Exploiting local intensity information in Chan–Vese model for noisy image segmentation

• This manuscript presents an improved region-based active contour model for noisy image segmentation.
• We define a local energy according to intensity information within the neighborhood of each point in image domain.
• By introducing a kernel function, our method employs intensity information in local region to guide the motion of active contour.
• Experiments on synthetic and real world images show that our model is robust to image noise while preserving the segmentation efficacy.

Due to the material property and imperfections of imaging devices, noise often exists in real-world images. This paper presents an improved region-based active contour model—Robust Chan–Vese (RCV) model for noisy image segmentation. First, for each point in a region, a local energy is defined according to the difference between the intensities of all points within its neighborhood and the intensity average of the region. Then, for the whole image domain, a global energy is defined as a data term to integrate the local energy with respect to the neighborhood center. Finally, the overall energy is represented by a level set formulation, from which a curve evolution equation is derived for energy minimization. Due to a kernel function in the data term, intensity information in local region is taken into account to guide the motion of contour, which enables RCV model to cope with noise. The improved method has been evaluated on synthetic image and industrial CT images. Compared with several popular level set methods, the experimental results show that RCV model is not only less sensitive to contour initialization, but also more robust to image noise while preserving the segmentation efficacy.