A superpixel-level active contour model with global and local constrains

Turbopixels are a powerful tool for image over segmentation, and have become increasingly popular in computer vision applications. They are constrained to have compactness, uniform size, and adherence to object boundaries. In this paper we advocate the use of turbopixels as the basic unit of active contour model, instead of operating at the pixel level. Firstly a color image is over-segmented into turbopixels, and then we represent color cues of turbopixels as a sparse color histogram in 8-d HSV color space. Finally we introduce edge indicator function and localizing region idea to obtain an improved active contour model. Practical experiments prove that our algorithm can obtain better segmented results than the state-of-art active contour model in color image segmentation.