Research On Multiscale Markov Random Fields For Image Segmentation

Image segmentation is the basis for many high level computer vision tasks, such as object identification, motion analysis and scene reconstruction. However, in the real-world applications in mechanical engineering, the a priori information of the image at fine scale is not reliable enough for classification because of nonuniform illumination, shading and background occlusions. Multiscale reconstruction provides an alternative method to cope with environmental changes across spatial locations and time, largely due to the higher the scale, the more reliable is the classification. The problem that arises is how to integrate the classification information residing in each scale in a segmentation framework. Consequently, multiscale image segmentation under dynamic background is a more challenging work than static background ones.In this sense, more robust multiscale fusion is required by dynamic image segmentation algorithms to allow an accurate separation between foreground and background. On the other hand, consistent segmentation result with different initial conditions is also required. Therefore, key issues in multiscale dynamic image segmentation problem, like the accuracy and stability of the inferred classification solutions against small variations in the background scene and initial parameter sets, namely, the robustness and consistency abilities, have to be addressed.Many multiscale approaches have been investigated, among which wavelet-domain hidden Markov trees are perhaps the most famous and widely used methods and have been successfully applied to image segmentation. The key advantage of these approaches is that they are capable of performing communication both intra and inter scales; that is, they allow the fusion of regional and boundary information by using robust higher scale classification likelihoods and accurate fine scale classification likelihoods, respectively. For this reason, WHMT-based approaches have also been exploited in numerous computer vision tasks such as infrared object classification, steam recognition, edge detection, writer identification and, more recently, texture retrieval.Although the previous image segmentation approaches based on WHMT multiscale fusion are successful to some extent, there are still some open problems. Generally, most of the WHMT-based methods have exploited the Expectation-Maximization(EM) algorithm to compute the optimal estimates of the multiscale parameters that maximize the likelihood function. However, the main problem with the EM algorithm is that it is a "greedy" method that converges to a local maximum on the log-likelihood surface. In fact, in our previous works on WHMT-based image segmentation, we find that the obvious perturbations of segmentation results occur due to different initial conditions of the WHMT model. On the other hand, dynamic programming algorithm allows optimal path states estimation along WHMT for efficient coding of object boundary, but could not fusion regional information at the same time. More importantly, the relaxation of unary and pairwise marginal constraints during energy minimization process is not reasonable. Following these observations, in this paper, four aspects are focused to solve the key issues in multiscale Markov random field image segmentation.1) Finest scale Markov random field image segmentation. This section will focus on Markov random field modeling at the finest scale. On this basis, research on the theory of labeling field and observation field theory as well as formulating image segmentation as maximum a posteriori estimation of labeling field. Moreover, adopt iterative algorithm for the purpose of image segmentation at the finest scale. Apply the finest scale Markov random field algorithm for tobacco target segmentation under dynamic backgrounds. Compare the segmentation results to that of Otsu's threshold segmentation method and making evaluations and discussions.2) Image segmentation through wavelet-domain multiscale hidden Markov random field. Research on the methods of multiscale modeling via 2D discrete wavelet transform. On this basis, the interscale modeling and intrascale communication method of the multiscale image model. In particular, modeling the statistical distribution through Gauss mixture at each scale, and modeling the depending relation between different wavelet coefficients across scales by hidden Markov tree. Moreover, research on how to fusion a priori information of the image during multiscale segmentation process as well as the maximum expectation method for estimating statistical classification likelihood of multiscale pixels. Apply the algorithm for noisy tobacco image segmentation under dynamic backgrounds to verify this proposed method.3) Multiscale Markov random field segmentation through graph cut. This chapter focuses on computing multiscale classification likelihood information through expectation maximization algorithm as well as modeling multiscale energy function by multiscale classification likelihoods. The key problem is how to fusion multiscale regional and boundary information into the energy function. Then graph cut based energy minimization method is employed for the purpose of approximate global optimization of the multiscale energy function. Tobacco image segmentation under dynamic backgroud, again, is adopted for evaluating the proposed method.4) Multiscale Markov field image segmentation through unconstrained convex optimization. This chapter focuses on global optimization methods of multiscale energy function. For this point, the modeling method of convex multiscale energy function is researched and the convexity of which is proved. In this sense, the image segmentation problem is cast into convex optimization problem. In addition, pseudomarginal constraints are considered during energy minimization process and the optimization problem is thus unconstrained. Hierarchical belief propagation algorithm is employed for unconstrained convex optimization estimation. Tobacco image segmentation under dynamic backgroud, again, is adopted for evaluating the proposed method.This paper aims to explore the multiscale image classification likelihood robust fusion algorithm under dynamic backgrounds, for the purpose of unconsistency difficulties in multiscale hidden Markov random field classification likelihood information fusion, as well as providing theoretical basis for multiscale Markov random field modeling and global optimization.