Image steganalysis so far has dealt only with detection of a hidden message and estimation of some of its parameters (e.g., message length and secret key). To our knowledge, so far there is no steganalysis method that can estimate the hidden message itself. Our goal in this paper is to bridge this gap. We propose a steganalysis approach to estimate the hidden message based on a Bayesian framework, modeling the image as a Markov random field and exploiting the analogy between images and statistical mechanical systems. Message embedding in bit planes of an image is modelled as a binary symmetric channel. The theoretical framework is presented in detail. Experimental results are provided to support the theory.