Image segmentation is a key technique for processing from image to image analysis. There are many kinds and methods of image segmentation, and some split algorithms can be used directly for any image, while others can only apply to images of special categories. Some algorithms need to be roughly divided on the image because they need the information extracted from the image. There is no unique standard method. The quality of the segmentation results need to be measured according to the specific occasion.
Early image segmentation methods can be divided into two categories. One is a boundary method. This method assumes that an image segmentation result will have an edge in the original image; one is a regional method, which assumes that the image segmentation result will have the same properties And the pixels in different regions have no common nature. Both methods have advantages and disadvantages, and some scholars consider combining both. Now, with the improvement of computer processing power, many methods continue to emerge, such as color component segmentation, texture image segmentation. The mathematical tools and analysis methods used are also constantly expanded, from time domain signals to frequency domain signal processing, wavelet transform, and the like.
Image segmentation mainly includes four technologies: parallel boundary segmentation technology, serial boundary split technology, parallel region segmentation technology, and serial area segmentation technology.
Below is a brief introduction to each item, respectively.
Parallel border segmentation
Different image grayscale are different, and the boundary generally has a significant edge, and the image can be divided by this feature. It should be noted that the boundaries between the edges and objects are not equivalent, and the edge refers to where the value of the pixels in the image has a mutation, and the boundary between the object refers to the boundary between the objects in the real scene. It is possible that there is a side of the edges, and there is also a possibility that the border is not bounded because the object in the real world is three-dimensional, while the image has only two-dimensional information, from three-dimensional to two-dimensional projection imaging inevitable will be lost. Information; additional, illumination and noise during imaging process are also an important factor. It is because of these reasons, the edge-based image segmentation is still world-class problems in the current image research, and current researchers are trying to add high-level semantic information in the edge extraction.
In actual image segmentation, it is often only used in the first order and second order derivatives. Although in principle, a higher order derivative can be used, however, because the noise affects, the three-order derivative information often has lost the application value. The second-order derivative can also explain the type of grayscale mutation. In some cases, such as a uniform image of grayscale, only the first order derivative may not find the boundary, and the second order derivative can provide very useful information. The second-order derivative is also sensitive to noise, and the method of solving is to smoothly filter the image, eliminate partial noise, and then perform edge detection. However, the algorithm using the second-order derivative information is based on cross zero detection, so the obtained edge point is less, which is conducive to the processing and identification of successive subsequent.
Roberts Operator: Edge is positioned, but sensitive to noise. Suitable for image segmentation of the edges with obvious edges and less noise.
Prewitt Operator: The noise is suppressed, and the principle of suppressing noise is to pass the pixel, but the pixel is equivalent to the low-pass filtering of the image, so the preWitt operator is not as good as the Roberts operator.
Sobel Operator: The Sobel Operator and Prewitt operator are weighted average, but the Sobel operator believes that the influence of pixels in the neighborhood is not equivalent, so different pixels have different weights, calculate The effects of sub-results are also different. In general, the farther distance is, the smaller the impact.
ISOTROPIC Sobel Operator: Weighted average operator, weight than the distance between the adjacent point and the central point, the gradient amplitude is consistent while detecting the edge in different directions, is usually the sameness.
The information of the first order derivative is utilized when the operator is above.
Laplacian Operator: At this time, second-order differential operator. It has the sameness, that is, it is independent of the direction of the coordinate axis, and the gradient result is unchanged after the coordinate axis rotation. However, it is more sensitive to noise, so the image is generally smoothed, because smooth processing is carried out with a template, so, the usual segmentation algorithm combines the LAPLACIAN operator and the slip operator to generate a new template. 2. Serial boundary segmentation
Parallel edge detection methods, the processing made on each point of the image does not depend on other point processing results. Serial boundary splitting information not only uses the information of its pixel when processing images, but also utilizes the result of the previously treated transmissive. Treatment of a pixel, and whether it is classified into boundary points, and information obtained from the processed processed by other points. Serial boundary segmentation techniques are typically operated by sequential search edge points, generally three steps 1. Determination of start edge points. 2. Search criteria will determine the next edge point according to this guideline. 3. Terminate the conditions, set the conditions that the search process ends.
Boundary tracking is a method of serial boundary segmentation.
Boundary tracking is a method of searching and connecting to an edge point in a gradient diagram to gradually detect all boundaries. In the parallel boundary segmentation method, the edge pixel is not necessarily combined into a closed curve, because the boundary may encounter a gap. The gap may be too large and cannot be used with a straight line or curve, or may not be a gap on a boundary. The method of border tracking can solve these problems to some extent, for some images, the segmentation results of this method are better. The specific algorithm is, first gradient operation, then perform a boundary tracking algorithm. 1. Pending point: Search for gradient graph, find the maximum gradient, and do the starting point for boundary tracking. 2. Growth rules: In this 8 neighborhood pixels, the maximum gradient is used as a boundary, while this point will be the starting point of the next search. 3. Termination Condition: Search until the gradient absolute value is less than a threshold, the search stops. Sometimes, in order to ensure the smoothness of the boundary, it is only selected in a certain range of pixels each time, which not only guarantees connectivity, but also guarantees smooth.
3. Parallel area segmentation
There are two main methods in parallel region: threshold segmentation and clustering.
Direct threshold segmentation generally does not apply to the correct segmentation of complex scenes, such as natural scenes, because of the image of complex scenery, some areas difficult to determine whether the prospect is still background. However, the threshold segmentation is especially useful when handling the foreground and background, it is especially useful, and the computational complexity required at this time is small. When the grayscale level of the object is concentrated, a simple setting of the gray level threshold extracting object is an effective way.
The threshold method is divided into two types of global thresholds and partial thresholds. If the threshold value used for each pixel on the image is equal during the split, it is a global threshold method; if the threshold used in each pixel may be different, it is a local threshold. method. Common methods for the identification of optimal global thresholds generally have the following: Test Method, histogram, minimum error method (this method is assumed that the grayscale distribution of background and foreground is normal distribution). When the light is uneven, there is a sudden noise, or the background gradation variation is relatively large, the entire image segment will not be a single threshold because a single threshold cannot take the actual situation of the respective pixels. At this time, the image can be divided into the coordinate block, and each block is divided separately, and the threshold value related to the coordinate is referred to as a dynamic threshold method, also referred to as an adaptive threshold method. This type of method is relatively large, but the anti-noise ability is relatively strong, and there is a good effect on the image that is not easy to divide with global thresholds. When the relatively simple method selected for the adaptive threshold, determine each pixel in which the maximum and minimum values of the pixels in the window are calculated, and then their mean is made as threshold. Each sub-block after the image block can be analyzed by a histogram. If there is a target and background in a sub-block, the histogram is a double peak. If there is only a target or background in the block, the histogram does not have a double peak, which can be split according to the parameter interpolation obtained by the neighborhood. The actual adaptive threshold segmentation can set a threshold for each pixel according to the actual properties of the image, but this process should take into account the actual requirements and computational complexity issues. 4. Serial area segmentation
Serial area segmentation can generally be divided into two methods: one is regional growth, and the other is split merge. Regional growth refers to the departure of a pixel from a pixel, and gradually adds adjacent pixels, and when a certain condition is satisfied, the regional growth is terminated. The quality of regional growth is determined at 1. Selection of the initial point (seed point) 2. Growth criteria 3. Termination conditions.
Regional growth is from one or some pixel points, and finally, the entire area is obtained, and the extraction of the target is achieved. The split merge is almost the inverse process of regional growth: from the entire image, it is constantly dividing each sub-area, and then merges the foreground area to achieve the target extraction. The assumption of split merge is for an image, the foreground area consists of some mutually connected pixels, so if a image is split into a pixel level, then it can be determined whether the pixel is the foreground pixel, and when all pixel points or sub-area After the judgment is completed, the foreground area or pixel can be obtained to obtain the foreground target.
