Typical Active Contour Models

Mingqi Gao

Here some typical and suggestive research works dedicating in Active Contours or Level Sets are provided, grouped according to the ways by which they capture image information, including learning-based models, edge-baesd models, region-based models and shape-prior-based models. Also, some ACMs for global optimization are introduced, can be found at the end of this page.

Learning-based Models

These models segment images using the features captured by Deep Convolutional Neural Network (DCNN). As a result, more hierarchical image information can be learned, leading to more accurate results. But the dataset containing large amounts of annotated images is required, like the following two papers:

Edge-based Models

These models segment images based on the assumption that objects to be segmented have distinct boundaries. To this end, their energy functionals are designed to move active contours towards the place where intensities change sharply, in which gradient information is the most frequently used cue. But segmentation process is prone to be disturbed by noise and clutter scene.

Region-based Models

Generally, these models use a certain kind of descriptor (e.g. intensity, color, texture) or their combinations to describe different regions in input images. The goal of these models is to divide input images into several regions in which descriptors are distributed evenly. Because segmentation process depends on the image information from all pixels, more robust results can be achieved than edge-based models. For intensity and texture descriptors, some typical models are listed as follows:

- Intensity descriptor

- Texture descriptor

Shape-based Models

ACMs for Global Minimization