the following rate:
t s maxtI Ud (c) Ud (c) t (c) min t maxtI Ud tI Ud (c)

Fig. 1. Overlapping ellipses

2.4. Data energy In amingo populations, each amingo can be modeled as a bright ellipse surrounded by a darker background. Thus, we dene the boundary of an ellipse F(x) as the set of P contained between the given ellipse x = (p, k), where k = (a, b, ) are the marks, and a concentric one x = (p, k ), with k = (a + , b + , ). This boundary will stand for

s I, b(s) = 1 + 9

(7)

The higher b(s), the more likely a amingo is in s.

Usually, amingo color is not homogeneous within a given image because of the focal of the camera. Thus, we present a local parameter estimation method. To compute d0 , we need to estimate the mean of amingo color in small squared area of the initial image. This estimation is decomposed in three steps: the construction of a weighted histogram of the initial image, a rst estimation of the parameters and a nal ltering of the parameter previously estimated over the image: - The weighted histogram is simply obtained by constructing the histogram of the initial image whose pixels are weighted by the respective pixels of the predetection map. - For each squared region, we extract the mean of the colour in the considered region by detecting the maximum of the weighted histogram. - Once done for every squared region of the image, we compute a spatial ltering of the previous parameters to correct biased estimates in low density areas. d0 is then computed from this amingo mean radiometry estimation. 3. BIRTH AND DEATH DYNAMICS For optimizing the model, we consider a simulated annealing based on a birth and death process. This process has rst been proposed in [5], where the proof of the convergence is given in the case of disks, the generalization to ellipses being straightforward. The algorithm simulating the process is dened as follows: Main program: initialize the inverse temperature parameter = 0 = 50 and the discretization step = 0 = 20000 and alternate birth and death steps Birth step: for each s S, if no object is already alive, we add an object in s with probability B(s) where B(s) is derived from the predetection map: s I, B(s) = zb(s) tI b(t) (8)

Convergence test: if the process has not converged, decrease the temperature and the discretization step by a given factor and go back to the birth step. The convergence is obtained when all the objects added during the birth step, and only these ones, have been killed during the death step. 4. RESULTS AND DISCUSSIONS We present in gure 4 and gure 3 two results on real aerial images. Computation took a couple of minutes for each result, which depends on both the number of objects to be extracted and also on the size of the image.

where z is a parameter of the process. Sorting step: once the birth step is nished, we s compute the data term Ud (c) of the current conguration objects uc . Then, we sort them, from increasing, according to their data energy. Death step: for each object taken in this order, we compute the death rate as follows: d(uc ) = a (uc ) 1 + a (uc ) (9)

Fig. 3. Top: initial population image. Bottom: Sample of the extraction To evaluate these results, we have computed the detection rate compared to a manual extraction, executed by an expert.The different datasets have been divided in three classes

where a (uc ) = exp(U (uc )), then the object uc dies with probability d(uc ).

Fig. 4. Sample of the extraction on a Mauritanian colony (only the ellipse center is pointed)

depending on the image quality and the population density. The false alarm rate is always negligible and the detection rate is given in gure 5. In most cases, the desagreement between the expert and our detection corresponds to ambiguous cases where no decision can reasonably be taken from the image alone.

Fig. 5. Resulting statistics 7. REFERENCES [1] P. Soille, Morphological Image Analysis, Springer Verlag, 1999. [2] R. Pollock, Individual Tree Recognition based on Synthetic Tree Crowns Image Model, in Proc. of the International Forum on Automated Interpretation of High Spatial Resolution Digital Imagery for For estry, D.A. Hill and D.G. Leckie, Eds., February 1998, pp. 2534. [3] M.N.M. Van Lieshout, Markov point processes and their applications, Imperial College Press, London, 2000. [4] G. Perrin, X. Descombes, and J. Zerubia, 2d and 3d vegetation resource parameters assessment using marked point processes, in Proc. International Conference on Pattern Recognition (ICPR), Hong-Kong, ao t 2006. u [5] X. Descombes, R. Minlos, and E. Zhizhina, Object extraction using a stochastic birth-and-death dynamics in continuum, in Research Report No 6135, 2007.

5. CONCLUSION In this paper, we proposed an algorithm to automatically extract amingo populations from aerial images. Based on a stochastic geometry approach, we have shown the efciency of the detection on low resolution images where it is even tricky for the human eye to distinguish amingos between themselves. The automatic paramater estimation gives a major advantage over other current techniques of detection because no preprocessing is needed. Futhermore, execution time requirements are reasonable to obtain amingo population extractions because of the proposed birth and death process which outperform classical RJMCMC schemes [4].

6. ACKNOWLEDGEMENTS The authors would like to thank Antoine Arnaud from the Tour du Valat for his help in validating the results by an expert. This work has been partially supported by the COLOR Flamant funded by INRIA Sophia Antipolis.