B-Spline based globally optimal segmentation combining low-level and high-level information

- The low-level and high-level information are combined to build an energy functional using multi-scale. Rough contour achieved at coarsest scale by multiple Gaussian kernel gray equalization is used as the prior shape of object. This shape is updated and used as the constraint of evolving contour in following fine-scale.
- A new edge stopping function based on the TV regularization is proposed which is beneficial to both global-based method and fast decrease of energy minimization.
- We proposed a statistical based globally optimal segmentation model using cubic B-Spline basis functions. These functions are used to explicitly represent the relaxation characteristic function which contributes to fast convergence and intrinsic smoothing globally optimal segmentation results.

Image segmentation is an important step for large-scale image analysis and object recognition. Variational-based segmentation methods are widely studied due to their good performance, but they still suffer from incapability to deal with images bearing weak contrast, overlapped noise and cluttered texture. To tackle this problem, we propose a new statistical information analysis based multi-scale and global optimization method for image segmentation. This multi-scale processing which is consistent with human's cognition mechanism enables us identify target at coarse scale. The high-level prior is obtained by the multiple Gaussian kernel gray equalization and used as shape constraint in following fine-scale. An efficient energy functional is proposed with convexity and improved TV regularization in order to segment inhomogeneous target from noisy background. A convex relaxation function is explicitly represented by cubic B-Spline basis for fast convergence and intrinsic smooth segmentation. Finally, the energy functional is minimized by standard methods of Split Bregman, Gradient Descent Flow and the corresponding Euler-Lagrange Equation. Experimental results on synthetic and real world images validate the robustness and high accuracy boundaries detection for low contrast, noisy and texture images.