Image segmentation Overview (favorites)

Image segmentation Overview (favorites)

Image segmentation is a key technology from image processing to image analysis. There are many types and methods of image segmentation. Some segmentation algorithms can be used directly for any image, while others can only be used to segment images of special categories. Some algorithms must first perform rough image segmentation because they need to extract information from the image. There is no unique standard method. The quality of the splitting result must be measured according to the specific requirements.

Early image segmentation methods can be divided into two categories. One is the boundary method. This method assumes that a subarea of the image segmentation result must have an edge in the original image. The other is the region method, this method assumes that a subarea of the image segmentation result must have the same properties, while pixels in different regions do not have the same properties. Both methods have advantages and disadvantages. Some scholars consider combining them for research. Nowadays, with the improvement of computer processing capabilities, many methods are emerging, such as color component segmentation and texture image segmentation. The mathematical tools and analysis methods used are constantly extended, from time-domain signals to frequency-domain signal processing and wavelet transformation.

Image segmentation mainly includes four technologies: Parallel Boundary segmentation, serial boundary segmentation, parallel area segmentation, and serial area segmentation.

The following is a brief introduction to each item.

1. Parallel Boundary Division

Different images have different gray scales, and the border usually has obvious edges. This feature can be used to separate the images. It must be noted that the boundary between the edge and the object is not the same. The edge refers to the position where the pixel value in the image changes, the boundary between objects refers to the boundary between objects in real scenarios. It is possible that the edge is not the boundary, or the boundary is not the edge, because objects in the real world are three-dimensional, and images only have two-dimensional information, from 3D to 2D, projection imaging will inevitably lose part of the information. In addition, illumination and noise during imaging are also an inevitable important factor. For these reasons, edge-based image segmentation is still a world-class problem in image research. Researchers are trying to add high-level semantic information to edge extraction.

In actual image segmentation, only the first and second derivatives are used. Although, in principle, the higher derivatives can be used. However, due to the influence of noise, derivative information above level 3 often loses application value. The second derivative can also indicate the type of gray-scale mutation. In some cases, for images with even gray variation, the first derivative may not be able to find the boundary. In this case, the second derivative can provide useful information. The second derivative is also sensitive to noise. The solution is to perform smooth filtering on the image, eliminate some noise, and then perform edge detection. However, the second-order derivative information algorithm is based on zero-crossing detection. Therefore, the obtained algorithm has fewer edge points, which is conducive to subsequent processing and recognition.

Robert ts OPERATOR: accurate edge location, but sensitive to noise. Suitable for image segmentation with obvious edges and less noise.

Prewitt operator: It can suppress noise. The principle of noise suppression is through pixel average, but the pixel average is equivalent to low-pass filtering of the image. Therefore, the Prewitt operator is not as good as the Roberts operator in edge location.

Sobel operator: Sobel operator and Prewitt operator are both weighted average, but Sobel operator believes that the influence of neighboring pixels on the current pixel is not equivalent, therefore, different pixels have different weights, which have different effects on the operator results. In general, the farther the distance is, the less impact it has.

Isotropic Sobel operator: weighted average operator. The weight value is inversely proportional to the distance between the neighboring point and the center point. When edge is detected in different directions, the gradient amplitude is the same, which is usually referred to as the isotropic Sobel operator.

The above calculation uses the information of the first derivative.

Laplacian operator: Second-order differential operator. It has an undirected relationship, that is, it has nothing to do with the axis direction, and the gradient result remains unchanged after the axis is rotated. However, the image is sensitive to noise. Therefore, the image is first processed smoothly because the template is used for smooth processing, generally, the splitting algorithm combines the Laplacian operator and the smoothing operator to generate a new template.

2. Serial boundary Segmentation

The parallel edge detection method does not depend on the processing results of other points on each point of the image. Serial boundary segmentation not only utilizes the pixel information, but also uses the previously processed pixel results. The processing of a pixel and whether to classify it into a boundary point are related to the information obtained from the previous processing of other points. The serial boundary division technique usually works by searching edge points in sequence. Generally, there are three steps 1. Determining the starting edge point. 2. search criteria. The next edge point will be determined based on this criterion. 3. Set the condition for ending the search process.

Border Tracking is a serial border segmentation method.

Border Tracking is a method that starts from an edge point in a gradient graph and searches for and connects edge points to gradually detect all boundaries. In the parallel boundary division method, edge pixels may not be combined into closed curves because there may be gaps on the boundary. The gap may be too large to be connected by a straight line or curve, or it may not be a gap on a boundary. The border tracking method can solve these problems to some extent. For some images, this method has better segmentation results. The specific algorithm is to first perform gradient operations on the original image, and then perform the Border Tracking Algorithm. 1. Start Point: Search for the gradient chart and find the maximum gradient point as the start point of the Border Tracking. 2. Growth rules: in the eight-neighbor pixel of a vertex, the vertex with the largest gradient is treated as a boundary, and the vertex is also used as the start point of the next search. 3. termination condition: Follow the 2 criterion to continue searching until the absolute gradient value is less than a threshold value. Sometimes, to ensure the smoothness of the boundary, we only select pixels within a certain range each time. The obtained boundary points not only ensure connectivity, but also ensure smoothness.

 

3. Parallel area Segmentation

There are two main methods for parallel area segmentation: Threshold Value Segmentation and clustering.

Direct threshold segmentation is generally not suitable for the correct segmentation of complex scenes, such as natural scenes. Because of the images of complex scenes, it is difficult to determine whether the scene is the foreground or background in some regions. However, threshold value segmentation is particularly useful when there is a strong contrast between the processing prospect and the background, and requires less computing complexity. When the gray level of an object is concentrated, it is an effective method to simply set the gray level threshold value extract body.

The threshold method can be divided into two types: global threshold and local threshold. If the thresholds used by each pixel in the image are the same during the segmentation process, the global threshold method is used; if the thresholds used by each pixel may be different, the local threshold method is used. The following common methods are used to determine the optimal global threshold: test method, histogram method, and minimum error method (this method assumes that the gray distribution of the background and foreground is normal ). When the illumination is uneven, there is sudden noise, or the background gray scale changes are large, the entire image is not split into a single threshold, because a single threshold cannot take into account the actual situation of each pixel of the image. At this time, the image can be segmented by coordinate, and each segment can be selected with a threshold value. This coordinate-related threshold is called the dynamic threshold method, also known as the adaptive threshold method. The time and space complexity of these methods are relatively large, but the anti-noise capability is relatively strong, which has a better effect on images that are not easy to split using global thresholds. When a simple method is used to select an adaptive threshold, a neighboring window centered on each pixel is determined, the maximum and minimum values of pixels in the window are calculated, and their mean value is used as the threshold value. Histogram analysis can be used for each sub-block after an image is segmented. If a sub-block contains a target and a background, the histogram shows a double peak. If there is only a target or a background in the block, the histogram does not have a dual peak. you can separate the histogram Based on the parameter interpolation obtained by each block in the neighborhood. The actual adaptive threshold segmentation can set a threshold for each pixel based on the actual nature of the image. However, the actual requirements and computing complexity must be taken into account in this process.

4. Serial area Segmentation

Serial region segmentation can be divided into two methods: area growth and split and merge. Regional Growth refers to the gradual addition of neighboring pixels starting from a certain pixel according to certain standards. when certain conditions are met, the growth of the region ends. The growth condition of a region is determined by 1. Selection of the initial vertex (seed vertex) 2. Growth Criterion 3. termination condition.

The growth of a region starts from a certain pixel or a certain number of pixels, and finally obtains the entire region to extract the target. Splitting and merging are similar to the inverse process of region Growth: Starting from the entire image, the system continuously splits into sub-regions and then merges the foreground regions to extract the target. The assumption of splitting and merging is that for an image, the foreground area is composed of several interconnected pixels. Therefore, if an image is split to a pixel level, then you can determine whether the pixel is a foreground pixel. After all the pixels or sub-regions are determined, combine the foreground regions or pixels to obtain the foreground target.