Bayesian framework based additive intrinsic components optimization deformable model for image segmentation

Abstract

The effectiveness of image segmentation can be greatly compromised by factors like inhomogeneity, low-resolution, and noise. Aiming at these challenges, we propose a new segmentation-oriented additive decomposition model for images. Firstly, the model assumes that the to be segmented image is the sum of three components: true image, bias field, and noise. Secondly, we pursue the true image in the image domain base on Bayesian framework, and establish the active contour model. In this model, the conditional probability is assumed to follow a local Gaussian distribution. The prior probability is constructed jointly by the following three assumptions. Specifically, we describe the true image as a Markov field defined as the Gibbs energy function. The bias field $b$ is modeled as a Gaussian distribution with mean 0 and variance ${\sigma }_i$. In addition, as an alternative, we employ regularization to the evolution curve by means of heat kernel convolution function. Finally, the proposed multi-objective optimization model is solved numerically using variational and gradient descent algorithms. The effectiveness of the proposed model has been validated through experiments conducted on various images, including natural, degraded text document, and others. The results show that compared to the classical active contour model, our model improve across four evaluation metrics. Among these, the smallest increase is in the P value, at 5%, while the most significant improvement is in the JSC value, reaching 14%.