This type of method first divides the image into several "initial" regions, then splits or merges these regions based on a certain index, and then gradually improves the index of region segmentation until finally the image is divided into the least number (or meets a certain requirement) of " Basically the same" area. These methods are relatively insensitive to noise and have high computational complexity. Such methods mainly include region growing algorithm, watershed algorithm, active contour models (active contour models, ACM) and so on.

 

The basic idea of the region growing algorithm is based on a certain similarity criterion, starting from a group of growing points, the adjacent pixels or regions similar to the growing point are merged with the growing point to form a larger region, and this process is repeated until the stopping criterion is satisfied. Selecting the appropriate growing point, determining the similarity criterion and the growth stop condition are the three keys of the region growing algorithm. The region growing algorithm can usually segment the connected regions with the same characteristics, which has the advantages of flexible control of the generation process and clear boundary information, but the computational complexity is often large. Noise and uneven gray scale in the image can easily lead to holes and over-segmentation.

 

The watershed algorithm is a mathematical morphology segmentation method based on the extension step theory. It regards the image as a topological landform in geodesy. The gray value of each pixel in the image represents the altitude of the point. Each local minimum value and its affected area are called catchment basins, and the boundary of catchment basins is A watershed is formed. The formation of the watershed can be illustrated by the process of "simulating immersion", that is, a small hole is pierced on the surface of each local minimum, and then the entire model is slowly immersed in water. As the immersion deepens, each local minimum The influence domain of the value gradually expands outward, and a watershed is formed by building a dam at the confluence of two catchment basins. The watershed algorithm has a good response to weak edges, so it is prone to over-segmentation. This problem can be improved by using prior knowledge to delete irrelevant edge information or modifying the gradient function so that the catchment basin only responds to the detected target.

 

The basic idea of ACM to achieve image segmentation is to regard the edge of the image object as a continuous smooth closed curve, and then evolve the curve based on information such as image edge, area, object shape, etc., so that it gradually moves and topologically changes from the initial position within the image definition domain. , and finally stop forming the target contour at the target edge, thereby segmenting the target from the image. Because the closed curve can better describe the target contour, it is usually called the active contour curve. In the early parametric ACM, the contour curve is expressed parametrically, which is gradually deformed under the drive of the energy function composed of the prior model and image data, and finally reaches the target boundary to realize image segmentation. In this process, there is a defect that it is difficult to effectively deal with the topological changes of curves (that is, the splitting and merging of curves). The subsequent appearance of geometric ACM (29.70) resolved the shortcomings of parametric ACM. Geometry The contour curve in ACM is expressed non-parametrically, and its evolution process is based on the curve's geometric characteristic measure (such as normal vector, curvature, etc.). In addition, geometric ACM is usually combined with the level set method, so that the active contour curve is expressed as the zero level set of the level set function (LSF), and then the model realizes the Flexible handling of curve topology changes. Variational level set active contour models belong to another class of geometric ACMs. This type of model is derived from some kind of energy function related to the image (or segmented target). Since the energy function can naturally incorporate additional constraint information such as image edges, regions, motion or object shape prior knowledge, the ability to solve complex image segmentation problems is stronger. The geometric ACMs that realize the curve evolution by evolving LSF are collectively referred to as level set based active contour models (LSAC).