A review on image segmentation

1. Definition

To partition an image into several regions. There are 2 versions:

1. Get desired patterns from the image with undifferentiated background;
2. Partition an image into several regions.

The first 1 can be viewed as a special case of 2.

2. Models for image patterns

There is no universal model for patterns on the image. Some theories assume the patterns follow Gaussian distribution, some assume the patterns have unique texture, etc. For the images obtained from the the real world, they are all projections of one or several objects on the background, which have specific geometric boundary to separate themselves from other objects.

In the image acquisition, the image background may not be uniform either because the acquisition device introdues the distortion, or the illumination, or the nature of background. The uneven background makes the segmentation more difficult. On the other hand, all images are digital, the digitization may create artifacts, which blur the object boundary or texture. For nonuniform background compensation, 2 main approaches are available: estimate the background nonuniformity from the image directly, or estimate it in the segmentation. In the first approach, spatial filtering is a simple method but it may create artifacts, a robust method is called N3 (nonparametric, non-uniform normalization), which is based on the assumption that INU blurs the PDF of image (estimated by image histogram).

There're many methods to segment the images. 2 popular schools of methods are thresholding and edge based method. Based on those 2 types of methods, there are region based segmentation, classification based method, and clustering based methods. What's more, deformable models also have been successfully used in image segmentation, which includes parametric deformable models, geometric deformable models, and etc.

3. Thresholding
The main topic of thresholding is to find an appropriate threshold value to separate the image in different regions according to their intensity or other features. The first technique used widely is based on histogram shape. In the 2 peaks case, the local minimum between 2 peaks is the optimal threshold. Some other methods includes optimal thresholding to minimize some of the objective functions, such as inter-region contrast, mean square error, EM algorithm, k-mean clustering, fuzzy c-mean, entropy, etc.

4. Edge based method
There are many edge detectors, such as Sobel, SUSAN, Canny, Laplacian of Gaussian, etc. Also, morphological method can be used to find the skeletons of the image.

To look for a global edge trace, many methods are proposed. One school of the method is based on template matching, Hough transform can be used to find parametric curves, such as lin, cycle, eclipse, etc. Border tracing to look for the border of an object on the image. Graph searching uses methods such as A*, dynamic programming to find the optimal trace.

5. Region based method
There are many methods available, to name a few, region growing ( which is used to get spot in Mcc), region merging and splitting (used to get spot and cluster in Mcc), connected component labeling.

6. Classification and clustering based method

• K-mean clustering
• fuzzy c-mean clustering
• K-nearest neighboring
• Adapative fuzzy c-mean
• Decision tree
• ANN

7. Deformable models
Parametric deformable model

Geometric deformable model