Comparison of two-dimensional object shape recognition methods

Abstract aiming at the shape recognition of two-dimensional object in pattern recognition, the object shape in two-valued image is the main object, and the shape recognition method is comprehensively reviewed from two main aspects of feature extraction and classifier design, and the research actuality at home and abroad is analyzed. In particular, the latest research results obtained in recent years. Finally, the existing problems and the future research directions are pointed out.

keywords object shape recognition; feature extraction; classifier design

Chinese Figure method Classification number TP391.41

Comparison on methods of 2D object shape recognition

Abstract: In view of two-dimensional object shape recognition question in pattern recognition, the object shape in binary image was Taken as the main object. It summarizes the shape recognition methods based on the contour, the region and the neural network separately, Furthermor E, analyzes the situation of present in domestic and foreign simultaneously, especially obtained the newest Resea RCH results in recent years. Finally, the paper points out the current question existed as well as the direction of future.

Key Words: Object shape recognition; Feature extraction; Classifier Design

The object shape recognition [1] is a basic problem in pattern recognition, it is also an important problem, it is widely used in image analysis, computer vision and target recognition and other fields. Humans can easily identify the shape of an object, but it is very difficult for a computer to automatically recognize the shape of any object. The shape of an object is the basis of human visual system analysis and recognition of objects. Generally speaking, we focus more on the shape of the object, and the texture and color of the object is the second, so how to represent the difference between the shape and the comparison shape is very important in the field of machine vision application and research.

Object shape recognition generally includes objects in the image that are rotated, scaled, panned, distorted, obscured, affine, projective transformed, and identified by the shape of the object in the noisy image. Because of the complexity of the problem and the difficulty of its realization, most of the methods reported in this paper are only discussed in one or several of the above transformations.

At present, the representation and description of object shape recognition have been put forward in many ways at home and abroad. The following sections of this article are organized as follows: The 1th section introduces feature extraction, section 2nd introduces the design of classifiers, and the results of simulation experiments in section 3rd;

1 Feature Extraction

In object recognition, the original data volume of the image itself is quite large. If all the original characteristics are sent to the classifier, the classifier will be unusually complicated and the computational amount is huge. Therefore, it is necessary to decompose the shape of an object to produce a primitive and symbolize it, to form a characteristic vector or a symbol string, a graph, and thus to produce a pattern representing the object, a process called feature extraction.

1.1 Simple geometric invariance

Many methods are proposed to identify objects using various geometrical invariance:

Using corner feature [2], it is usually defined as a point where the curvature is high enough on the image boundary. Corner features have translation, rotation, and scaling invariance. It is only applicable to objects whose boundary angle is many and can represent the characteristic point of the object's shape.

The shape is represented by the equivalent curve class [3], which has the properties of translation, rotation, and scale invariant. It is used to determine whether they are of the same type, using the uniform variation of the polar radii of the shape's boundary points. It has no sensitivity to the disturbance of the boundary and is unaffected by the size, position and orientation of the shape. It should be based on the size of the shape and the actual accuracy of the need to select the appropriate scale, through the General Assembly to introduce redundant computation, slow execution speed, and too small to make the recognition accuracy is too coarse, leading to false identification, so the choice of appropriate scale is the key to this method.

The implicit polynomial curve [4] has many good properties to the object description, the invariant which is obtained based on the high-level implicit polynomial curve has better robustness to the object recognition, can overcome the influence of noise, and can recognize the target object with partial occlusion, but it is difficult to find the hidden polynomial curve invariants.

There is also a curvature function method [5], the contour of the object is represented by their curve function, and the arc length and tangent angle [6] are used to identify the object and so on.

It is easy to realize object recognition by geometric invariance, but the description of contour is too abstract to realize accurate identification or retrieval.

1.2 Gauss Descriptor

Gaussian descriptor [7] is a boundary-based shape feature, with a high recognition or matching rate, relative to the translation, rotation, scaling unchanged, less computational, moderate edge change and noise is not sensitive to a wide range of advantages.

Later, in the literature [8], the Gaussian descriptor is generalized, and a local Gaussian descriptor is proposed, which is applied to the object shape recognition to obtain a higher recognition rate. However, both Gaussian descriptors and modified local Gaussian descriptors have no affine invariance and need to be further improved.

1.3 Fourier descriptors

Fourier descriptor [9] has the characteristics of simple and high efficiency, and has become one of the important methods to identify the shape of the object. Its basic idea is that the shape of the object is assumed to be a closed curve, and the coordinate change of the P (L) on a moving point along the boundary curve X (L) +jy (L) (P (l) coordinates are represented in the plural form) is a function that takes shape boundary circumference as a period. This periodic function can be expanded into Fourier series representations. A series of coefficients in Fourier series are directly related to the shape of the boundary curve, called Fourier descriptors. When the coefficient item is taken to a sufficient order, it can extract and recover the shape information of the object completely.

Fourier descriptor is the Fourier transform coefficient of the shape boundary curve of an object, which is the result of the frequency domain analysis of the object boundary curve signal. According to the properties of Fourier transform, Fourier descriptors are related to shape scale, direction and starting point. Therefore, in order to identify shapes with rotation, translation and scale invariance, Fourier descriptors need to be normalized. In short, Fourier descriptor method is used to identify the image contour, only applicable to the closed boundary, and can not reflect the internal characteristics of the region, in the face of more complex images, the image has more than the target or the contour of the image, this method of recognition is not ideal.

1.4 wavelet Descriptors

Wavelet transform [10-11] has the advantages of space-frequency locality, directivity and multi-resolution, and is applied in many fields such as signal processing, image processing, pattern recognition and so on. Wavelet transform is a multi-resolution transform, which decomposes images on different scales. When applying wavelet to contour representation, it is necessary to select the limit series of wavelet coefficients to describe the contour, and to normalized the wavelet coefficients to achieve the requirements of translation, scaling and rotation invariance. The wavelet coefficients are described as wavelet descriptors, which make up some shortcomings of Fourier transform. In the object shape recognition, the single wavelet transform is based on the boundary and the computational amount is large.

The four methods mentioned above are boundary-based. In the method of object shape recognition based on boundary, it is difficult to obtain the ideal effect in application because the detection, representation and subsequent calculation of contour are often unstable. How to overcome the difficulty of the existing boundary-based feature representation and propose a new stable and feasible method is a challenging problem in the field of image shape representation and recognition.

1.5 Methods of the skeleton

The core idea of the Skeleton method [12] is to describe its shape using the topological relationship of the object's middle axis or skeleton. It can be used to identify objects that are rotated, panned, scaled, and anti-noise. The disadvantage of the skeleton method is that the skeleton itself is not easy to get, the stability is not ideal, especially for the more complex shape of the object.

1.6 Moment invariants

Moment invariants are the characteristics of moments in which the image of an object is shifted, rotated, and the scaling transformation remains constant. The recognition of object shape by using moment invariants is an important method in pattern recognition. HU[13] In 1961, the definition of the moment of continuous function and the basic properties of the moment are presented, and the properties of translation, rotation and scaling invariance of the moment are proved, and the expressions of seven invariant moments with translation, rotation and scaling invariance are given in detail. Hu established a moment invariant, requiring all pixels in the target area to participate in the calculation, although some scholars have studied the fast algorithm of moments, but they are quite time-consuming. LI[14] Using the invariance of the Fourier-mellin transform, a method of constructing arbitrary order moment invariants is deduced, and it is pointed out that Hu's moment invariants are a special case. TEAGUE[15] It is suggested to use orthogonal polynomial to construct the orthogonal moment to overcome the disadvantage that Hu's moment invariants contain a large amount of redundant information. Orthogonal moment is better than other kinds of moments in information redundancy, image expression and recognition effect. Zernike moment is a kind of orthogonal invariant, because it has orthogonal base, and it is easy to construct higher moments, so it is widely used. On the basis of the region-based Hu moment invariants, the moment is generalized [16] to construct some new moment invariants, such as the contour moment invariants [17], the polar radius invariant moment [18], the relative boundary moment [19], and so on. The moment features are calculated in the whole image space, and the global characteristics of the image are obtained, which is susceptible to noise disturbance. And it only applies to images with obvious differences, thus improving the ability to differentiate similar objects becomes the key to solving such problems.

The above mentioned two methods are based on the region, the region-based representation and the gray value of the image is closely related to the non-uniform illumination and other factors, and due to the calculation of the entire region, the computational capacity is very large. On the other hand, because it considers the internal structure of an object, it contains more information about the shape of the object than the boundary-based method, so it is more stable and has high recognition rate.

1.7 Small Wave Moment

The wavelet moment [20-21], which combines moment characteristics and Potte, reflects both the global information of the image and the local information of the image. The algorithm not only solves the problem that the feature quantity changes with the image rotation, translation and zooming, but also improves the recognition ability of the approximate object, and has strong robustness, which greatly strengthens the analysis ability of the image fine degree. It has the advantages of high recognition rate, especially in the case of small difference of similar objects, strong anti-noise, but it does not recognize objects with occlusion and distortion.

1.8 Independent Component Analysis (ICA) [22]

With n-type objects to be recognized, the image pixels of the same kind object in the ICA pretreatment are arranged into n koriyuki vectors, and the k*n matrix is formed for K-similar training samples. This set of observation signals is composed of a linear mixture of the D independent components. The independent component analysis of this matrix separates the D independent component, which forms a set of the characteristic space, and the subspace of the D-base vector spanned forms the characteristic space of the Class I object. Because of the N-class objects, we can get the characteristic space composed of the independent components of N group, and each group base describes the characteristics of the corresponding object class.

ICA in the case of small sample training, with the ability to quickly extract sample characteristics, can identify the image with missing and deformed objects, and insensitive to noise disturbance, which is very important in the practical application of target recognition.

1.9 Principal component Analysis (PCA) [23]

In the field of image recognition, the dimension of the original data x of the input is n, it is hoped that by preprocessing the M (<n< span= "" >) Dimension Data y, if there is no restriction, the x is simply truncated, then the mean squared error will be equal to the sum of the variance of each component. To achieve this, we hope to obtain a linear transform w, so that the truncation of WX is optimal under the minimum mean square error, which requires that the discarded component has a lower variance, while the reserved component has a higher variance, and PCA is the method to find the linear transformation. It is based on the K-L decomposition, the purpose is to find a set of vectors in the data space to interpret the variance of the data as much as possible, through a special vector matrix, the data from the original high-dimensional space projection into a low-dimensional vector space, the reduction of dimension after saving the main information of the data, making the data easier to process.

PCA is the optimal transformation in the sense of minimum mean square error, it can achieve the best effect in eliminating the correlation between the pattern features and highlighting its difference. But the image features extracted by PCA do not have the invariance of displacement, scale and rotation. In the process of PCA recognition, all the pixels in the whole image are involved in the operation, which is more suitable for complex image recognition, but the image size is consistent.

1.10 Method of Circle decomposition

The circular decomposition method [24] can be regarded as an extension of the traditional boundary-based method, which preserves the advantages of such methods in detail description, and extends their descriptions from the boundary to the whole shape area, thus can be regarded as the combination of the boundary-based and region-based methods.

The circular decomposition method has a strong ability to describe both global and local information, which has the features of translation, rotation, scaling invariance, and good resistance to deformation, occlusion and random noise, but not for distorted object shape recognition and recognition rate is not very good.

1.11Radon transformations

The Radon transform [25-26] is a transformation method that calculates the projection of the image f (x, y) along a specified angular direction. Projecting in the specified direction is the line integral of the two-dimensional function in that direction, as well as the projection on the horizontal axis after the angle of the image is rotated clockwise. Due to its inherent advantages of good noise resistance, it is advantageous to use it as an effective method of image analysis in the environment with noise source.

After the radon transformation of the image, the main advantage is that the identification problem can be reduced from two to one dimension, so that it can be treated as a one-dimensional signal, greatly improving the processing speed, and the extracted image features have rotation, translation and scaling invariance, but after processing the general need and wavelet, moment and other techniques combined practical will be more effective.

2 Classifier Design

On the premise that the D-dimensional feature space has been determined, the classifier design problem is a choice of what criteria, using what method, the identified D-dimensional feature space is divided into decision-making domain problem. The classifier with high classification accuracy, low error rate and good reliability is the ultimate goal of recognition.

2.1 BP Neural network

BP neural network [27] is the most widely used neural network model in Pattern recognition classification, and the network with hidden layer can complete arbitrary segmentation of multidimensional space. BP network adopts error-Reverse propagation learning algorithm, which is widely used in function approximation, pattern recognition, classification, data compression and so on. It is a multilayer neural network with three or three layers, each of which consists of a number of neurons. It is trained according to the teacher Learning mode, when a pair of learning modes are provided to the network, the activation value of its neurons will propagate from the input layer to the output layer through the intermediate layers, and the outputs of each neuron in the output layer correspond to the network response of the input mode. Then, according to reduce the hope output and actual output error principle, from the output layer through the middle layer, and finally back to the input layer to correct the connection rights.

BP algorithm is a local optimization method in essence, and there are two big defects in the convergence process: one is the slow convergence speed and the other is the problem of "local minimum point". In the learning process, sometimes occurs when the learning is repeated to a certain number of times, although the actual output of the network and hope that there is still a large gap in the output, but no matter how to learn, the network global error reduction speed has become very slow, or no longer change, this phenomenon is due to the network convergence to local minimum point. If the number of cells in the middle layer of the BP network is improved properly, or a small random number is added to each connection right, it is possible to avoid the local minimum point in the convergence process.

2.2 Genetic BP Neural network [28-29]

Because the BP neural network algorithm uses the gradient under the method, it is easy to fall into the local minimum and the training time is longer. Genetic algorithm (GA) uses heuristic search technique to find the optimal solution, which has the advantages of good robustness, high search efficiency, less restriction on the target function, and easy parallel parallel machine operation. The Genetic BP network algorithm synthesizes the global optimization of genetic algorithm and the parallel computation of neural network, which can overcome the disadvantage that the genetic algorithm eventually evolves to the optimal solution and the neural network is prone to local solution, and has better global and convergent speed.

The basic idea of this algorithm is: First, GA solves the optimization problem, because GA is a group of points that search the solution space at the same time, and constitute the continuous evolution of the group sequence, so after the evolution of certain algebra, we can get some global better, from these good points, and then use neural network to solve, and then get the global excellent solution.

2.3 Wavelet Neural Network

wavelet neural network [30-31] is a feedforward network with wavelet function as the neuron excitation function, which can be regarded as a function-connected network based on wavelet function, and also can be regarded as the generalization of radial basis function network.

The wavelet neural network model includes input layer, output layer and hidden layer. The hidden layer consists of two nodes: the wavelet base node and the scale function node. Wavelet Neural Network is a neural network based on wavelet analysis, it makes full use of the good localization property of wavelet transform and combines the self-learning function of neural network, so it has strong approximation and fault-tolerant ability, it avoids the blindness and local optimal nonlinear optimization problem of BP neural network structure design, and simplifies training greatly. Has a strong function of learning ability. It has good time-frequency localization characteristics, and its application to object shape recognition has better recognition effect than traditional neural network, but it is affected by noise.

2.4 Self-organizing competitive artificial neural network (SCNN)

SCNN[32] Based on the "lateral inhibition" structure of biological neurons, consisting of a single-layer neural network, the input node and the output node are fully connected. Because the competition characteristic of the network in the learning process is manifested in the output layer, its output layer is also called the competition layer, and the active function of the competition network is a two value type function. SCNN uses the Kohollen learning rules to train and judge the input patterns, and finally to identify and classify the various targets.

It can overcome the disadvantage that BP network is not easy to converge, the learning time is long and so on, the training and recognition is completed simultaneously, it has real-time performance and high efficiency.

2.5 convolutional Neural Network (CNNs)

CNNS[33] is a method used in the field of two-dimensional image processing, pattern recognition and machine vision in recent years.

Most of the classification methods are feature-based, which means that certain features must be extracted before they can be resolved. However, the feature extraction shown is not easy and is not always reliable in some application issues. The CNNs classifier avoids display feature extraction and can be implicitly learned from the training data. It combines feature extraction function into multi-layer perceptron through structural recombination and reduction of weight value. The output of a layer, a feature map, forms the input of the next layer, sharing a common set of weights (also known as convolution letters) with the neurons associated with the same feature graph. At the last layer, the feature map is categorized by a single-layer perceptron, which is usually fully connected. This makes it different from other neural network based classifiers, and it can directly work in grayscale images, so that it can be directly used to deal with the classification of object images.

2.6 Support Vector Machine (SVM)

SVM[34-35] is a machine learning method based on the principle of structural risk minimization, which uses kernel functions to map input vectors to a high-dimensional feature space, and then constructs an optimal super-plane in the space to approximate the classification function. Its basic idea can be summed up as follows: firstly, the input space is transformed into a high-dimensional space by nonlinear transformation, then the optimal linear classification surface is obtained in this new space, and the nonlinear transformation is realized by defining the appropriate intrinsic product function.

SVM overcomes many inherent defects of neural network, such as easy to learn or fall into local minima, and has excellent learning ability and generalization ability for data analysis of small sample data. It has a good recognition rate for the translation and rotation of the object, and it has strong robustness, but the recognition rate of the scaled object is somewhat lower. To solve the problem of object recognition and classification of big data, how to improve the real-time of data processing and shorten the time of training sample is still an urgent problem.

3 Simulation Experiment

in the environment of MATLAB 7, some simulation experiments are carried out: feature extraction with wavelet moment, and according to the characteristics of high wavelet moment feature dimension, this paper adopts the feature selection method of combining dispersion degree and sequential forward method proposed by aspect [36], respectively using BP neural Network (BP), Genetic BP Neural Network (GABP) as a classifier. In this experiment, two types of two-valued aircraft images are used for identification, 1 are F1 and F2, the image size is 512*512 pixels and the format is BMP. Each type of aircraft are taken different translation, scaling, rotation of the transformation of a total of 40 images, and then from each type of aircraft selected three geometric transformation of each of the 10 are added mean 0, the variance is 0.01, 0.05, 0.1 Gaussian noise total 60. Therefore, 100 images were obtained in this experiment, of which 20 (10 pieces per class) were not added to the noise of the image for training, the other 80 were used for testing.

Summarize

This paper introduces the methods of object shape recognition from two aspects of feature extraction and classifier design, and each method references a certain reference, so this paper omits the formula of correlative method, and only shows the thinking and advantages and disadvantages of the method in recognizing the object. Through the above introduction, it is known that, so far, there is still no universal method to identify objects in various environments. These methods can recognize the corresponding object only under certain conditions, and most of the methods are not implemented by a single technique. More importantly, all algorithms need to be improved to enable them to adapt to common occasions and improve recognition rates and to recognize three-dimensional objects. Due to the detection and representation of the contour, the boundary-based method is often unstable; because the computation is oriented to the whole region, the computation of the region-based method is very large, and the neural network is an intelligent method, and many of its technologies are immature. Therefore, the main work in the future is to continuously improve and improve these methods.

For beginners in object recognition, this article is of great reference value.

Reference documents

[1] Ding Xianfeng, Wu Hong, Hongjiang, MA Songde. A survey of shape matching [J]. Journal of Automation, 2001,27 (5): 678-694.

[2] Mehrotra R,nichani s.corner detection[j]. Pattern recogni tion, 1990,23 (11): 1223-1233.

[3] Chen Xiaochun, Ye Yingdong, Nussen min. A method of shape recognition [J]. Pattern recognition and artificial intelligence, 2006,9 (6): 758-763.

[4] Wu Gang, Li Daolun. Target recognition based on implicit polynomial curve affine invariants [J]. Journal of Electronics, 2004,32 (12): 1987-1992.

[5] Bandera A, Urdiales C. 2D object recognition based on curvature functions obtained from local histograms of the Contou R chain Code[j]. Pattern recognition letters,1999,20 (1) 4955.

[6] Dong Xu,wenli Xu. Description and recognition of object contours using arc length and tangent orientation[j]. Pattern recognition Letters, 2005,26 (7): 855-864.

[7] Liu Yishu, using Gaussian depiction of characters for character recognition [J]. Computer Applications, 2006,26 (11): 2778-2780.

[8] Liu Yishu. Regional Gaussian delineation and its application in object recognition [J]. Computer Engineering and Design, 2007,28 (15): 3650-3652.

[9] Wang Tao, Liu Wenyin, Sun Jia Guang, Hongjiang. Fourier descriptor recognition of the shape of the object [J]. Computer research and Development, 2002,39 (12): 1714-1719.

[Ten] Tieng qm,boles WW. Recognition of 2D object contours using the wavelet transform zero-crossing representation[j]. IEEE Transactions on Pattern analysis and Machine Intelligence, 1997,19 (8): 910-916.

[11] Chinese, Baiyun. Wavelet transform with translation, scale and rotation invariance [J]. Signal Processing, 2004,20 (2): 147-152.

[Cecilia] DR. Recognition of shapes by attributed skeletal graphs[j]. Pattern recognition,2004,37 (1): 21-31.

[] Hu MK. Visual pattern recognition by moment invariants[j]ieee transactions on Information theory,1962,8 (2): 179-18.

Li y.refarming The theory of invariant moments for pattern recognition[j]. Pattern recognition,1992,25 (7): 723-730.

[Teagure] MR. Image analysis via the general theory of Moments[j]. Journal of the Optical Society of AMERICA,1979,70 (8): 920-930.

[16] Liu Jin, Zhang Tianxi. Extension of image invariant moment [J]. Computer Science, 2004,27 (5): 668-674.

[17] Liu Yishu, Yang Lihua, Lynch. Profile moment invariants and its application in object shape recognition [[J]. Chinese Journal of Image Graphics, 2004,9 (3): 308-313.

[18] Cao Maoyong, Sun Nong bright, Yu dao silver. Polar Radius invariant moment for pattern recognition [J]. Journal of Computer Science, 2004,27 (6): 860-864.

[19] Li Lihong, Miao Jingli, Wang Jingji, An Qingbin. Application of relative boundary moment in pattern recognition [J]. Micro Computer Information, 2005,21 (7): 3-4.

[Dinggan] s,horace H S ip.discriminative wavelet shape descriptors for recognition of 2-d Patterns[j]. Pattern recognition,1999,32 (2): 151-165.

[21] Zhang Hong, Chen Wenxian. An image recognition method based on wavelet moment [J]. Journal of Beijing University of Technology, 2004,30 (4): 427-431.

[22] Rui, Shen Chunlin, Qi Tian, James Ding. Feature invariance target recognition based on ICA [J]. Small microcomputer system, 2005,26 (3): 505-50.8

[23] Dong, Gao, Hu Liangmei, Dong Wenwen. Research on Shape feature extraction and recognition based on principal component analysis [J]. Journal of Hefei University (natural Science Edition), 2003,26 (2): 176-179.

[24] Wang Yifei, Chen Yanqiu. Circular decomposition method for two-dimensional shape description [J]. Computer Science, 2006,33 (11): 228-233.

[25] Wang Yaoming, Shangwei, Shikuang. Image moment feature extraction based on radon transform and its application in image recognition [J]. Computer Engineering, 2001,27 (2): 82-84.

[26] Yx, deterministic, Zhang Hongzai, Tri Jianfeng, Tripeling. Research on image recognition based on radon transform [J], Computer application research, 2004,21 (5): 207-209.

[27] Zhang Hua, Zhang Miao, Liu Wei, Mengxiang increase. Image shape recognition based on BP neural network. Computer Science, 2006,33 (1): 269-271.

Maniezzo v. Genetic evolution of the topology and weightdistribution of neural network[j]. IEEE Transactions on Neural networks,1994,5 (1): 39-53.

[29] Guo Xiaoting, Zhu Yan. Evolutionary neural network based on genetic algorithm [J]. Journal of Tsinghua University (natural Science Edition), 2000,40 (10): 116-119.

[Qinghua,benveniste.a] Zhang Wavelet network[j]. IEEE Trans on Neural networks,1992,3 (6): 889-898.

[31] Wang Aming, Liu Tian Fang, Wang Yu. Wavelet Neural network model for image and pattern recognition [J]. Journal of Chinese Mining University, 2002,31 (5): 382-385.

[32] Bang, Feng Daiyi, Yang Baihui. Multi-object image recognition based on wavelet moment and SCNN [J]. Computer Engineering and applications, 2006,42 (17): 90-93.

[33] Li Baoqing. Pattern classifier based on convolutional neural network [J]. Journal of Dalian University, 2003,24 (2): 19-23.

[34] He, Tao Qing, Fangtingjian. Genetic algorithm for optimizing perceptron based on statistical learning theory [J]. Pattern recognition and artificial intelligence, 2001,14 (2): 211-215.

[35] Zhang Xue. On statistical learning theory and support vector machine [J]. Journal of Automation, 2000,26 (1): 32-42.

[36] aspect, Zhou Yuanhua. Pattern Recognition method based on wavelet invariant moment [J]. Journal of Infrared and Millimeter waves, 2000,19 (3): 215-218.

Reprint to Http://blog.chinaunix.net/uid-24517893-id-3140811.html

Comparison of two-dimensional object shape recognition methods