Image Segmentation Research on Image Segmentation Technology
Image segmentation is one of the basic problems of image processing and computer vision, and it is the key step of image processing and image analysis. This paper introduces the threshold-based segmentation method and the evaluation and application status of image segmentation performance. Finally, the development trend of image segmentation is summarized.
Keywords: image segmentation, threshold, edge detection, region segmentation
China Library Classification Number: TN957.52 Document Identification Number: A.
1 Introduction
With the in-depth study of image segmentation technology, its application is increasingly extensive. Any work that needs to extract and measure image targets is inseparable from image segmentation. Image segmentation is a very important and difficult problem in many fields such as image processing, pattern recognition and artificial intelligence, and it is the first and very important key step in computer vision technology. The result of image segmentation directly affects the understanding of images in computer vision. Most of the existing methods are designed for specific applications, which have great pertinence and limitations. So far, there is no universal method or objective standard to judge whether the segmentation is successful or not. Therefore, the study of image segmentation still lacks a unified theoretical system, which makes the study of image segmentation still a very challenging topic.
2 image segmentation method
Image segmentation is simply to divide a digital image into different regions, which have the same properties under certain standards, such as gray, color, texture and so on. However, there are obvious differences in attributes between any adjacent areas.
2. 1 Threshold segmentation method based on gray features
Threshold segmentation technology is one of the classic and popular image segmentation methods. It divides the gray level of an image into several parts with one or several thresholds, and considers that pixels belonging to the same part are the same object.
This method mainly includes the following:
(1) Single threshold method, which uses global threshold to distinguish background from target. When the histogram of the image has obvious double peaks, the valley between the two peaks is selected as the threshold.
(2) Double threshold method, which uses two thresholds to distinguish the background from the target. By setting two thresholds, it is prevented that a single threshold is set too high or too low, and the target pixel is mistaken for the background pixel or the background pixel is mistaken for the target pixel.
(3) Multi-threshold method, when there are uneven illumination, sudden noise and other factors or the background gray level changes greatly, there is no suitable single threshold for the whole image, and a single threshold cannot take into account the specific conditions of different regions of the image. At this time, the image can be divided into blocks, and a threshold can be set for each block.
2.2 Edge detection segmentation method
Edge detection technology can be divided into parallel edge detection and serial edge detection according to the processing order. Common edge detection methods include: difference method, template matching method and statistical method. Because the change law of edge gray level is generally stepped or pulsed. The relationship between edge and difference can be summarized as two cases. One is that the edge appears at the maximum or minimum value of the difference; The second is that the edge appears at the zero crossing.
2.3 Region-based segmentation method
Region-based segmentation method takes advantage of the spatial characteristics of images. This method considers that the segmented regions have similar properties. The commonly used methods are region growing method and region splitting and merging method. This method is effective for segmenting images of complex scenes or natural scenes with insufficient prior knowledge.
The region growing method starts with dividing the image into many small regions, which can be small neighborhoods or even single pixels. In each region, the features that can reflect the consistency of pixels in the object are calculated as the criteria for region merging. The first step of region merging is to assign a set of parameters, namely features, to each region. Next, check all the boundaries of adjacent areas. If the eigenvalues on both sides of a given boundary are obviously different, then this boundary is strong, and vice versa. Strong boundaries are allowed to continue to exist, weak boundaries are eliminated, and adjacent areas are merged. When there is no erasable weak boundary, the region merging process ends and the image segmentation is completed.
2.4 the combination of image segmentation technology and specific tools
Since the late 1980s, with the emergence and maturity of some special theories, such as mathematical morphology, fractal theory, fuzzy mathematics, wavelet analysis, pattern recognition and genetic algorithm, a large number of scholars have devoted themselves to applying new concepts and methods to image segmentation, which has effectively improved the segmentation effect. Many new segmentation algorithms have been generated. The following are some simple summaries of these algorithms.
2.4. 1 Segmentation algorithm based on mathematical morphology
Watershed algorithm is a classical segmentation method based on mathematical morphology theory. In this method, the image is compared with the terrain with different height values, and the high gray value is considered as a ridge and the low gray value is considered as a valley. When a drop of water flows from any point, it will flow to the bottom of the terrain and finally gather at a local lowest point. Finally, all the drops will gather in different suction basins, so that the corresponding image will be divided into several parts. Watershed algorithm has simple operation and excellent performance, and can extract the contour of moving target and get the edge of moving target accurately. However, segmentation needs gradient information and is sensitive to noise.
2.4.2 Segmentation algorithm based on fuzzy mathematics
At present, a remarkable feature of the application of fuzzy technology in image segmentation is that it can be combined with many existing image segmentation methods to form a series of comprehensive fuzzy segmentation technologies, such as fuzzy clustering, fuzzy threshold, fuzzy edge detection and so on.
This method mainly includes two segmentation algorithms: generalized fuzzy operator and fuzzy threshold method.
(1) The generalized fuzzy operator processes the image within the scope of the generalized fuzzy set, so that the real edge is at a lower gray level, but some pixels are not edges at a lower gray level. Although the calculation of the algorithm is simple and the edge is delicate, the edge image will be damaged.
(2) The fuzzy threshold method introduces the fuzzy mathematical description of the gray image, selects the segmentation threshold of the image by calculating the fuzzy entropy of the image, and then processes the image with the threshold method to get the boundary.
2.4.3 Segmentation method based on genetic algorithm
This algorithm is a solution to the optimization problem proposed by the idea of biological evolution. It uses parameter coding set instead of parameter itself, and uses the strategy of survival of the fittest to simulate evolution to search the solution space of function. It searches for optimization in a group of points rather than a single point. Genetic algorithm uses random transformation rules instead of deterministic rules to work in the solution process. The only information it needs is the fitness value, and the search process is completed by simple replication, hybridization and mutation. Because this method can search the global minimum of the energy function, reduce the dimension of the search space, reduce the sensitivity of the algorithm to the initial position of the template and greatly reduce the calculation time. Its disadvantage is that it is easy to converge to local optimum.
2.4.4 Segmentation algorithm based on neural network
Artificial neural network has self-organization, self-learning, adaptive performance and very strong nonlinear mapping ability, which is suitable for solving the classification problems with unclear background knowledge, unclear reasoning rules and complex problems, so it is also suitable for solving complex image segmentation problems. In principle, most segmentation methods can be realized by ANN (Attention Neural Network). The application of artificial neural network in segmentation started late. At present, only multilayer feedforward neural network, multilayer error back propagation neural network, self-organizing neural network, Hopfield neural network and CSNN constraint fact neural network have been applied. Multi-layer feedforward neural network for image segmentation. The number of neurons in the input layer depends on the number of input features, while the number of neurons in the output layer is equal to the number of classifications.
2.5 Other methods of image segmentation
Four common image segmentation methods are introduced. There are many related image segmentation methods and documents, and new methods are constantly produced. Some of these methods are only effective under certain circumstances, and some are combined with several methods, which are collectively called the fifth category.
(1) annotation is a method based on statistics. In this method, several areas to be divided in an image are represented by different labels, each pixel in the image is marked in a certain way, and pixels with the same label are merged into the areas represented by the labels.
(2) Segmentation method based on Snake model. The segmentation based on Snake model approximates the true contour of the image object by dynamically optimizing the energy function.
(3) Texture segmentation. Due to the introduction of new mathematical tools, texture segmentation technology has made some progress. Zhang Peng and others applied wavelet analysis to texture primitive extraction.
(4) The knowledge-based image segmentation method is directly based on prior knowledge, which makes the segmentation more in line with the characteristics of the actual image. The difficulty of this method lies in the correct and reasonable representation and utilization of knowledge.
Evaluation of image segmentation performance
The evaluation of image segmentation mainly includes two aspects: one is to study the performance of various segmentation algorithms in different situations and master how to choose and control their parameter settings to meet different needs. The second is to analyze the performance of various segmentation algorithms when segmenting the same image, and compare the advantages and disadvantages, so as to choose the appropriate algorithm in practical application. Segmentation evaluation methods are divided into two categories: analytical method and experimental method. The analysis method directly analyzes the principle and performance of the segmentation algorithm itself, and the experimental method evaluates the algorithm by testing the segmentation results of the image. These two methods have their own advantages and disadvantages. Due to the lack of reliable theoretical basis, not all segmentation algorithms can analyze their performance through analysis. Each evaluation method is put forward for some consideration, and different evaluation methods can only reflect some performance of the segmentation algorithm. On the other hand, the performance of each segmentation algorithm is determined by many factors, so it may need a variety of criteria for comprehensive evaluation.
4 the development trend of image segmentation technology
With the wide application of neural network, genetic algorithm, statistical theory, wavelet theory and fractal theory in image segmentation, image segmentation technology presents the following development trend: (1) multi-feature fusion. (2) Combination of various segmentation methods. (3) New theories and methods.
refer to
[1][ America ]RC Gonzalez. Digital Image Processing (2nd Edition) [M]. Ruan, wait for translation. Beijing: Electronic Industry Press, 2003.
Zhang Yujin. Image segmentation [M]. Beijing: Science Press, 200 1.
Li, Peng, Peng Bo, et al. Intelligent image processing technology [M]. Beijing: Electronic Industry Press, 2004.
Yang Hui, Qu. Overview of image segmentation methods [J]. Computer development and application. 2005, 18(3):2 1-23.
Click the next page for details >>& gt Paper on image segmentation technology.