Image identification and estimation using the maximum entropy principle
Introduction
Digital processing of noisy images has become an interesting field of investigations. Many authors have considered the problem of image estimation and modeling. To ensure that the estimation algorithms are efficient both in computation time and storage, recursive estimators have gained much attention Kaufman et al., 1983, Koch et al., 1995, Kadaba et al., 1998, Woods et al., 1987, Wu and Kundu, 1992. The optimal estimation of images corrupted by additive noise has been considered (see for example Jain, 1989, Rajala and de Figueiredo, 1981, Kuan et al., 1985, Nahi, 1972. The first requirement is that a model is available. In (Kaufman et al., 1983), recursive identification and estimation of images has been considered. An autoregressive (AR) model was used for image modeling. Different parameter identification algorithms were compared. The identified AR model is then used in a reduced update Kalman filter (RUKF) for image estimation. However, the authors assumed that the AR model and the measurement noise covariance were a priori known. Recently, in (Kadaba et al., 1998), the authors proposed a recursive estimation algorithm inspired from the RUKF developed in (Kaufman et al., 1983) for non-gaussian AR models. They supposed that the process noise density is known as in (Kaufman et al., 1983), that the model order, the measurement noise covariance were also a priori known.
In the present work, image identification and estimation are reconsidered in the same framework as in (Kaufman et al., 1983). The only difference is that no a priori information is assumed. We consider an AR model for the given noisy image. Then the model order, the model coefficients, the measurement noise covariance are determined through an exhaustive search such that the entropy of the estimation error reaches its maximum. This state corresponds to the minimum mean squared estimation error (MSE). Experiments on simulated and real noisy images are given for illustrating the effectiveness of the proposed method. To this effect, Section 2 presents the image model. Section 3 gives a review of the RUKF used in (Kaufman et al., 1983). Section 4 briefly reviews the estimation procedures used. The maximum entropy principle (MEP) is dealt with in Section 5. Finally, Section 6 presents and discusses the experimental results.
Section snippets
Image modeling
In establishing the basic fundamentals for mathematical image modeling, we will assume that any interesting image can be defined on mapping of pixel coordinates into density values over a discrete 2D square region consisting of an N×N regularly spaced lattice, i.e, with N2 different picture elements or pixels. A conventional raster type scan will be assumed in the following manner: left to right scan, advance one line, repeat. Thus, at any point in the picture, some points will be “in the
RUKF
In one-dimension, the Kalman filter offers an attractive solution to the linear filtering and prediction problem. The extension of one-dimensional Kalman filtering to the system defined by (1) and (4) involves an enormous amount of data storage and transfer due to the high dimension of the resulting state vector. Hence a straightforward extension is of limited success, and thus it becomes desirable to consider computationally effective approximations. Here one such approximation, the 2D RUKF as
Estimation procedures
Since implementation of the RUKF requires the coupling coefficient vector c and the process noise covariances Qw and Qv, we will briefly review the procedure for identifying Qw and c directly from the noisy observations. Details can be found in (Kaufman et al., 1983).
Estimation of the probability density function (p.d.f.)
The RUKF used in (Kaufman et al., 1983; Koch et al., 1995) was applied to images assuming that the model order θ(p,q) (Eq. (3)) is a priori known. In the following, we will address the question of the 2D AR model order determination. This determination is based on the MEP. As the latter is defined for a p.d.f., we will use the following formulation.
Let be the estimated image density function for a given model order θ. The estimation error is defined as being the difference between the
Experimental procedures and results
In order to test the identification of the best model parameters c calculated by the least-squares method with bias compensation and at the same time the procedure for the model order determination based on the MEP, two random homogeneous fields were initially generated. The first AR model of order θ(1,1) (RF1) considered in (Kaufman et al., 1983) is defined by the following parameters:with a model noise covariance Qw=91.The second random
Conclusion
Identification and recursive filtering based upon a reduced Kalman filter for noisy images have been considered. A 2D AR model was used for image modeling as in (Kaufman et al., 1983). The MEP was proposed as a solution to the case, where no a priori information is available on the AR model order, on the model and observation function noise covariance, which were assumed to be known in Kaufman et al., 1983, Kadaba et al., 1998. It is shown that the estimation and/or filtering error with maximum
References (25)
Modeling by shortest data description
Automatica
(1978)A new look at the statistical model identification
IEEE Trans. Automat. Control
(1974)- Aksasse, B., Badidi, L., Radouane, L., 1999. A rank test based approach to order estimation part_I: 2-D AR models...
- Aksasse, B., Radouane, L., 1999a. Two-dimensional autoregressive (2-D AR) model order estimation, IEEE Trans. Signal...
- Akssase, B., Radouane, L., 1999b. Two-dimensional autoregressive (2-D AR) model order estimation using ϕβ criterion....
- Bouzouba, K., Radouane, L., 1999. Two-dimensional autoregressive (2-D AR) model order estimation via the maximum...
- Cadzow, J.A., Baseghi, B., Hsu, T., 1983. Singular-value decomposition approach to time series modelling. In: Proc....
- Isaksson, A.J., 1994. Frequency domain accuracy of identified 2-D causal AR-models. IEEE Trans. Signal Process. 42 (2)...
Fundamental of Digital Image Processing
(1989)- Jaynes, E.T., 1982. On the rationale of maximum entropy methods. Proc. IEEE 70...
Cited by (10)
Automatic image segmentation based on PCNN with adaptive threshold time constant
2011, NeurocomputingAsymptotic behavior of RA-estimates in autoregressive 2D processes
2009, Journal of Statistical Planning and InferenceCitation Excerpt :In this case, the sample size is usually larger than the one available in filters and, frequently, structure and topology do not impose heavy requirements or constraints. In some cases, strong hypothesis are made on the laws governing the observed process (Allende and Pizarro, 2003; Brunelli and Messelodi, 1995; Bustos et al., 2002; Butler, 1998; Dryden et al., 2002; Van de Weijer and Van den Boomgaard, 2005); other approaches can be seen in the works by Bouzouba and Radouane (2000), Brandle et al. (2003), Nirel et al. (1998), Sim et al. (2004), Tohka et al. (2004), Xu (2005) and Zervakis et al. (1995). High-level image analysis, or vision, also benefits from the use of robust estimation techniques, as can be seen in Black and Rangarajan (1996), Black et al. (1997), Chen et al. (2003), Comport et al. (2006), Glendinning (1999), Gottardo et al. (2006), Hasler et al. (2003), Kim and Han (2006), Li et al. (1998), Meer et al. (1991), Mirza and Boyer (1993), Prastawa et al. (2004), Roth (2006), Singh et al. (2004), Stewart (1999), Torr and Zisserman (2000) and Wang and Suter (2004a, b).
Kalman Filter Employment in Image Processing
2020, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)A novel method of image segmentation based on PCNN with spiral optimization
2014, International Conference on Signal Processing Proceedings, ICSP