Image segmentation - based on the divided image areas

Looking directly divided regions is based on region segmentation technique, in fact, similar to the same boundary based image segmentation using the similarity of the object and the background intensity distribution.

Generally based image segmentation region can be divided into two categories:

1. Region growing
2. Regional split and merge

A region growing

According to certain criteria or sub-pixel region larger area city polymerization process. The key is to select the region growing method suitable growth criteria, different criteria can affect the growth of the region growing process, results. Growth criteria may be formulated according to different principles, guidelines most region growing using local image properties.

1.1 Basic Methods

Begins with a set of seed points, the seed of a similar nature (e.g., gray level) is attached to each field of the pixel region grown from seed

1.1.1 seed produced

• Depending on the nature of the problem of selecting one or more starting point
• Without prior information, the same feature set is calculated for each pixel, the feature set for the pixel growth in an area belonging to
• If the results of these calculations are presented the values ​​of different clusters of pixels near the center of the cluster can be used as seed

1.1.2 terminates rules
if no conditions added to the pixels of a region are met, the area stop growing, termination rule-makers need prior knowledge or prior model.

1.1.3 similarity criteria

• Grayscale similar criteria
• Texture similar guidelines
• Color similar guidelines

1.2 Algorithm for Region Growing

1.2.1 based on the area gradation difference region growing method
which is a basic unit of a pixel operating method, the following steps:

1. The image is scanned line by line to find out still images without attribution
2. In this check it is centered on the pixel art pixel, i.e. the pixel-by-field comparison with it, if the gradation difference is smaller than the threshold value $T$ , they will merge
3. With the new merged pixel center, go back to step 2 to check for new pixels in the neighborhood, until the area can not expand further
4. Repeat steps 1 to 3 until it find the pixel does not belong whole growth process is ended

4 a field threshold $T=1$ region growing example:

Improvement of the method :

• Average gray difference obtained in all areas, and combined with the adjacent region of the minimum gradation difference
• Consider the history of the region growing
• Consider the growth of shape

1.2.2 based region growing region of the gradation distribution statistics
consider the similarity of the intensity distribution as a criterion to determine the growth merge region, the following steps:

1. The image is divided into non-overlapping small regions
2. Histogram comparing adjacent area, merging the similarity of the intensity distribution of regional
3. Setting the termination criterion, the respective areas are sequentially combined by repeating step 2, until the termination criterion

Assume $h_1(z)$ . $h_2(z)$ , respectively two adjacent area histogram, the intensity distribution of the similarity criteria: $\max|\ h_1(z)-h_2(z)\ |$ $\sum|\ h_1(z)-h_2(z)\ |$ if the detection result is less than a given threshold value $T$ , is about to merge the two regions
should be noted that the size of divided areas to the right, the reliability decreases too small, too large an area shape is not measured over

2 region split and merge

The basic idea of ​​the region splitting and merging algorithm similar to the differential, that is infinitely divided, then the area will be divided to meet similarity criteria are combined.

2.1 Implementation steps

1. For different gray levels in the image area, divided into four sub-regions
2. If the neighboring sub-regions of the same gray level for all the pixels, it is merged
3. Repeat steps 1 and 2, until there are no new split and merge

2.1.1 improved method of
similarity criteria $P(R_i)$ Is defined as:

1. More than 80% of the area of ​​the pixel satisfies Inequality $P(R_i)=|\ z_{ij}-m_i\ |\leqslant 2\sigma_i$among them $Z_ {ij}$It represents a region $R_{ij}$B $j$ gradation dots; $m_i$It represents a region $R_i$The average gray level; $\sigma_i$It represents a region $R_i$Grayscale standard deviation
2. If the $P(R_i)\leqslant 2\sigma_i$, Then the region of the gray level of all pixels is set $m_i$