CN113191361B - A Shape Recognition Method - Google Patents

Abstract

The present invention proposes a shape recognition method, which extracts contour key points of shape samples; defines approximate offset curvature values at each key point and judges the concavity and convexity of key points to obtain candidate segmentation points; adjusts the curvature screening threshold to obtain the shape Segmentation points; calculate the minimum segmentation cost for shape segmentation, and obtain several sub-shape parts; construct the topology structure of shape samples; use the full-scale visual representation method of the shape to obtain the feature expression image of the corresponding sub-shape part; input each feature expression image into the volume Construct the feature matrix of the shape sample; construct the graph convolutional neural network; train the graph convolutional neural network to obtain the feature matrix and adjacency matrix of the test sample, and input it to In the trained graph convolutional network model, shape classification and recognition are realized.

Description

A shape recognition method

Technical Field

The invention relates to a shape recognition method and belongs to the technical field of shape recognition.

Background Art

Contour shape recognition is an important research direction in the field of machine vision. Using object shape features for target recognition is the main research topic of machine vision. The main result of this research is to fully extract the target shape features by improving the shape matching algorithm or designing effective shape descriptors for better similarity measurement. This has been widely used in engineering, such as radar, infrared imaging detection, image and video matching and retrieval, robot automatic navigation, scene semantic segmentation, texture recognition and data mining.

Usually, the expression and retrieval of contour shapes are based on manually designed shape descriptors to extract target contour features, such as Shape Contexts, Shape Vocabulary, and Bag of contour fragments. However, the shape information extracted by manual descriptors is usually incomplete, and it cannot be guaranteed that the descriptors are robust to changes in local changes, occlusions, and overall deformation of the target shape. Designing too many descriptors will lead to redundant feature extraction and high computational complexity. Therefore, the recognition accuracy and efficiency are low. In recent years, as convolutional neural networks have achieved good results in image recognition tasks, they have begun to be applied in shape recognition tasks. However, since contour shapes lack information such as surface texture and color that images have, the recognition effect of directly applying convolutional neural networks is low.

In view of the above problems of shape recognition algorithms, how to provide a target recognition method that can accurately classify the target contour shape is a problem that urgently needs to be solved by those skilled in the art.

Summary of the invention

The present invention is proposed to solve the problems in the prior art, and the technical solution is as follows:

A shape recognition method, the method comprising the following steps:

Step 1: Extract the contour key points of the shape sample;

Step 2: define the approximate bias curvature value at each key point and determine the concavity and convexity of the curve at the key point to obtain candidate shape segmentation points;

Step 3: Adjust the curvature screening threshold to obtain the shape segmentation point;

Step 4: segment the shape based on the principle that the segmentation line segments are located within the shape and do not cross each other, and obtain several sub-shape parts with the minimum segmentation cost;

Step 5: construct the topological structure of the shape sample;

Step 6: Use the full-scale visualization method of the shape to obtain a feature expression image of the corresponding sub-shape part;

Step 7: Input each feature expression image into the convolutional neural network for training, and learn to obtain the feature vector of each sub-shape part;

Step 8: construct a feature matrix of shape samples;

Step 9: Construct a graph convolutional neural network;

Step 10: Train the graph convolutional neural network, perform shape segmentation on the test sample, obtain the feature vector of each sub-shape part, calculate the feature matrix and adjacency matrix of the test sample, and input them into the trained graph convolutional network model to realize shape classification and recognition.

Preferably, in step 1, the method for extracting contour key points is:

The contour of each shape sample is composed of a series of sampling points. For any shape sample S, sampling n points of the contour yields:

S={(p _x (i), p _y (i))|i∈[1, n]},

Where p _x (i), p _y (i) are the horizontal and vertical coordinates of the contour sampling point p (i) in the two-dimensional plane, and n is the contour length, that is, the number of contour sampling points;

The contour curve of the shape sample is evolved to extract key points. In each evolution process, the point that contributes the least to target recognition is deleted, where the contribution of each point p(i) is defined as:

Among them, h(i, i-1) is the length of the curve between points p(i) and p(i-1), h(i, i+1) is the length of the curve between points p(i) and p(i+1), H ₁ (i) is the angle between line segment p(i)p(i-1) and line segment p(i)p(i+1), and the length h is normalized according to the contour perimeter; the larger the Con(i) value, the greater the contribution of the point p(i) to the shape feature;

This method uses a region-based adaptive end function F(t) to overcome the problem of extracting too many or too few contour key points:

Where _S0 is the area of the original shape, _S1 is the area of the shape after i evolutions, and _n0 is the total number of points on the original shape contour; when the end function value F(t) exceeds the set threshold, the contour key point extraction ends and n ^* contour key points are obtained.

Furthermore, in step 2, the specific method of defining the approximate bias curvature value at each key point and judging the concavity and convexity of the curve at the key point to obtain the candidate segmentation point is:

In order to calculate the approximate bias curvature value of any key point p(i) in the shape sample S, take the contour points p(i-ε) and p(i+ε) before and after p(i), where ε is an empirical value; because:

cosHε(i)∝cur(p(i)),

Where Hε(i) is the angle between the line segment p(i)p(i-ε) and the line segment p(i)p(i+ε), cur(p(i)) is the curvature at point p(i); the approximate biased curvature value cur～(p(i)) at point p(i) is defined as:

cur～(p(i))＝cosHε(i)+1，

Where Hε(i) is the angle between the line segment p(i)p(i-ε) and the line segment p(i)p(i+ε), cosHε(i) ranges from -1 to 1, and cur～(p(i)) ranges from 0 to 2;

According to the shape segmentation method that conforms to visual naturalness, the shape segmentation points are all located at the concave curve of the contour; therefore, when screening the candidate segmentation points for shape segmentation, a method for judging the concavity of the curve at the key point p(i) is defined:

For the binary image of the shape, the values of the pixels inside the contour of the shape sample S are all 255, and the values of the pixels outside the contour of the shape sample S are all 0; the equidistant sampling line segments p(i-ε)p(i+ε) obtain R discrete points. If the pixel values of these R discrete points are all 255, then the line segments p(i-ε)p(i+ε) are all inside the shape contour, that is, the curve at p(i) is convex; if the pixel values of these R discrete points are all 0, then the line segments p(i-ε)p(i+ε) are all outside the shape contour, that is, the curve at p(i) is concave; the key point p(i) where all curves are concave is recorded as the candidate segmentation point P(j).

Furthermore, in step 3, the steps of adjusting the curvature screening threshold Th and obtaining the shape segmentation points are as follows:

(1) For all candidate segmentation points P(j) obtained in step 2, their average approximate bias curvature value is used as the initial threshold Th ₀ :

Where J is the total number of candidate segmentation points;

(2) For the threshold Th _τ during the τth adjustment, according to the relationship between the approximate bias curvature value of each candidate segmentation point P(j) and Th _τ , P(j) can be divided into two categories: candidate segmentation points with approximate bias curvature values greater than Th τ and candidate segmentation points with approximate bias curvature values greater than Th _τ . Candidate segmentation points with approximate bias curvature values less than or equal to Th _τ Calculate and record the segmentation discrimination D _τ under the current threshold:

in,

in They represent the positive and negative curvature deviations of each candidate segmentation point P(j) under the threshold Th _τ , represents the minimum value of the positive curvature deviation of all candidate segmentation points, Represents the maximum value of the negative curvature deviation of all candidate segmentation points;

Determine whether there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ . If not, no adjustment is made and the process goes to step (4). If there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ , the process goes to step (3) and continues to adjust the threshold.

(3) Continue to adjust the threshold. The new threshold Th _τ+1 is the minimum value of the positive curvature deviation of all candidate segmentation points in the last threshold adjustment process, which can be expressed by the following formula:

According to the threshold Th _τ+1, the positive and negative curvature deviations of each candidate segmentation point under the τ+1th adjustment are calculated. and the segmentation discrimination D _τ+1 and record them; determine whether there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ+1. If not, no adjustment is made and go to step (4); if there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ+1 , set τ=τ+1 and repeat the current step to continue adjusting the threshold;

(4) If the threshold is adjusted multiple times, there will be multiple segmentation distinctions. The threshold corresponding to the maximum segmentation distinction is the final curvature screening threshold Th, and the point whose approximate bias curvature value is less than the threshold Th is the final shape segmentation point.

Furthermore, in the step 4, the shape segmentation is performed based on the principle that the segmentation line segments are located within the shape and do not intersect each other, and the specific method of segmenting to obtain a plurality of sub-shape parts with the minimum segmentation cost is:

(1) For any two shape segmentation points P(e ₁ ) and P(e ₂ ), the segmentation line segments P(e ₁ ) and P(e ₂ ) are sampled at equal intervals to obtain C discrete points. If there is a discrete point with a pixel value of 0 among the C points, the segment P(e ₁ ) and P(e ₂ ) exist outside the shape contour and are not selected as segmentation line segments.

(2) For any two shape segmentation points P(e ₃ ) and P(e ₄ ), if there exists a shape segmentation line segment P(e ₅ ) and P(e ₆ ) such that:

or

Then the line segment P(e ₃ )P(e ₄ ) intersects with the existing segmentation line segment P(e ₅ )P(e ₆ ), and the line segment P(e ₃ )P(e ₄ ) is not selected as the segmentation line segment;

(3) The segmentation line set that meets the above two principles is further screened, and segmentation is achieved at the minimum segmentation cost by defining three metrics I to evaluate the quality of the segmentation line segments:

Among them, D ^* (u,v), L ^* (u,v), and S ^* (u,v) are three segmentation metrics, namely, normalized segmentation length, segmentation arc length, and segmentation residual area. u and v are the sequence numbers of any two shape segmentation points. is the total number of split points;

For any shape segmentation line segment P(u)P(v), the three segmentation evaluation indicators are calculated as follows:

Where D _max is the length of the longest segment among all segmentation segments, and the value range of D ^* (u, v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect;

in is the contour curve between points P(u) and P(v) The length of L ^* (u,v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect;

Where _Sd is the area of the shape segmented by the segmentation line segment P(u)P(v), that is, the area of the shape segmented by the segment P(u)P(v) and the contour curve The area of the closed region formed, the value range of S ^* (u,v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect;

According to the above steps, the segmentation cost Cost for the segmentation line segment P(u)P(v) is calculated as:

Cost=αD ^* (u, v) + βL ^* (u, v) + γS ^* (u, v),

Among them, α, β, and γ are the weights of each metric;

Calculate the segmentation cost Cost of the segmentation line segments in the filtered segmentation line segment set; sort all the calculated costs from small to large, and finally select N-1 segmentation line segments with the smallest Cost according to the number N of segmentation sub-shape parts set for the category to which the shape sample S belongs, so as to achieve optimal segmentation and obtain N sub-shape parts; the number N of segmentation sub-shape parts depends on the category to which the current shape sample S belongs. For shapes of different categories, the corresponding number of segmentation sub-shape parts is manually set.

Furthermore, in step 5, the specific method of constructing the topological structure of the shape sample is as follows: for any N sub-shape parts obtained by segmenting the shape sample S, the central shape part is recorded as the starting vertex v ₁ , and the remaining adjacent shape parts are sorted in a clockwise direction and recorded as vertices {v _o |o∈[2,N]}; the edge connecting v ₁ to the remaining vertices v _o is recorded as (v ₁ ,v _o ), thereby forming a shape directed graph that satisfies the topological order:

G ₁ =(V ₁ ,E ₁ ),

Where V ₁ ={v _o |o∈[1, N]}, E ₁ ={(v ₁ , v _o )|o∈[2, N]};

After all training shape samples are optimally segmented, the maximum number of sub-shape parts obtained by segmenting the training shape samples is recorded as For any shape sample S, its adjacency matrix The calculation method is:

in express A real matrix of order,

Furthermore, in step 6, the specific method of using the full-scale visualization representation method of the shape to obtain the color feature expression image of the corresponding sub-shape part is:

For any sub-shape part S ¹ of a shape sample S:

in, are the horizontal and vertical coordinates of the contour sampling point p ¹ (i) of the sub-shape part in the two-dimensional plane, n ¹ is the contour length, that is, the number of contour sampling points;

First, the feature function M composed of three shape descriptors is used to describe the outline of the sub-shape part S1:

M={s _k (i), l _k (i), c _k (i)|k∈[1, m], i∈[1, n ¹ ]},

Among them, s _k , l _k , c _k are the three invariant parameters of normalized area s, arc length l and centroid distance c in scale k, k is the scale label, and m is the total number of scales; these three shape invariant descriptors are defined respectively:

With a contour sampling point p ¹ (i) as the center and an initial radius A preset circle C ₁ (i) is made, which is the initial semi-global scale for calculating the parameters of the corresponding contour points. After the preset circle C ₁ (i) is obtained according to the above steps, the calculation methods of the three shape descriptors under the scale k=1 are as follows:

When calculating the descriptor s ₁ (i), the area of the region Z ₁ (i) in the preset circle C ₁ (i) that is directly connected to the target contour point p ¹ (i) is recorded as Then we have:

Where B(Z ₁ (i), z) is an indicator function, defined as

The ratio of the area of Z ₁ (i) to the area of the preset circle C ₁ (i) is used as the area parameter s ₁ (i) of the target contour point descriptor:

The value range of s ₁ (i) should be between 0 and 1;

When calculating the c ₁ (i) descriptor, the centroid of the region directly connected to the target contour point p ¹ (i) is first calculated. Specifically, the coordinate values of all pixels in the region are averaged. The result is the coordinate value of the centroid of the region, which can be expressed as:

Among them, w ₁ (i) is the centroid of the above area;

Then calculate the distance between the target contour point p ¹ (i) and the center of gravity w ₁ (i) It can be expressed as:

Finally The ratio of the radius of the preset circle C ₁ (i) of the target contour point p ¹ (i) is used as the centroid distance parameter c ₁ (i) of the target contour point descriptor:

The value range of c ₁ (i) should be between 0 and 1;

When calculating the l ₁ (i) descriptor, the length of the arc segment directly connected to the target contour point p ¹ (i) within the preset circle C ₁ (i) is recorded as and will The ratio to the circumference of the preset circle C ₁ (i) is used as the arc length parameter l ₁ (i) of the target contour point descriptor:

Among them, the value range of l ₁ (i) should be between 0 and 1;

According to the above steps, we can calculate the initial radius at scale label k = 1. The characteristic function M ₁ of the sub-shape part S ¹ of the shape sample S at the semi-global scale is:

M ₁ ={s ₁ (i), l ₁ (i), c ₁ (i)|i∈[1, n ¹ ]},

Since a digital image uses a pixel as the smallest unit, a single pixel is selected as the continuous scale change interval in the full scale space; that is, for the kth scale label, the radius r _k of the circle C _k is set as:

That is, when the initial scale k = 1, After that, the radius r _k is reduced m-1 times in units of one pixel until the minimum scale k=m is reached; the characteristic functions at other scales are calculated in the same way as the characteristic function M ₁ at scale k=1, and finally the characteristic functions of the sub-shape part S ¹ of the shape sample S at all scales are obtained:

M={s _k (i), l _k (i), c _k (i)|k∈[1, m], i∈[1, n ¹ ]},

The characteristic functions at each scale are stored in matrices S ^M , L ^M , and C ^M respectively. S ^M is used to store _sk (i). The k-th row and i-th column of S ^M stores the area parameter _sk (i) of point p ¹ (i) at the k-th scale. L ^M is used to store l _k (i). The k-th row and i-th column of L ^M stores the arc length parameter l _k (i) of point p ¹ (i) at the k-th scale. C ^M is used to store c _k (i). The k-th row and i-th column of C ^M stores the centroid distance parameter c _k (i) of point p ¹ (i) at the k-th scale. S ^M , L ^M , and C ^M are finally used as grayscale images of the three shape features of the sub-shape part S ¹ of the shape sample S in the full-scale space:

GM ¹ ={S ^M , L ^M , C ^M },

Among them, S ^M , L ^M , and C ^M are matrices of size m×n, each representing a grayscale image;

Then, the three grayscale images of the sub-shape part ^S1 are used as the three RGB channels to obtain a color image as the feature expression image of the sub-shape part ^S1.

Furthermore, in step seven, the feature expression image samples of each sub-shape part of all training shape samples are input into the convolutional neural network to train the convolutional neural network model; different sub-shape parts of each type of shape have different category labels; after the convolutional neural network is trained to convergence, for any shape sample S, the feature expression images {T ^num | num∈[1, N]} corresponding to the N sub-shape parts formed by segmentation are respectively input into the trained convolutional neural network, and the output of the second fully connected layer of the network is the feature vector of the corresponding sub-shape part Where Vec is the number of neurons in the second fully connected layer;

The structure of the convolutional neural network includes an input layer, a pre-training layer and a fully connected layer; the pre-training layer is composed of the first four modules of the VGG16 network model, and the parameters obtained after training the four modules in the imagenet dataset are used as initialization parameters, and three fully connected layers are connected after the pre-training layer;

The first module in the pre-training layer specifically includes 2 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 64, the size is 3×3, and the size of the pooling layer is 2×2; the second module specifically includes 2 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 128, the size is 3×3, and the size of the pooling layer is 2×2; the third module specifically includes 3 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 256, the size is 3×3, and the size of the pooling layer is 2×2; the fourth module specifically includes 3 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 512, the size is 3×3, and the size of the pooling layer is 2×2; the calculation formula for each convolutional layer is:

C _O =φ _relu (W _C ·C _I +θ _C ),

Among them, θ _C is the bias vector of the convolutional layer; W _C is the weight of the convolutional layer; C _I is the input of the convolutional layer; C _O is the output of the convolutional layer;

The fully connected layer module specifically includes 3 fully connected layers, where the first fully connected layer contains 512 nodes, the second fully connected layer contains Vec nodes, and the third fully connected layer contains _NT nodes; _NT is the sum of the number of segmented sub-shape parts of all categories of shapes; the calculation formula for the first two fully connected layers is:

F _O =φ _tanh (W _F ·F _I +θ _F ),

Among them, φ _tanh is the tanh activation function, θ _F is the bias vector of the fully connected layer; W _F is the weight of the fully connected layer; _FI is the input of the fully connected layer; _FO is the output of the fully connected layer;

The last fully connected layer is the output layer, and its output calculation formula is:

Y _O =φ _softmax (W _Y ·Y _I +θ _Y ),

Among them, φ _softmax is the softmax activation function, θ _Y is the bias vector of the output layer, each neuron in the output layer represents a corresponding sub-shape part category, W _Y is the weight of the output layer, Y _I is the input of the output layer; Y _O is the output of the output layer.

Furthermore, the specific method of constructing the feature matrix of the shape sample in step eight is:

For any shape sample S, the N sub-shape parts formed by segmentation have the corresponding feature matrix expression The calculation formula is:

Wherein, _Fa represents the a-th row vector of the matrix F, and ^fa is the eigenvector of the a-th sub-shape part outputted in step 7. Represents a zero vector of dimension size Vec.

Furthermore, in step nine, the structure of the graph convolutional neural network is constructed, including a preprocessing input layer, a hidden layer and a classification output layer, and the adjacency matrix is performed in the preprocessing input layer. Normalization preprocessing, specifically:

in I _N is the identity matrix, is the degree matrix, After normalization preprocessing

The hidden layer includes two graph convolution layers, and the calculation formula of each graph convolution layer is:

in, is the weight of the graph convolution layer; H _I is the input of the graph convolution layer, and the input of the first convolution layer is the feature matrix of the shape sample H _O is the output of the graph convolutional layer;

The calculation formula of the classification output layer is:

Among them, φ _softmax is the softmax activation function, G _I is the input of the output layer, that is, the output of the second graph convolutional layer, G _W is the weight of the output layer; G _O is the output of the output layer; each neuron in the output layer represents a corresponding shape category.

Furthermore, the specific method for implementing contour shape classification and recognition in step 10 is: training the graph convolutional neural network model until convergence; for any test shape sample, first extract the key points of the shape contour, calculate the curvature value at each key point, judge its concavity and convexity, obtain the candidate segmentation points, and then adjust the curvature screening threshold to obtain the shape segmentation points; according to the two principles (1) and (2) in step 5, obtain the segmentation line segment set, and calculate the segmentation cost of the segmentation line segments in the segmentation line segment set; if the number of segmentation line segments is less than Then all segmentation line segments are used to segment the shape; otherwise, according to the minimum segmentation cost The shape is segmented by segmentation line segments; the color feature expression image of each sub-shape part is calculated and input into the trained convolutional neural network, and the output of the second fully connected layer of the convolutional neural network is used as the feature vector of the sub-shape part; the shape directed graph of the test shape sample is constructed, and its adjacency matrix and feature matrix are calculated and input into the trained graph convolutional neural network model. The shape category corresponding to the maximum value in the output vector is judged as the shape type of the test sample, thereby realizing shape classification and recognition.

The present invention proposes a new shape recognition method, and designs a new shape classification method using graph convolutional neural network; the proposed shape feature topological graph expression is a directed graph structure constructed based on graph segmentation, which not only distinguishes the shape hierarchy, but also makes full use of the stable topological feature relationship between the parts of each level of the shape to replace the geometric position relationship. Compared with the method in the background technology that only calculates and compares the corresponding salient point features for matching, the present invention can be more robustly applied to interference such as shape articulation transformation, partial occlusion, and rigid body transformation; the full-scale visualization representation method of the shape used can fully express all the information of each sub-shape part, and then use the continuous convolution calculation of the neural network to extract the features of each part in the full-scale space; compared with the direct application of the convolutional neural network, the designed graph convolutional neural network has greatly reduced training parameters and higher calculation efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flowchart of a shape recognition method of the present invention.

FIG. 2 is a schematic diagram of some samples of target shapes in the shape sample set.

FIG3 is a schematic diagram of segmentation of shape samples.

FIG4 is a schematic diagram of the full-scale space.

FIG. 5 is a schematic diagram of a target shape after being cut off at a preset scale.

FIG. 6 is a schematic diagram of a target shape segmented at a preset scale.

FIG. 7 is a schematic diagram of a characteristic function of a sub-shape portion of a target shape at a single scale.

FIG8 is a schematic diagram of a feature matrix of a sub-shape portion of a target shape in the full-scale space.

FIG. 9 is a schematic diagram of three grayscale images and a synthesized color image calculated from a sub-shape portion of a target shape.

FIG10 is a diagram showing the structure of a convolutional neural network used to train images expressing the features of each sub-shape portion.

FIG. 11 is a diagram showing the characteristic structure of each sub-shape portion of the target shape.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

As shown in FIG1 , a shape recognition method includes the following process:

1. The total number of samples in the shape sample set is 1400, with a total of 70 shape categories, and each shape category has 20 shape samples. Figure 2 shows a schematic diagram of some samples of the target shape in the shape sample set. Half of the samples in each shape category are randomly selected to be included in the training set, and the remaining half are included in the test set, resulting in a total of 700 training samples and 700 test samples. 100 contour points are sampled for each shape sample. Take a shape sample S as an example:

S＝{p _x (i), p _y (i)|i∈[1, 100]},

Among them, p _x (i), _py (i) are the horizontal and vertical coordinates of the contour sampling point p(i) in the two-dimensional plane.

The contour curve of the shape sample is evolved to extract key points. In each evolution process, the point that contributes the least to target recognition is deleted. The contribution of each point p(i) is defined as:

Among them, h(i, i-1) is the length of the curve between points p(i) and p(i-1), h(i, i+1) is the length of the curve between points p(i) and p(i+1), H ₁ (i) is the angle between line segment p(i)p(i-1) and line segment p(i)p(i+1), and the length h is normalized according to the contour perimeter. The larger the Con(i) value, the greater the contribution of the point p(i) to the shape feature.

This method uses a region-based adaptive end function F(t) to overcome the problem of extracting too many or too few contour key points:

Where _S0 is the area of the original shape, _S1 is the area after i evolutions, and _n0 is the total number of points on the original shape contour. When the end function value F(t) exceeds the set threshold, the contour key point extraction ends. For the shape sample S shown in Figure 3, a total of 24 contour key points are extracted.

2. Calculate the approximate bias curvature value and curve concavity at each key point of the shape sample. Taking the shape sample S as an example, the calculation formula of the approximate bias curvature value cur ^～ (p(i)) of its contour key point p(i) is as follows:

cur ^~ (p(i))=cosH _ε (i)+1,

Where H _ε (i) is the angle between the line segment p(i)p(i-ε) and the line segment p(i)p(i+ε), ε=3.

The method for judging the concavity of the curve at the contour key point p(i) is as follows:

The line segments p(i-ε)p(i+ε) are sampled equidistantly to obtain R discrete points. If the pixel values of these R discrete points are all 255, then the line segments p(i-ε)p(i+ε) are all inside the shape contour, that is, the curve at p(i) is convex; if the pixel values of these R discrete points are all 0, then the line segments p(i-ε)p(i+ε) are all outside the shape contour, that is, the curve at p(i) is concave. The key point p(i) where all curves are concave is recorded as the candidate segmentation point P(j). For the shape sample S, a total of 11 candidate segmentation points are extracted.

3. Adjust the curvature screening threshold Th and obtain the shape segmentation point. For the 11 candidate segmentation points P(j) of the shape sample S, their average approximate bias curvature value is used as the initial threshold Th ₀ :

The average approximate bias curvature values cur ^~ (P(j)) of the 11 candidate segmentation points P(j) are 0.1, 0.2, 0.25, 0.35, 0.4, 0.5, 0.5, 0.64, 0.7, 0.7, and 0.8 respectively.

Increase the threshold in the following order:

(1) For the threshold Th _τ during the τth adjustment, according to the relationship between the approximate bias curvature value of each candidate segmentation point P(j) and Th _τ , P(j) can be divided into two categories: candidate segmentation points with approximate bias curvature values greater than Th _{τ and} Candidate segmentation points with approximate bias curvature values less than or equal to Th _τ Calculate and record the segmentation discrimination D _τ under the current threshold:

in,

in They represent the positive and negative curvature deviations of each candidate segmentation point P(j) under the threshold Th _τ , represents the minimum value of the positive curvature deviation of all candidate segmentation points, Represents the maximum value of the negative curvature deviation of all candidate segmentation points.

Determine whether there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ . If not, no adjustment is made and the process goes to step (3). If there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ , the process goes to step (2) and continues to adjust the threshold.

(2) Continue to adjust the threshold. The new threshold Th _τ+1 is the minimum value of the positive curvature deviation of all candidate segmentation points in the last threshold adjustment process, which can be expressed by the following formula:

According to the threshold Th _τ+1, the positive and negative curvature deviations of each candidate segmentation point under the τ+1th adjustment are calculated. And the segmentation discrimination D _τ+1 is recorded. Determine whether there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ+1. If not, no adjustment is made and go to step (3). If there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ+1 , set τ=τ+1 and repeat the current step to continue adjusting the threshold.

(3) If the threshold is adjusted multiple times, there will be multiple segmentation distinctions. The threshold corresponding to the maximum segmentation distinction is the final curvature screening threshold Th, and the point whose approximate bias curvature value is greater than the threshold Th is the final shape segmentation point.

For the shape sample S, the segmentation discrimination and threshold values recorded in the four threshold adjustment processes are:

Therefore, the threshold Th1 corresponding to the maximum segmentation discrimination _D1 is the final curvature screening threshold, that is, Th = 0.5. The five candidate segmentation points whose approximate bias curvature values are less than _Th are the final shape segmentation points, and the corresponding approximate bias curvatures are 0.1, 0.2, 0.25, 0.35, and 0.4 respectively.

4. For the shape sample S, connect the five shape segmentation points in pairs to form 10 line segments, and retain the seven line segments that are located within the shape and do not intersect each other as the segmentation line segment set. Use the metric I to calculate the segmentation cost of each segmentation line segment according to the following method:

Where D ^* (u,v), L ^* (u,v), S ^* (u,v) are the three segmentation metrics of normalized segmentation length, segmentation arc length, and segmentation residual area, respectively. u and v are the serial numbers of any two shape segmentation points. is the total number of split points.

For any shape segmentation line segment P(u)P(v), the three segmentation evaluation indicators are calculated as follows:

D _max is the length of the longest segment among all segmentation segments, and the value range of D ^* (u, v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect.

in is the contour curve between points P(u) and P(v) The length of L ^* (u, v) should range from 0 to 1, and the smaller the value, the more significant the segmentation effect.

Where _Sd is the area of the shape segmented by the segmentation line segment P(u)P(v), that is, the area of the shape segmented by the segment P(u)P(v) and the contour curve The area of the closed region, S ^* (u, v) should range from 0 to 1, and the smaller the value, the more significant the segmentation effect.

According to the above steps, the segmentation cost Cost for the segmentation line segment P(u)P(v) is calculated as:

Cost=αD ^* (u, v) + βL ^* (u, v) + γS ^* (u, v),

Among them, α, β, and γ are the weights of each metric.

As shown in FIG3 , for the shape sample S, two segmentation line segments with the minimum and second minimum segmentation costs are selected as the final optimal segmentation line segments, and three sub-shape parts are obtained.

5. For the shape sample S, record the central shape part as the starting vertex v ₁ , and sort the other two adjacent shape parts in a clockwise direction, and record them as vertices {v ₂ , v ₃ }. The edges connecting v ₁ to vertices v ₂ and v ₃ are recorded as (v ₁ , v ₂ ) and (v ₁ , v ₃ ), respectively, thus forming a shape directed graph that satisfies the topological order:

G ₁ =(V ₁ ,E ₁ ),

Where V ₁ ={v ₁ , v ₂ , v ₃ }, E ₁ ={(v ₁ , v ₂ ), (v ₁ , v ₃ )}.

Since the maximum number of sub-shape parts obtained by segmenting each training sample in the contour shape set is 11, the adjacency matrix of the shape sample S is It is expressed as:

Where a∈[1，11], b∈[1，11].

6. Perform full-scale visualization on the three sub-shape parts obtained by segmentation, where the specific method of full-scale visualization is as follows:

(1) For the contour of any sub-shape part, sample the contour to obtain 100 contour sampling points. As shown in Figure 4, the total number of scales in the full-scale space is set to 100, and the normalized area, arc length and centroid distance of each contour point corresponding to each layer of scale are calculated based on the coordinates of the 100 contour sampling points. Taking the sub-shape part ^S1 as an example, the specific calculation method is as follows:

The sampling point p ¹ (i) of the sub-shape part S ¹ is taken as the center of the circle and the initial radius is Make a preset circle C ₁ (i), which is the initial semi-global scale for calculating the corresponding contour point parameters. After obtaining the preset circle C ₁ (i) according to the above steps, a part of the target shape must fall within the preset circle, as shown in Figure 5. If the part of the target shape that falls within the preset circle is a separate area, then the separate area is the area that has a direct connection with the target contour point p ¹ (i), denoted as Z ₁ (i); if the part of the target shape that falls within the preset circle is divided into several unconnected areas, such as area A and area B as shown in Figure 5, then the area on the contour of the target contour point p ¹ (i) is determined to be the area that has a direct connection with the target contour point p ¹ (i), which is area A in Figure 5, denoted as Z ₁ (i). Based on this, the area of area Z ₁ (i) in the preset circle C ₁ (i) that has a direct connection with the target contour point p ¹ (i) is denoted as Then we have:

Where B(Z ₁ (i), z) is an indicator function, defined as

The ratio of the area of Z ₁ (i) to the area of the preset circle C ₁ (i) is used as the area parameter s ₁ (i) of the descriptor of the target contour point p ¹ (i):

The value range of s ₁ (i) should be between 0 and 1.

When calculating the centroid of the region directly connected to the target contour point p ¹ (i), the coordinate values of all the pixels in the region are averaged, and the result is the coordinate value of the centroid of the region, which can be expressed as:

Among them, w ₁ (i) is the centroid of the above area.

The distance between the target contour point and the center of gravity w ₁ (i) is calculated It can be expressed as:

and will The ratio of the radius of the preset circle of the target contour point is used as the centroid distance parameter c ₁ (i) of the descriptor of the target contour point p ¹ (i):

The value range of c ₁ (i) should be between 0 and 1.

After the contour of the target shape is cut by the preset circle, there will inevitably be one or more arc segments within the preset circle, as shown in Figure 6. If the target shape has only one arc segment within the preset circle, then the arc segment is determined to be an arc segment that has a direct connection relationship with the target contour point p ¹ (i). If the target shape has multiple arc segments within the preset circle, such as arc segment A (Segment A), arc segment B (Segment B), and arc segment C (Segment C) in Figure 6, then the arc segment where the target contour point p ¹ (i) is located is determined to be an arc segment that has a direct connection relationship with the target contour point p ¹ (i), which is arc segment A (Segment A) in Figure 6. Based on this, the length of the arc segment that has a direct connection relationship with the target contour point p ¹ (i) within the preset circle C ₁ (i) is recorded as and will The ratio of the circumference of the preset circle C ₁ (i) is used as the arc length parameter l ₁ (i) of the descriptor of the target contour point:

Among them, the value range of l ₁ (i) should be between 0 and 1.

According to the above steps, we can calculate the initial radius at scale label k = 1. The characteristic function M ₁ of the sub-shape part S ¹ of the shape sample S at the semi-global scale is:

M ₁ ={s ₁ (i), l ₁ (i), c ₁ (i)|i∈[1, 100]},

As shown in FIG7 , the characteristic functions of each of the 100 scales in the full-scale space are calculated respectively, where for the k-th scale label, the radius r _k of the circle C _k is set as:

That is, when the initial scale k = 1, After that, the radius r _k is reduced by 99 times in units of one pixel until the minimum scale k = 100. The characteristic function of the sub-shape part S ¹ of the shape sample S in all scale spaces is calculated:

M={s _k (i), l _k (i), c _k (i)|k∈[1, 100], i∈[1, 100]},

(2) As shown in FIG8 , the feature functions of the 100 scales in the full-scale space of the sub-shape part S ¹ are merged into three feature matrices in the full-scale space in scale order:

GM ¹ ={S ^M , L ^M , C ^M },

Among them, S ^M , L ^M , and ^CM are all grayscale matrices of size m×n, each representing a grayscale image.

(3) As shown in FIG. 9 , the three grayscale images of the sub-shape portion S ¹ are used as the three RGB channels to synthesize a color image as the feature expression image of the sub-shape portion S ¹

7. Construct a convolutional neural network, including an input layer, a pre-training layer, and a fully connected layer. The present invention inputs the feature expression image samples of each sub-shape part of all training shape samples into the convolutional neural network to train the convolutional neural network model. Different sub-shape parts of each type of shape have different category labels. After the convolutional neural network is trained to convergence, taking the shape sample S as an example, the feature expression images {T ^num | num∈[1,3]} of size 100×100 corresponding to the three sub-shape parts formed by segmentation are respectively input into the trained convolutional neural network, and the output of the second fully connected layer of the network is the feature vector of the corresponding sub-shape part. Where Vec is the number of neurons in the second fully connected layer, and Vec is set to 200.

The present invention uses the sgd optimizer, the learning rate is set to 0.001, the delay rate is set to 1e-6, the cross entropy is selected as the loss function, and the batch size is selected to be 128. As shown in Figure 10, the pre-training layer consists of the first 4 modules of the VGG16 network model, and the parameters obtained after training the 4 modules in the imagenet dataset are used as initialization parameters. Three fully connected layers are connected after the pre-training layer.

The first module in the pre-training layer specifically includes 2 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 64, the size is 3×3, and the size of the pooling layer is 2×2; the second module specifically includes 2 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 128, the size is 3×3, and the size of the pooling layer is 2×2; the third module specifically includes 3 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 256, the size is 3×3, and the size of the pooling layer is 2×2; the fourth module specifically includes 3 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 512, the size is 3×3, and the size of the pooling layer is 2×2. The calculation formula for each convolutional layer is:

C _O =φ _relu (W _C ·C _I +θ _C ),

Among them, θ _C is the bias vector of the convolutional layer; W _C is the weight of the convolutional layer; C _I is the input of the convolutional layer; C _o is the output of the convolutional layer;

The fully connected layer module specifically includes 3 fully connected layers, of which the first fully connected layer contains 512 nodes, the second fully connected layer contains 200 nodes, and the third fully connected layer contains 770 nodes. The calculation formula for the first two fully connected layers is:

F _O =φ _tanh (W _F ·F _I +θ _F ),

Among them, φ _tan h is the tanh activation function, θ _F is the bias vector of the fully connected layer; W _F is the weight of the fully connected layer; _FI is the input of the fully connected layer; _FO is the output of the fully connected layer;

The last fully connected layer is the output layer, and its output calculation formula is:

Y _O =φ _softmax (W _Y ·Y _I +θ _Y ),

Among them, φ _softmax is the softmax activation function, θ _Y is the bias vector of the output layer, each neuron in the output layer represents a corresponding sub-shape part category, W _Y is the weight of the output layer, Y _I is the input of the output layer; Y _O is the output of the output layer;

8. As shown in Figure 11, the feature matrix of the shape sample is constructed based on the three sub-shape feature vectors of the shape sample

Wherein, _Fa represents the a-th row vector of the matrix F, and ^fa is the eigenvector of the a-th sub-shape part outputted in the above step. Represents a zero vector of dimension size 200.

9. Construct a graph convolutional neural network, including a preprocessing input layer, a hidden layer, and a classification output layer. The present invention converts the adjacency matrix of the shape sample topology graph into and the feature matrix The input graph convolutional neural network structure model is trained. The present invention uses the sgd optimizer, the learning rate is set to 0.001, the delay rate is set to 1e-6, the loss function uses cross entropy, and the batch size is selected as 128.

Preprocess the adjacency matrix in the input layer Normalization preprocessing, specifically:

in is the identity matrix, is the degree matrix, After normalization preprocessing

The hidden layer includes two graph convolution layers, and the calculation formula of each graph convolution layer is:

in, is the weight of the graph convolution layer; H _I is the input of the graph convolution layer, and the input of the first convolution layer is the feature matrix of the shape sample H _O is the output of the graph convolutional layer;

The calculation formula of the classification output layer is:

Among them, φ _softmax is the softmax activation function, G _I is the input of the output layer, that is, the output of the second graph convolutional layer, G _W is the weight of the output layer; G _O is the output of the output layer; each neuron in the output layer represents a corresponding shape category.

10. Input all training samples into the graph convolutional neural network to train the graph convolutional neural network model. For any test shape sample, first extract the key points of the shape contour, calculate the curvature value at each key point, determine its convexity, obtain the candidate segmentation points, and then adjust the curvature screening threshold to obtain the shape segmentation points. Connect the shape segmentation points in pairs to form segmentation segments, retain the segments that are located in the shape and do not cross each other as segmentation segments, obtain a segmentation segment set, and calculate the segmentation cost of the segmentation segments in the segmentation segment set. If the number of segmentation segments is less than 10, all segmentation segments are used to segment the shape. Otherwise, the shape is segmented according to the 10 segmentation segments with the smallest segmentation cost. Calculate the color feature expression image of each sub-shape part and input it into the trained convolutional neural network. The output of the second fully connected layer of the convolutional neural network is used as the feature vector of the sub-shape part. Construct a shape directed graph of the test shape sample, calculate its adjacency matrix and feature matrix, and input them into the trained graph convolutional neural network model. The shape category corresponding to the maximum value in the output vector is judged as the shape type of the test sample to achieve shape classification and recognition.

Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims (11)

1. A shape recognition method, characterized in that the method comprises the following steps: Step 1: Extract the contour key points of the shape sample; Step 2: define the approximate bias curvature value at each key point and determine the concavity and convexity of the curve at the key point to obtain candidate shape segmentation points; Step 3: Adjust the curvature screening threshold to obtain the shape segmentation point; Step 4: segment the shape based on the principle that the segmentation line segments are located within the shape and do not cross each other, and obtain several sub-shape parts with the minimum segmentation cost; Step 5: Construct the topological structure of the shape sample; Step 6: Use the full-scale visualization method of the shape to obtain a feature expression image of the corresponding sub-shape part; Step 7: Input each feature expression image into the convolutional neural network for training, and learn to obtain the feature vector of each sub-shape part; Step 8: construct a feature matrix of shape samples; Step 9: Construct a graph convolutional neural network; Step 10: Train the graph convolutional neural network, perform shape segmentation on the test sample, obtain the feature vector of each sub-shape part, calculate the feature matrix and adjacency matrix of the test sample, and input them into the trained graph convolutional network model to realize shape classification and recognition. 2. A shape recognition method according to claim 1, characterized in that: in the step 1, the method for extracting contour key points is: The contour of each shape sample is composed of a series of sampling points. For any shape sample S, sampling n points of the contour yields: S={(p _x (i), p _y (i))|i∈[1, n]}, Where p _x (i), p _y (i) are the horizontal and vertical coordinates of the contour sampling point p (i) in the two-dimensional plane, and n is the contour length, that is, the number of contour sampling points; The contour curve of the shape sample is evolved to extract key points. In each evolution process, the point that contributes the least to target recognition is deleted, where the contribution of each point p(i) is defined as: Among them, h(i, i-1) is the length of the curve between points p(i) and p(i-1), h(i, i+1) is the length of the curve between points p(i) and p(i+1), H ₁ (i) is the angle between line segment p(i)p(i-1) and line segment p(i)p(i+1), and the length h is normalized according to the contour perimeter; the larger the Con(i) value, the greater the contribution of the point p(i) to the shape feature; This method uses a region-based adaptive end function F(t) to overcome the problem of extracting too many or too few contour key points: Where _S0 is the area of the original shape, _S1 is the area of the shape after i evolutions, and _n0 is the total number of points on the original shape contour; when the end function value F(t) exceeds the set threshold, the contour key point extraction ends and n ^* contour key points are obtained. 3. A shape recognition method according to claim 2, characterized in that: in the step 2, the specific method of defining the approximate bias curvature value at each key point and judging the concavity and convexity of the curve at the key point to obtain the candidate segmentation point is: In order to calculate the approximate bias curvature value of any key point p(i) in the shape sample S, take the contour points p(i-ε) and p(i+ε) before and after p(i), where ε is an empirical value; due to: cosH _ε (i)∝cur(p(i)), Where H _ε (i) is the angle between the line segment p(i)p(i-ε) and the line segment p(i)p(i+ε), and cur(p(i)) is the curvature at point p(i); The approximate bias curvature value cur~(p(i)) at the definition point p(i) is: cur~(p(i))=cosH _ε (i)+1, Where H _ε (i) is the angle between the line segment p(i)p(i-ε) and the line segment p(i)p(i+ε), cosH _ε (i) ranges from -1 to 1, and cur～(p(i)) ranges from 0 to 2; According to the shape segmentation method that conforms to visual naturalness, the shape segmentation points are all located at the concave curve of the contour; therefore, when screening the candidate segmentation points for shape segmentation, a method for judging the concavity of the curve at the key point p(i) is defined: For the binary image of the shape, the values of the pixels inside the contour of the shape sample S are all 255, and the values of the pixels outside the contour of the shape sample S are all 0; the equidistant sampling line segments p(i-ε)p(i+ε) obtain R discrete points. If the pixel values of these R discrete points are all 255, then the line segments p(i-ε)p(i+ε) are all inside the shape contour, that is, the curve at p(i) is convex; if the pixel values of these R discrete points are all 0, then the line segments p(i-ε)p(i+ε) are all outside the shape contour, that is, the curve at p(i) is concave; the key point p(i) where all curves are concave is recorded as the candidate segmentation point P(j). 4. A shape recognition method according to claim 3, characterized in that: in the step 3, the step of adjusting the curvature screening threshold Th and obtaining the shape segmentation point is as follows: (1) For all candidate segmentation points P(j) obtained in step 2, their average approximate bias curvature value is used as the initial threshold Th ₀ : Where J is the total number of candidate segmentation points; (2) For the threshold Th _τ during the τth adjustment, according to the relationship between the approximate bias curvature value of each candidate segmentation point P(j) and Th _τ , P(j) can be divided into two categories: candidate segmentation points with approximate bias curvature values greater than Th τ and candidate segmentation points with approximate bias curvature values greater than Th _τ . Candidate segmentation points with approximate bias curvature values less than or equal to Th _τ Calculate and record the segmentation discrimination D _τ under the current threshold: in, in They represent the positive and negative curvature deviations of each candidate segmentation point P(j) under the threshold Th _τ , represents the minimum value of the positive curvature deviation of all candidate segmentation points, Represents the maximum value of the negative curvature deviation of all candidate segmentation points; Determine whether there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ . If not, no adjustment is made and the process goes to step (4). If there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ , the process goes to step (3) and continues to adjust the threshold. (3) Continue to adjust the threshold. The new threshold Th _τ+1 is the minimum value of the positive curvature deviation of all candidate segmentation points in the last threshold adjustment process, which can be expressed by the following formula: According to the threshold Th _τ+1, the positive and negative curvature deviations of each candidate segmentation point under the τ+1th adjustment are calculated. and the segmentation discrimination D _τ+1 and record them; determine whether there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ+1. If not, no adjustment is made and go to step (4); if there is a candidate segmentation point whose approximate bias curvature value is greater than the threshold Th _τ+1 , set τ=τ+1 and repeat the current step to continue adjusting the threshold; (4) If the threshold is adjusted multiple times, there will be multiple segmentation distinctions. The threshold corresponding to the maximum segmentation distinction is the final curvature screening threshold Th, and the point whose approximate bias curvature value is less than the threshold Th is the final shape segmentation point. 5. A shape recognition method according to claim 4, characterized in that: in said step 4, the shape segmentation is performed based on the principle that the segmentation line segments are located within the shape and do not intersect each other, and the specific method of segmenting to obtain a plurality of sub-shape parts with the minimum segmentation cost is: (1) For any two shape segmentation points P(e ₁ ) and P(e ₂ ), the segmentation line segments P(e ₁ ) and P(e ₂ ) are sampled at equal intervals to obtain C discrete points. If there is a discrete point with a pixel value of 0 among the C points, the segment P(e ₁ ) and P(e ₂ ) exist outside the shape contour and are not selected as segmentation line segments. (2) For any two shape segmentation points P(e ₃ ) and P(e ₄ ), if there exists a shape segmentation line segment P(e ₅ ) and P(e ₆ ) such that: or Then the line segment P(e ₃ )P(e ₄ ) intersects with the existing segmentation line segment P(e ₅ )P(e ₆ ), and the line segment P(e ₃ )P(e ₄ ) is not selected as the segmentation line segment; (3) The segmentation line set that meets the above two principles is further screened, and segmentation is achieved at the minimum segmentation cost by defining three metrics I to evaluate the quality of the segmentation line segments: Among them, D ^* (u,v), L ^* (u,v), and S ^* (u,v) are three segmentation metrics, namely, normalized segmentation length, segmentation arc length, and segmentation residual area. u and v are the sequence numbers of any two shape segmentation points. is the total number of split points; For any shape segmentation line segment P(u)P(v), the three segmentation evaluation indicators are calculated as follows: Where D _max is the length of the longest segment among all segmentation segments, and the value range of D ^* (u, v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect; in is the contour curve between points P(u) and P(v) The length of L ^* (u,v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect; Where _Sd is the area of the shape segmented by the segmentation line segment P(u)P(v), that is, the area of the shape segmented by the segment P(u)P(v) and the contour curve The area of the closed region formed, the value range of S ^* (u,v) should be between 0 and 1, and the smaller the value, the more significant the segmentation effect; According to the above steps, the segmentation cost Cost for the segmentation line segment P(u)P(v) is calculated as follows: Cost＝αD ^* (u, v) + βL ^* (u, v) + γS ^* (u, v), Among them, α, β, and γ are the weights of each metric; Calculate the segmentation cost Cost of the segmentation line segments in the filtered segmentation line segment set; sort all the calculated costs from small to large, and finally select N-1 segmentation line segments with the smallest Cost according to the number N of segmentation sub-shape parts set for the category to which the shape sample S belongs, so as to achieve optimal segmentation and obtain N sub-shape parts; the number N of segmentation sub-shape parts depends on the category to which the current shape sample S belongs. For shapes of different categories, the corresponding number of segmentation sub-shape parts is manually set. 6. A shape recognition method according to claim 5, characterized in that: in said step 5, the specific method of constructing the topological structure of the shape sample is: for any N sub-shape parts obtained by segmenting the shape sample S, the central shape part is recorded as the starting vertex v ₁ , and the remaining adjacent shape parts are sorted in a clockwise direction and recorded as vertices {v _o |o∈[2,N]}; the edges connecting v ₁ to the remaining vertices v _o are recorded as (v ₁ ,v _o ), thereby forming a shape directed graph that satisfies the topological order: G ₁ =(V ₁ ,E ₁ ), Where V ₁ ={v _o |o∈[1, N]}, E ₁ ={(v ₁ , v _o )|o∈[2, N]}; After all training shape samples are optimally segmented, the maximum number of sub-shape parts obtained by segmenting the training shape samples is recorded as For any shape sample S, its adjacency matrix The calculation method is: in express A real matrix of order, 7. A shape recognition method according to claim 6, characterized in that: in the step 6, the specific method of using the full-scale visual representation method of the shape to obtain the color feature expression image of the corresponding sub-shape part is: For any sub-shape part S ¹ of a shape sample S: in, are the horizontal and vertical coordinates of the contour sampling point p ¹ (i) of the sub-shape part in the two-dimensional plane, n ¹ is the contour length, that is, the number of contour sampling points; First, the feature function M composed of three shape descriptors is used to describe the outline of the sub-shape part ^S1 : M={s _k (i), l _k (i), c _k (i)|k∈[1, m], i∈[1, n ¹ ]}, Among them, s _k , l _k , c _k are the three invariant parameters of normalized area s, arc length l and centroid distance c in scale k, k is the scale label, and m is the total number of scales; these three shape invariant descriptors are defined respectively: With a contour sampling point p ¹ (i) as the center and an initial radius A preset circle C ₁ (i) is made, which is the initial semi-global scale for calculating the parameters of the corresponding contour points. After the preset circle C ₁ (i) is obtained according to the above steps, the calculation methods of the three shape descriptors under the scale k=1 are as follows: When calculating the descriptor s ₁ (i), the area of the region Z ₁ (i) in the preset circle C ₁ (i) that is directly connected to the target contour point p ¹ (i) is recorded as Then we have: Where B(Z ₁ (i), z) is an indicator function, defined as The ratio of the area of Z ₁ (i) to the area of the preset circle C ₁ (i) is used as the area parameter s ₁ (i) of the target contour point descriptor: The value range of s ₁ (i) should be between 0 and 1; When calculating the c ₁ (i) descriptor, the centroid of the region directly connected to the target contour point p ¹ (i) is first calculated. Specifically, the coordinate values of all pixels in the region are averaged. The result is the coordinate value of the centroid of the region, which can be expressed as: Among them, w ₁ (i) is the centroid of the above area; Then calculate the distance between the target contour point p ¹ (i) and the center of gravity w ₁ (i) It can be expressed as: Finally The ratio of the radius of the preset circle C ₁ (i) of the target contour point p ¹ (i) is used as the centroid distance parameter c ₁ (i) of the target contour point descriptor: The value range of c ₁ (i) should be between 0 and 1; When calculating the l ₁ (i) descriptor, the length of the arc segment directly connected to the target contour point p ¹ (i) within the preset circle C ₁ (i) is recorded as and will The ratio to the circumference of the preset circle C ₁ (i) is used as the arc length parameter l ₁ (i) of the target contour point descriptor: Among them, the value range of l ₁ (i) should be between 0 and 1; According to the above steps, we can calculate the initial radius at scale label k = 1. The characteristic function M ₁ of the sub-shape part S ¹ of the shape sample S at the semi-global scale is: M ₁ ={s ₁ (i), l ₁ (i), c ₁ (i)|i∈[1, n ¹ ]}, Since a digital image uses a pixel as the smallest unit, a single pixel is selected as the continuous scale change interval in the full scale space; that is, for the kth scale label, the radius r _k of the circle C _k is set as: That is, when the initial scale k = 1, After that, the radius r _k is reduced m-1 times in units of one pixel until the minimum scale k=m is reached; the characteristic functions at other scales are calculated in the same way as the characteristic function M ₁ at scale k=1, and finally the characteristic functions of the sub-shape part S ¹ of the shape sample S at all scales are obtained: M={s _k (i), l _k (i), c _k (i)|k∈[1, m], i∈[1, n ¹ ]}, The characteristic functions at each scale are stored in matrices S ^M , L ^M , and C ^M respectively. S ^M is used to store _sk (i). The k-th row and i-th column of S ^M stores the area parameter _sk (i) of point p ¹ (i) at the k-th scale. L ^M is used to store l _k (i). The k-th row and i-th column of L ^M stores the arc length parameter l _k (i) of point p ¹ (i) at the k-th scale. C ^M is used to store c _k (i). The k-th row and i-th column of C ^M stores the centroid distance parameter c _k (i) of point p ¹ (i) at the k-th scale. S ^M , L ^M , and C ^M are finally used as grayscale images of the three shape features of the sub-shape part S ¹ of the shape sample S in the full-scale space: GM ¹ ={S ^M , L ^M , C ^M }, Among them, S ^M , L ^M , and C ^M are matrices of size m×n, each representing a grayscale image; Then, the three grayscale images of the sub-shape part ^S1 are used as the three RGB channels to obtain a color image as the feature expression image of the sub-shape part ^S1. 8. A shape recognition method according to claim 7, characterized in that: in the step 7, the feature expression image samples of each sub-shape part of all training shape samples are input into the convolutional neural network to train the convolutional neural network model; different sub-shape parts of each type of shape have different category labels; after the convolutional neural network is trained to convergence, for any shape sample S, the feature expression images {T ^num | num∈[1, N]} corresponding to the N sub-shape parts formed by segmentation are respectively input into the trained convolutional neural network, and the output of the second fully connected layer of the network is the feature vector of the corresponding sub-shape part Where Vec is the number of neurons in the second fully connected layer; The structure of the convolutional neural network includes an input layer, a pre-training layer and a fully connected layer; the pre-training layer is composed of the first four modules of the VGG16 network model, and the parameters obtained after training the four modules in the imagenet dataset are used as initialization parameters, and three fully connected layers are connected after the pre-training layer; The first module in the pre-training layer specifically includes 2 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 64, the size is 3×3, and the size of the pooling layer is 2×2; the second module specifically includes 2 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 128, the size is 3×3, and the size of the pooling layer is 2×2; the third module specifically includes 3 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 256, the size is 3×3, and the size of the pooling layer is 2×2; the fourth module specifically includes 3 convolutional layers and 1 maximum pooling layer, where the number of convolutional kernels in the convolutional layer is 512, the size is 3×3, and the size of the pooling layer is 2×2; the calculation formula for each convolutional layer is: C _O =φ _relu (W _C ·C _I +θ _C ), Among them, θ _C is the bias vector of the convolutional layer; W _C is the weight of the convolutional layer; C _I is the input of the convolutional layer; C _O is the output of the convolutional layer; The fully connected layer module specifically includes 3 fully connected layers, where the first fully connected layer contains 512 nodes, the second fully connected layer contains Vec nodes, and the third fully connected layer contains _NT nodes; _NT is the sum of the number of segmented sub-shape parts of all categories of shapes; the calculation formula for the first two fully connected layers is: F _O =φ _tanh (W _F ·F _I +θ _F ), Among them, φ _tanh is the tanh activation function, θ _F is the bias vector of the fully connected layer; W _F is the weight of the fully connected layer; _FI is the input of the fully connected layer; _FO is the output of the fully connected layer; The last fully connected layer is the output layer, and its output calculation formula is: Y _O =φ _softmax (W _Y ·Y _I +θ _Y ), Among them, φ _softmax is the softmax activation function, θ _Y is the bias vector of the output layer, each neuron in the output layer represents a corresponding sub-shape part category, W _Y is the weight of the output layer, Y _I is the input of the output layer; Y _O is the output of the output layer. 9. A shape recognition method according to claim 8, characterized in that: the specific method of constructing the feature matrix of the shape sample in step eight is: For any shape sample S, the N sub-shape parts formed by segmentation have the corresponding feature matrix expression The calculation formula is: Wherein, _Fa represents the a-th row vector of the matrix F, and ^fa is the eigenvector of the a-th sub-shape part outputted in step 7. Represents a zero vector of dimension size Vec. 10. A shape recognition method according to claim 9, characterized in that: in the step 9, a graph convolutional neural network structure is constructed, including a preprocessing input layer, a hidden layer and a classification output layer, and an adjacency matrix is performed in the preprocessing input layer Normalization preprocessing, specifically: in I _N is the identity matrix, is the degree matrix, After normalization preprocessing The hidden layer includes two graph convolution layers, and the calculation formula of each graph convolution layer is: in, is the weight of the convolutional layer of the graph; H _I is the input of the convolutional layer of the graph, and the input of the first convolutional layer is the feature matrix of the shape sample H _O is the output of the graph convolutional layer; The calculation formula of the classification output layer is: Among them, φ _softmax is the softmax activation function, G _I is the input of the output layer, that is, the output of the second graph convolutional layer, G _W is the weight of the output layer; G _O is the output of the output layer; each neuron in the output layer represents a corresponding shape category. 11. A shape recognition method according to claim 10, characterized in that: the specific method for implementing contour shape classification and recognition in step 10 is: training the graph convolutional neural network model until convergence; for any test shape sample, first extract the key points of the shape contour, calculate the curvature value at each key point, judge its concavity and convexity, obtain the candidate segmentation points, and then adjust the curvature screening threshold to obtain the shape segmentation points; according to the two principles (1) and (2) in step 5, obtain the segmentation line segment set, calculate the segmentation cost of the segmentation line segment in the segmentation line segment set; if the number of segmentation line segments is less than Then all segmentation line segments are used to segment the shape; otherwise, according to the minimum segmentation cost The shape is segmented by segmentation line segments; the color feature expression image of each sub-shape part is calculated and input into the trained convolutional neural network, and the output of the second fully connected layer of the convolutional neural network is used as the feature vector of the sub-shape part; the shape directed graph of the test shape sample is constructed, and its adjacency matrix and feature matrix are calculated and input into the trained graph convolutional neural network model. The shape category corresponding to the maximum value in the output vector is judged as the shape type of the test sample, thereby realizing shape classification and recognition.