A survey of image segmentation methods

A survey of image segmentation methods

Tri Yue

(School of Software, Xi ' an Jiaotong University, Shaanxi 710049)

absrtact: By retrieving the image segmentation method which has been improved in recent years, the following kinds of methods are more active. are: threshold-based segmentation, segmentation based on region growth, segmentation based on wavelet transform, segmentation method based on neural network, ability functional-based segmentation method, probability-based segmentation method and segmentation method based on specific theory. In this paper, the basic principles of various methods are expounded, and some new research results based on these segmentation methods are introduced.

Keywords: image segmentation image processing

Image segmentation refers to the process of dividing images into different regions according to certain similarity criteria, which is one of the fundamental problems in the fields of computer vision and image processing, and it is the preprocessing of image classification, scene parsing, object detection, image 3D reconstruction and so on. The research has been one of the hot topics since the 1960s. , and is widely used in medical image analysis, traffic control, meteorological prediction, geological exploration, face and fingerprint recognition and many other fields.

First, the threshold-based segmentation method

Gray threshold Segmentation is one of the most commonly used parallel region techniques, which is the most applied number in image segmentation. The threshold segmentation method is actually the following transformation of the input image F to the output image g:

where T is the threshold value, for the object's image element, g (i,j) = 1, for the background of the image element, g (i,j) = 0.

Therefore, the key of threshold segmentation algorithm is to determine the threshold value, if a suitable threshold can be determined to accurately separate the image. After the threshold value is determined, the gray value comparison and pixel segmentation of the threshold and pixel points can be carried out in parallel to each pixel, and the result of segmentation is given directly to the image area.

The advantages of threshold segmentation are simple calculation, high operation efficiency and fast speed. It is widely used in applications where computational efficiency is a concern (e.g. for software implementations).

The common threshold segmentation method mainly includes the maximal inter-class variance method (Otsu algorithm), which is based on the threshold segmentation method of entropy, the minimum error method, the Symbiosis matrix method, the moment holding method, the probability relaxation method, the fuzzy set method and the threshold segmentation method combined with other methods.

Based on Otsu algorithm, high sensitivity [1], after a lot of experiments and analysis, explained the essence of Otsu segmentation failure in complex background, put forward the idea of restricting the pixel and gray level of background area, improved the Otsu algorithm, and achieved good results in practical application.

Rong Jianwu [2] is an adaptive threshold segmentation algorithm for estimating background in Gaussian scale space for effectively segmenting heterogeneous illumination images. Firstly, a Gaussian scale space is constructed by using the two Gaussian function to deal with the image, and the background estimation is carried out in this space, and the background difference method is used to eliminate the non-uniform illumination interference to extract the target image.

Liu Ding "5" and so on to improve the segmentation accuracy of the bright halo of the silicon diameter detection image, a two-dimensional histogram region oblique Division multi-Threshold segmentation method based on multi-objective artificial fish swarm algorithm is proposed.

The segmentation method based on region growth

The basic idea of regional growth is to assemble pixels with similar properties to form a region. In particular, a seed pixel is used as the starting point for each area that needs to be segmented, and then the pixels surrounding the seed pixels with the same or similar properties as the seed pixels (judged by some predetermined growth or similarity criteria) are merged into the region where the seed pixels are located. Use these new pixels as new seed pixels to continue with the above process until there are no more pixels that meet the criteria to be included. Such an area would have grown.

Li Qi [13] By combining the gradient vector flow (GVF) field with the seed region growth (SRG) method, a new fast automatic image segmentation algorithm is proposed. This method first constructs a flow scalar field based on the gradient vector flow field, then proposes a new fast seed region growth segmentation Method--Fast Sweep method (Fastscanning method, FSM) for initial segmentation of scalar field, Finally, the region adjacency graph is used to combine the results of the initial segmentation to obtain the final result. The method is characterized by fast segmentation speed.

Three, the segmentation method based on wavelet transform

Wavelet transform is a widely used mathematical tool in recent years, it has good localization in both time and frequency domain, and wavelet transform has multi-scale characteristic, it can analyze signal at different scales, so it is applied in many aspects such as image processing and analysis.

The basic idea of thresholding image segmentation based on wavelet transform is that the histogram of image is decomposed into different levels of wavelet coefficients by two-step wavelet transform, then the threshold thresholds are selected based on the given segmentation criteria and wavelets coefficients, and finally the region of image segmentation is marked by the threshold value. The whole segmentation process is from coarse to fine, controlled by the scale change, that is, the starting partition is realized by the histogram projected on the rough L2 (R) subspace, if the segmentation is not ideal, the image segmentation is gradually refined by using the wavelet coefficients of the histogram in fine subspace. The calculation of the segmentation algorithm will change linearly with the size of the image.

Sun Chao Male [14] and so on by the color image of the temperature paint, wavelet transform processing, extracting wavelet eigenvalue, and color information together as a characteristic value of fuzzy clustering. Compared with the traditional method of image segmentation by using wavelet transform or fuzzy clustering alone, this algorithm has a good effect on the segmentation of color image with temperature paint.

Iv. the segmentation method based on neural network

In recent years, artificial neural network recognition technology has aroused wide attention and applied to image segmentation. The basic idea of the segmentation method based on neural network is to get the linear decision function by training the multilayer perceptron, then classify the pixels by the decision function to achieve the purpose of segmentation. This approach requires a lot of training data. There are a huge amount of connections in neural networks, and it is easy to introduce spatial information, which can solve the problem of noise and inhomogeneity in image. Which network structure to choose is the main problem that this method solves.

In order to further extend the application of the pulse-coupled neural network (Plulse coupled neural Network, PCNN) in the image segmentation, Zhou Dong [10] has simplified and improved the PCNN model, using the relationship between the threshold value and the corresponding region mean of the pulse output, A method to optimize the connection coefficients is proposed, which results in the segmentation of the model by iterative method.

Tang Siyuan [15] in order to improve the traditional BP neural network for medical image segmentation of the initial weight value of the existing sensitivity, learning rate fixed, slow convergence and easy to fall into the local minimum, a method based on improved particle swarm optimization algorithm BP neural network medical image segmentation.

Five, based on the energy functional partition method

The method mainly refers to the active contour model (active contour model) and the algorithm developed on it, the basic idea is to use continuous curve to express the target edge, and define an energy function to make its independent variable including The edge curve, Therefore, the segmentation process is transformed into the process of solving the minimum value of the energy function, which can be obtained by solving the Euler (Euler) corresponding to the functions. Lagrange) equation to achieve the curve position where the energy reaches the hour is the contour of the target.

Sho Chunxia [3] and others combine the advantages of two kinds of edge stop function (Edge stop function based on Gaussian mixture model color distribution and edge stop function defined on multi-scale image gradient), a hybrid model of edge stop function is proposed, which is based on the image Color and Edge feature adaptive guidance level set function evolution. This algorithm can not only detect the texture target area effectively, but also effectively calculate the precise and smooth boundary of the texture region.

Zhang Jichun [4] and other people, in order to improve the level set image segmentation quality and reduce the level set iteration number, proposed a new energy formula and level set function. Based on the discretization of rough set data, the Discretization method for the data is introduced, the new energy function and kernel function are weighted according to the information of the discrete region of the image, and the original discrete image is mapped to the high-dimensional space, so that the model can deal with many kinds of images and even a certain signal-to-noise ratio.

Object [6] A new multi-Atlas Active Contour Model framework is proposed, which effectively utilizes the prior information of the Atlas and the Gray information of the image to be segmented, and introduces the multi-map shape priori into the active contour model and uses the active contour model to correct the errors caused by the registration in the process of fusing the labeled images. Get smooth and accurate segmentation results.

Zhang Fan [9] and other materials in the study of the dislocation theory into the level set method, the use of dislocation dynamics mechanism to drive the level set curve evolution of the dislocation of the configuration force, can effectively avoid the local image gradient anomalies in the case of the curve stop evolution phenomenon, or avoid local boundary leaks at weak edges due to smaller image gradients.

Vi. segmentation method based on probability statistics

At present, the algorithm of image segmentation based on probability statistics can be divided into three kinds of models, one is standard Gaussian mixture model, the other is the implicit Markov random field model using Gibbs probability distribution, the spatial information is introduced through the Pixel neighborhood to mark the pixels in category. The third is to use Markov random field to calculate the prior distribution to obtain the category tag, taking into account the gray information and spatial information of the pixel [8].

In order to solve the problem that the result of MRF model is prone to smooth phenomenon, Song Yantao [8] proposes a new Markov random field image segmentation model based on image weight method. By introducing weights into the image slice, using KL distance to introduce the penalty of entropy, the algorithm has strong self-adaptability, can overcome the influence of noise on the segmentation result, and obtains the high segmentation precision.

Seven, the segmentation method based on the specific theory

There is no universal theory of image segmentation. With the development of new theories and methods of each subject, the method of image segmentation with some specific theories and methods is presented, including: Image segmentation method based on clustering analysis, segmentation method based on fuzzy set theory, etc.

Dong Zholi [7] A color image segmentation algorithm based on two-segment multi-component graph cutting is proposed. The new algorithm is based on the map and ml (maximum likelihood) estimation framework, using a multi-component strategy instead of a regional re-tagging, each map estimate, the same partition under the non-adjacent area is no longer marked with the new label, but processing as a plurality of components of the split, so that the number of labels is no longer incremented.

Chen Ziyang [11] An image segmentation method based on three-dimensional histogram and suppression fuzzy Kohonen Clustering Network (RFKCN) is proposed. The method firstly blurs the pixels, constructs two redundant images through the fuzzy mean and fuzzy median values, and then makes up a three-dimensional feature vector set by the redundant image and the original image, and uses the RFKCN clustering network to cluster the feature vectors to achieve the purpose of image segmentation.

Zhiwei Tang[12] An image segmentation algorithm based on the maximal inter-class variance method and the fuzzy theory is proposed. In this method, the image is divided into target area, background area and fuzzy region by pre-segmentation, then further processing of fuzzy region is divided into target area and background area. This method can preserve the detail in a large degree, and can achieve good segmentation effect when the SNR is low and the contrast is poor, but the operation time is longer.

Ix. Conclusion

In view of the current image segmentation methods, different methods have their own specific categories, and there is no universal segmentation method. In the case of the bottleneck of single theory research, the fusion of different methods and the combination of subject theory knowledge become the direction to seek breakthroughs and get good results. With the rise of deep learning and neural network, and the wide application, the field of image segmentation has achieved good results. With the emergence and popularization of unmanned driving, automatic navigation, face recognition and so on, the system has more and more requirements for image segmentation, and image segmentation will be one of the hotspots in the future, which has a broad prospect.

References:

[1] Gao min, Li Hui sheng, Zhou Yulong, Fontaine, Zheng full. Background constraints on the target segmentation of tanks under infrared complex background. Journal of Automation, 2016, 42 (3): 416-430

[2] Rong Jianwu, Hyun Jing, Huihui, Chen Haipeng. Adaptive threshold segmentation algorithm for estimating background under Gaussian scale space. Journal of Automation, 2014, 40 (8): 1773-1782

[3] Sho Chunxia, first rain, Zhang Qing. Gaussian mixed function region matching guided level set texture image segmentation. Chinese Journal of Computer Science, 2010,33 (7)

[4] Zhang Jichun, Guo Wu. Horizontal set image segmentation based on rough set and new energy formula. Journal of Automation, 2015, 41 (11): 1913-1925

[5] Liu Ding, Zhang Xin Rain, Chen Yajun. Threshold segmentation method of silicon single crystal diameter detection based on multi-objective artificial fish swarm algorithm. Journal of Automation, 2016, 42 (3): 431-442

[6] object, Lu Zhentai, Zhang, Yang Wei, Chen Wufan, Zhang Yu. Segmentation of brain image based on multi-atlas active contour model. Chinese Journal of Computer Science, 2016,39 (7)

[7] Dong Zholi, Li Lei, Zhang Dexian. An unsupervised color image segmentation algorithm based on two-stage multi-component graph cutting. Journal of Automation, 2014, 40 (6): 1223-1232

[8] Song Yantao, Guize, Sun Guanseng. A brain MR image segmentation algorithm based on the Markov random field of image chip. Journal of Automation, 2014, 40 (8): 1754-1763

[9] Sail, Zhang. An image segmentation algorithm based on dislocation theory for distance regularization level set. Journal of Automation, 2018, 44 (5): 943-952

[10] Zhou Dong, climax, Guo Yongjia. A simplified PCNN image segmentation method with parameter self-adaptation. Journal of Automation, 2014, 40 (6): 1191-1197

[11] Chen Ziyang, Wang Paoping. An image segmentation method based on three-dimensional histogram and RFKCN. Chinese Journal of Computer Science, 2011,34 (8)

[Zhiwei] TANG, Yixuan WU. One image segmentation method based on Otsu and fuzzy theory seeking image segment threshold. International Conference on Electronics, Communications and Control (ICECC). 2011, 2170-2173

[13] Li Qi, Luo, Shand. Image segmentation based on flow direction scalar field and fast sweep method. Journal of Automation, 2008,34 (8)

[14] Sun Chao Male, Ishin, Tri Li. The application of wavelet transform combined with fuzzy clustering in color image segmentation of temperature-warming paint is presented. Journal of Software,2012,23 (Suppl. (2)): 64?68

[15] Tang Siyuan, Jun Junfeng, Yang Min. A new method of medical image segmentation based on BP neural network. Computer science. 2017,44 (6A)

A survey of image segmentation methods