News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Image recognition segmentation method
Image recognition segmentation method
Threshold-based segmentation
As a common area parallel technique, the threshold segmentation method divides the gray histogram of an image into several categories with one or several thresholds, so that the pixels whose gray value in the same class in the image belong to the same object. Since the gradation characteristics of the image are directly applied, the calculation is simple and applicable. Obviously, the key and difficult part of the threshold segmentation approach is how to obtain an appropriate threshold. In practical applications, the threshold setting is susceptible to noise and lightness. In recent years, methods have been used to select thresholds with maximum correlation criteria, methods based on image topological stability, Yager measure minimization methods, gray level co-occurrence matrix methods, variance methods, entropy methods, peak and valley analysis methods, etc. Among them, the self-compliance threshold method, the maximum entropy method, the ambiguous threshold method, and the inter-class threshold method are several algorithms that are more successful in improving the traditional threshold method. In more cases, the choice of thresholds can be combined with two or more methods, which is also a trend in image segmentation.
Characteristic
The advantage of threshold segmentation is that the calculation is simple, the operation efficiency is high, and the speed is fast. Global Thresholds can stop effective segmentation for different purposes and backgrounds where grayscales differ greatly. When the gray level difference of the image is not obvious or the gray value range of different purposes is stacked, partial threshold or dynamic threshold segmentation method should be adopted. On the other hand, this method only considers the gray value of the pixel itself, and generally does not think about the spatial feature, so it is very sensitive to noise. In practical applications, the threshold method is usually used separately from other methods.
Edge-based segmentation
Edge segmentation-based segmentation attempts to detect segmentation problems by detecting edges containing different regions are one of the most common approaches. Usually the variation of the pixel gray value on the edge between different regions is often violent, which is one of the main assumptions that edge detection is completed. The first-order or second-order differential operators of commonly used gray levels stop edge detection. Commonly used differential operators have a differential (sobel operator, Robert operator, etc.), a second derivative (Laplacian, etc.) and a template operation (Prewit operator, Kirsch operator, etc.).
Characteristic
The difficulty of edge-based segmentation is the contradiction between noise immunity and detection accuracy in edge detection. If the detection accuracy is improved, the false edges generated by the noise will lead to an unreasonable contour; if the noise immunity is improved, the contour miss detection and the positional deviation will occur. To this end, various multi-scale edge detection methods are proposed, and the separation plan of multi-scale edge information is designed according to practical problems, so as to better coordinate the anti-noise and detection accuracy.
Region-based segmentation
The essence of regional segmentation is to connect objects with similar properties to form the final segmentation region. It applies partial spatial information of the image, which can effectively restrain the small continuous defects of the image segmentation space existing by other methods. In such an approach, if you start from the whole picture, the regional attribution of each pixel is determined according to the principle of regional attribute differences, which constitutes a regional map, often referred to as the regional growth segmentation method. If we start from the pixel, according to the principle of divergence of regional attribute characteristics, the connected pixels with close properties are combined into regions, which is the segmentation method of regional growth. If the above two methods are applied in combination, it is a method of unity and merger. It first divides the image into a number of small areas with strong divergence, and then blends the small areas into large areas according to certain rules to reach the purpose of segmenting the image.
Characteristic
Region-based segmentation often results in over-segmentation of images, while pure edge-based detection sometimes does not provide a good regional structure. For this reason, the region-based approach and edge detection can be separated to give full play to their respective advantages. Get better segmentation results.
Image segmentation method based on clustering analysis
The feature space clustering method stops the image segmentation by using the corresponding feature space points in the image space, and stops segmentation according to their collection in the feature space, and then maps them back to the original image space to obtain the segmentation result. Among them, K-means and ambiguous C-means clustering (FCM) algorithm are the most commonly used clustering algorithms. The K-means algorithm first selects K initial class means, then classifies each pixel into the class closest to it and calculates a new class mean. Iterate through the previous steps until the difference between the old and new class mean is less than a certain threshold. The vague C-means algorithm is the implementation of the K-means algorithm on the basis of ambiguous mathematics. It is optimized by a ambiguous function to complete clustering. It does not think that each point can belong to a certain class, just like K-means clustering. Give each point a degree of subordination of the various types, and use the degree of subordination to better describe the characteristics of the edge pixels, and to deal with the inherent uncertainty of the things. Applying the characteristics of ambiguous C-means (FCM) non-monitoring ambiguous clustering calibration to stop image segmentation can reduce human intervention, and there are characteristics of uncertainty and ambiguity in the appropriate image.
The clustering approach should pay attention to several issues:
(1) How to determine the number of clusters.
(2) How to affirm the validity principle of clustering.
(3) How to set the initial value when the location and characteristics of the cluster center are not known beforehand.
(4) Expenses for computing.
And the FCM algorithm is extremely sensitive to the initial parameters, and sometimes requires the intervention of manual intervention parameters to approach the global optimal solution and improve the segmentation speed. In addition, the traditional FCM algorithm does not consider spatial information and is not sensitive to noise and grayscale.
Segmentation method based on wavelet transform
The fundamental idea of threshold image segmentation based on wavelet transform is that the histogram of the image is first synthesized into wavelet coefficients of different levels by dyadic wavelet transform, then the threshold threshold is selected according to the given segmentation principle and wavelet coefficient, and finally the threshold is applied. The area where the image is divided. The whole segmentation process is controlled from coarse to fine with scale changes, that is, the initial segmentation is done by the histogram projected on the rough L2(R) subspace. If the segmentation is not ideal, the histogram is applied in the precise subspace. The wavelet coefficients on the surface gradually refine the image segmentation. The computational feed of the segmentation algorithm varies linearly with the image size. Wavelet transform provides an accurate and unified framework for the analysis and characterization of signals at different scales. From the perspective of image segmentation, wavelet synthesis provides a mathematically complete depiction; wavelet transform can greatly reduce or eliminate the correlation between the extracted different features by selecting appropriate filters, not only with "zoom" Features, and there are fast algorithms on completion.
Characteristic
Wavelet transform is a multi-scale, multi-channel profiling tool. It is a local transform in the airspace and frequency domain. Therefore, it can effectively extract information from the signal. After the warfare warfare and other functions, the function or signal stops multi-scale analysis. Many problems that the Fourier transform cannot handle are handled. In recent years, multi-ary wavelets have been used for edge detection. In addition, wavelet transforms using orthogonal wavelet bases can also extract multi-scale edges, and can identify some types of edges by calculating and estimating the oddness of the image.
Mathematical morphology based segmentation
Mathematical morphology is a non-linear filtering method that can be used to suppress image processing problems such as noise, feature extraction, edge detection, and image segmentation. Mathematical morphology was first used to deal with binary images, and later used to deal with grayscale images. Nowadays, some scholars have begun to deal with computer vision problems with soft mathematical morphology and vague morphology. The characteristic of mathematical morphology is the ability to stop the synthesis of complex shapes and extract meaningful shape weights from useless information. Its fundamental idea is to apply a probe called a construction element to collect image information. When the probe moves from time to time in the image, not only can the structural features of the image be understood according to the interrelationship between the parts of the image, but also the application. Mathematical morphology fundamental operations can also structure many very effective methods of image processing and profiling. Its fundamental form operation is corrosion and shrinkage. Corrosion has the effect of reducing the purpose, increasing the purpose of the inner hole and eliminating the external isolated noise; and shrinking is the process of merging all the background points in the image that are in contact with the target object into the object, with the result that the purpose is increased and the aperture is reduced. The space in the destination can be added to form a connected domain. Another fundamental method of computation in mathematical morphology is open and closed operations. The open operation has the effect that the image is a small object, and the object is wary and slips away from the object at the border of the larger object; the closed operation has the function of filling the small space in the image of the object and connecting the adjacent object to the war-slip border.
Characteristic
Mathematical morphology is applied to image segmentation, and has the characteristics of good positioning effect, high segmentation precision and good anti-noise performance. At the same time, this method also has its own limitations: because in the preliminary work of image processing, the mathematical morphology of the open (closed) operation, after the image processing is stopped, there are still a lot of short lines and isolated points that do not match the purpose; Incomplete processing, it is also necessary to stop a series of point-based open (closed) operations, so the speed of operation is significantly reduced. How to combine mathematical morphology with other methods to restrain these defects will be the future work direction of mathematical morphology. The role of connecting the war-sliding borders of adjacent objects.
Segmentation method based on artificial neural network
In recent years, artificial neural network recognition technology has caused widespread concern and applied to image segmentation. The fundamental idea of the segmentation method based on neural network is to obtain a linear decision function by exercising the multi-layer perceptron, and then use the decision function to stop classifying the pixels to reach the purpose of segmentation.
Characteristic
Segmentation of images using artificial neural networks requires a large amount of exercise data. The neural network has a huge amount of convergence, easy to introduce spatial information, and can better deal with noise and unevenness in the image. The choice of network architecture is the main problem to be addressed in this approach.
Segmentation method based on genetic algorithm
Genetic algorithm (GA) is a search and optimization process that imitates natural selection and genetic mechanism. It has a strong global optimization search ability and is a self-compliant search method with universal applicability. It stops searching in the search space rather than on a single point in the search space, and it uses genetic manipulation rules rather than affirmative rules to work in the solution process. These characteristics make genetic algorithms well suited for image segmentation, especially threshold segmentation and region growing. Applying GA's global optimization ability and insensitivity to the initial position can improve the performance of image segmentation.
Characteristic
The difficulty of applying genetic algorithm to image segmentation is that the selection of the compliance function and the probability of interpolating and the probability of mutation are definitely determined. GA may also converge to partial optimality. It is conceivable to use a hybrid genetic algorithm that can self-adapt to the interpolation probability and mutation probability self-adapting genetic algorithm and the imitation annealing method.
As a common area parallel technique, the threshold segmentation method divides the gray histogram of an image into several categories with one or several thresholds, so that the pixels whose gray value in the same class in the image belong to the same object. Since the gradation characteristics of the image are directly applied, the calculation is simple and applicable. Obviously, the key and difficult part of the threshold segmentation approach is how to obtain an appropriate threshold. In practical applications, the threshold setting is susceptible to noise and lightness. In recent years, methods have been used to select thresholds with maximum correlation criteria, methods based on image topological stability, Yager measure minimization methods, gray level co-occurrence matrix methods, variance methods, entropy methods, peak and valley analysis methods, etc. Among them, the self-compliance threshold method, the maximum entropy method, the ambiguous threshold method, and the inter-class threshold method are several algorithms that are more successful in improving the traditional threshold method. In more cases, the choice of thresholds can be combined with two or more methods, which is also a trend in image segmentation.
Characteristic
The advantage of threshold segmentation is that the calculation is simple, the operation efficiency is high, and the speed is fast. Global Thresholds can stop effective segmentation for different purposes and backgrounds where grayscales differ greatly. When the gray level difference of the image is not obvious or the gray value range of different purposes is stacked, partial threshold or dynamic threshold segmentation method should be adopted. On the other hand, this method only considers the gray value of the pixel itself, and generally does not think about the spatial feature, so it is very sensitive to noise. In practical applications, the threshold method is usually used separately from other methods.
Edge-based segmentation
Edge segmentation-based segmentation attempts to detect segmentation problems by detecting edges containing different regions are one of the most common approaches. Usually the variation of the pixel gray value on the edge between different regions is often violent, which is one of the main assumptions that edge detection is completed. The first-order or second-order differential operators of commonly used gray levels stop edge detection. Commonly used differential operators have a differential (sobel operator, Robert operator, etc.), a second derivative (Laplacian, etc.) and a template operation (Prewit operator, Kirsch operator, etc.).
Characteristic
The difficulty of edge-based segmentation is the contradiction between noise immunity and detection accuracy in edge detection. If the detection accuracy is improved, the false edges generated by the noise will lead to an unreasonable contour; if the noise immunity is improved, the contour miss detection and the positional deviation will occur. To this end, various multi-scale edge detection methods are proposed, and the separation plan of multi-scale edge information is designed according to practical problems, so as to better coordinate the anti-noise and detection accuracy.
Region-based segmentation
The essence of regional segmentation is to connect objects with similar properties to form the final segmentation region. It applies partial spatial information of the image, which can effectively restrain the small continuous defects of the image segmentation space existing by other methods. In such an approach, if you start from the whole picture, the regional attribution of each pixel is determined according to the principle of regional attribute differences, which constitutes a regional map, often referred to as the regional growth segmentation method. If we start from the pixel, according to the principle of divergence of regional attribute characteristics, the connected pixels with close properties are combined into regions, which is the segmentation method of regional growth. If the above two methods are applied in combination, it is a method of unity and merger. It first divides the image into a number of small areas with strong divergence, and then blends the small areas into large areas according to certain rules to reach the purpose of segmenting the image.
Characteristic
Region-based segmentation often results in over-segmentation of images, while pure edge-based detection sometimes does not provide a good regional structure. For this reason, the region-based approach and edge detection can be separated to give full play to their respective advantages. Get better segmentation results.
Image segmentation method based on clustering analysis
The feature space clustering method stops the image segmentation by using the corresponding feature space points in the image space, and stops segmentation according to their collection in the feature space, and then maps them back to the original image space to obtain the segmentation result. Among them, K-means and ambiguous C-means clustering (FCM) algorithm are the most commonly used clustering algorithms. The K-means algorithm first selects K initial class means, then classifies each pixel into the class closest to it and calculates a new class mean. Iterate through the previous steps until the difference between the old and new class mean is less than a certain threshold. The vague C-means algorithm is the implementation of the K-means algorithm on the basis of ambiguous mathematics. It is optimized by a ambiguous function to complete clustering. It does not think that each point can belong to a certain class, just like K-means clustering. Give each point a degree of subordination of the various types, and use the degree of subordination to better describe the characteristics of the edge pixels, and to deal with the inherent uncertainty of the things. Applying the characteristics of ambiguous C-means (FCM) non-monitoring ambiguous clustering calibration to stop image segmentation can reduce human intervention, and there are characteristics of uncertainty and ambiguity in the appropriate image.
The clustering approach should pay attention to several issues:
(1) How to determine the number of clusters.
(2) How to affirm the validity principle of clustering.
(3) How to set the initial value when the location and characteristics of the cluster center are not known beforehand.
(4) Expenses for computing.
And the FCM algorithm is extremely sensitive to the initial parameters, and sometimes requires the intervention of manual intervention parameters to approach the global optimal solution and improve the segmentation speed. In addition, the traditional FCM algorithm does not consider spatial information and is not sensitive to noise and grayscale.
Segmentation method based on wavelet transform
The fundamental idea of threshold image segmentation based on wavelet transform is that the histogram of the image is first synthesized into wavelet coefficients of different levels by dyadic wavelet transform, then the threshold threshold is selected according to the given segmentation principle and wavelet coefficient, and finally the threshold is applied. The area where the image is divided. The whole segmentation process is controlled from coarse to fine with scale changes, that is, the initial segmentation is done by the histogram projected on the rough L2(R) subspace. If the segmentation is not ideal, the histogram is applied in the precise subspace. The wavelet coefficients on the surface gradually refine the image segmentation. The computational feed of the segmentation algorithm varies linearly with the image size. Wavelet transform provides an accurate and unified framework for the analysis and characterization of signals at different scales. From the perspective of image segmentation, wavelet synthesis provides a mathematically complete depiction; wavelet transform can greatly reduce or eliminate the correlation between the extracted different features by selecting appropriate filters, not only with "zoom" Features, and there are fast algorithms on completion.
Characteristic
Wavelet transform is a multi-scale, multi-channel profiling tool. It is a local transform in the airspace and frequency domain. Therefore, it can effectively extract information from the signal. After the warfare warfare and other functions, the function or signal stops multi-scale analysis. Many problems that the Fourier transform cannot handle are handled. In recent years, multi-ary wavelets have been used for edge detection. In addition, wavelet transforms using orthogonal wavelet bases can also extract multi-scale edges, and can identify some types of edges by calculating and estimating the oddness of the image.
Mathematical morphology based segmentation
Mathematical morphology is a non-linear filtering method that can be used to suppress image processing problems such as noise, feature extraction, edge detection, and image segmentation. Mathematical morphology was first used to deal with binary images, and later used to deal with grayscale images. Nowadays, some scholars have begun to deal with computer vision problems with soft mathematical morphology and vague morphology. The characteristic of mathematical morphology is the ability to stop the synthesis of complex shapes and extract meaningful shape weights from useless information. Its fundamental idea is to apply a probe called a construction element to collect image information. When the probe moves from time to time in the image, not only can the structural features of the image be understood according to the interrelationship between the parts of the image, but also the application. Mathematical morphology fundamental operations can also structure many very effective methods of image processing and profiling. Its fundamental form operation is corrosion and shrinkage. Corrosion has the effect of reducing the purpose, increasing the purpose of the inner hole and eliminating the external isolated noise; and shrinking is the process of merging all the background points in the image that are in contact with the target object into the object, with the result that the purpose is increased and the aperture is reduced. The space in the destination can be added to form a connected domain. Another fundamental method of computation in mathematical morphology is open and closed operations. The open operation has the effect that the image is a small object, and the object is wary and slips away from the object at the border of the larger object; the closed operation has the function of filling the small space in the image of the object and connecting the adjacent object to the war-slip border.
Characteristic
Mathematical morphology is applied to image segmentation, and has the characteristics of good positioning effect, high segmentation precision and good anti-noise performance. At the same time, this method also has its own limitations: because in the preliminary work of image processing, the mathematical morphology of the open (closed) operation, after the image processing is stopped, there are still a lot of short lines and isolated points that do not match the purpose; Incomplete processing, it is also necessary to stop a series of point-based open (closed) operations, so the speed of operation is significantly reduced. How to combine mathematical morphology with other methods to restrain these defects will be the future work direction of mathematical morphology. The role of connecting the war-sliding borders of adjacent objects.
Segmentation method based on artificial neural network
In recent years, artificial neural network recognition technology has caused widespread concern and applied to image segmentation. The fundamental idea of the segmentation method based on neural network is to obtain a linear decision function by exercising the multi-layer perceptron, and then use the decision function to stop classifying the pixels to reach the purpose of segmentation.
Characteristic
Segmentation of images using artificial neural networks requires a large amount of exercise data. The neural network has a huge amount of convergence, easy to introduce spatial information, and can better deal with noise and unevenness in the image. The choice of network architecture is the main problem to be addressed in this approach.
Segmentation method based on genetic algorithm
Genetic algorithm (GA) is a search and optimization process that imitates natural selection and genetic mechanism. It has a strong global optimization search ability and is a self-compliant search method with universal applicability. It stops searching in the search space rather than on a single point in the search space, and it uses genetic manipulation rules rather than affirmative rules to work in the solution process. These characteristics make genetic algorithms well suited for image segmentation, especially threshold segmentation and region growing. Applying GA's global optimization ability and insensitivity to the initial position can improve the performance of image segmentation.
Characteristic
The difficulty of applying genetic algorithm to image segmentation is that the selection of the compliance function and the probability of interpolating and the probability of mutation are definitely determined. GA may also converge to partial optimality. It is conceivable to use a hybrid genetic algorithm that can self-adapt to the interpolation probability and mutation probability self-adapting genetic algorithm and the imitation annealing method.