Dynamic Block-Based Parameter Estimation for MRF Classification of High-Resolution Images

Smoothing Parameter Estimation for Markov Random Field Classification of non-Gaussian Distribution Image

“…Due to the assumption of Gaussian class conditional densities, the value of ∆U ι αβ was defined as a Mahalanobis distance using the equal covariance matrix for all the classes in the case of (Tolpekin andStein, 2009, Li et al, 2012); or using the mean of the covariance of each pair of classes in the case of (Aghighi et al, 2014). However, this assumption may not be tenable for remotely sensed mixed pixels (Xu et al, 2005).…”

“…Markov random fields (MRFs) as undirected graphical models are common methods for incorporating both spectral and contextual information (Chen et al, 2010a). They are formulated as the minimization of an energy function which consists of spectral and spatial energy terms and an important term which plays a key controlling role known as smoothing parameter (Aghighi et al, 2014). Larger values of the smoothing parameter result in over-smoothed classified maps and too small values do not fully utilize the available spatial information (Tolpekin and Stein, 2009).…”

“…Moreover, both methods suffered from similar covariance assumptions for the classes. Thus, we proposed a robust smoothing parameter estimation framework to overcome their limitations (Aghighi et al, 2014). One of the limitations of the described methods is that they were developed based on the assumption of the Gaussian class conditional distribution (Aghighi et al, 2014), (Tolpekin and Stein, 2009) and (Li et al, 2012).…”

“…Thus, we proposed a robust smoothing parameter estimation framework to overcome their limitations (Aghighi et al, 2014). One of the limitations of the described methods is that they were developed based on the assumption of the Gaussian class conditional distribution (Aghighi et al, 2014), (Tolpekin and Stein, 2009) and (Li et al, 2012). Although the Gaussian distribution is widely applied in many of image labelling applications, this assumption may not be tenable for remotely sensed mixed pixels (Xu et al, 2005).…”

“…However, each image band may provide different or even conflicting information based on its wavelength. Therefore, some other methods such as fuzzy no-edge/edge function using Sobel mask (Tarabalka et al, 2010), edge probability map using Canny operator (Aghighi et al, 2014) and graduated increase edge penalty (Yu and Clausi, 2008) were introduced in previous studies.…”