CN104952065A - Method for building multilayer detailed skeleton model of garment images - Google Patents

Abstract

The invention discloses a method for establishing a multi-level detailed skeleton model of clothing images. For the clothing outline extracted from the image in advance, use the constrained Delaunay triangulation method to extract the clothing skeleton corresponding to the clothing outline, and smooth the extracted clothing skeleton; at the same time extract the symmetry axis of the clothing outline, and use the clothing outline symmetry axis to match the clothing skeleton Branch, use the improved web page sorting method to calculate the importance value of the key points of the skeleton; use the iterative cycle to calculate the importance value of the key points of the clothing skeleton and simplify the clothing skeleton to obtain the final multi-level detail skeleton model of the clothing image. The invention can sequentially simplify the clothing skeleton, and can obtain the multi-level detailed skeleton model of the clothing image from simple to complicated.

Description

A method for building multi-level details skeleton model of clothing image

technical field

The invention relates to a method for establishing a skeleton model of a clothing image, in particular to a method for establishing a multi-level detailed skeleton model of a clothing image.

Background technique

In the field of image processing, skeletons can describe objects in a simple and compact way. Especially in the application of clothing image processing, skeletons can embody a series of information of clothing, such as symmetry, directionality, and attributes. However, in practical applications, the skeleton is not only easily disturbed by the outline of the object to produce unreasonably small branches, but also may contain too many small branches. Therefore, skeleton simplification becomes an important means to extract effective features in skeletons with rich details. Most of the traditional skeleton simplification methods adopt a simple clipping method, that is, first select a parameter related to the skeleton attribute (usually the length is selected as the parameter), and then cut out the interference branch and the unimportant skeleton branch by setting the threshold of the parameter. . But this method can only get a relatively simplified skeleton, but cannot extract the most important skeleton prototype; in addition, the threshold of the parameters is not easy to determine, and often needs to be set and adjusted manually. This method has the shortcomings of large uncertainty and low efficiency, so it cannot meet the multiple and rapid application of the skeleton in practical applications.

Contents of the invention

Aiming at the shortcomings of large uncertainty and low efficiency in the traditional skeleton simplification method mentioned above, the present invention provides a method for establishing a multi-level detailed skeleton model of clothing images, by applying the modified PageRank method to the construction of clothing images In the skeleton, the importance level of the skeleton branch is evaluated, and different levels of classification are carried out to automatically build a multi-level detail model. This model has great application prospects in clothing classification and clothing matching.

As shown in Figure 1, the technical scheme adopted in the present invention comprises the following steps:

Step 1. Extract the clothing skeleton: extract the clothing outline from the clothing image with the OPENCV tool and sample the polygon at equal intervals to obtain the polygon of the clothing outline, use the polygons that make up the clothing outline as constraints, and perform Delaunay triangulation on the inside of the polygon, so that the triangles inside the polygon Keeping the properties of the Delaunay triangle, a triangle with two different types of edges is obtained; the two different types of edges are the interior edges existing in the interior of the clothing outline and the boundary edges existing on the clothing outline, and the boundary edge is composed of two adjacent clothing The connection lines between the contour points are composed, and then the divided triangles are divided into three types according to the number of internal edges and boundary edges, and the internal skeleton line segments of each triangle are extracted, and the end points of the internal skeleton line segments of all triangles are connected end to end. , get the clothing skeleton;

In the above clothing image, the background color of the image is single and has obvious color difference from that of the clothing, which can show the outline of the clothing.

Step 2. Smooth the extracted clothing skeleton:

Extract the clothing skeleton branches in the clothing skeleton. In order to smooth the clothing skeleton, perform Bezier curve fitting on each clothing skeleton branch respectively. The two end points of the clothing skeleton branch are used as the beginning and end points of the Bezier curve, and the skeleton connection point in the middle of the clothing skeleton branch is As the control point of the Bezier curve; this can not only achieve a smooth effect, but also keep the position of the key point unchanged, so as to ensure that the overall position of the clothing skeleton remains unchanged.

Step 3. Extract the symmetry axis of the clothing outline:

Most of the clothing has symmetry, so first use the principal component analysis PCA (Principal Component Analysis) method to reduce the dimensionality of the equidistant sampling points of the clothing outline to obtain two eigenvectors, and the two eigenvectors are respectively used as the main direction and the secondary direction; The center of gravity of the clothing outline is the passing point, and the main direction and the secondary direction are respectively used as the straight line direction to form two PCA axes. Each PCA axis divides the clothing outline into contour lines P and Q on both sides;

Taking any one of the PCA axes as the mirror axis, mirror the contour line P on one side to the other side to obtain the contour line P' of the mirrored part, and the contour line P' of the mirrored part and the contour of the other side obtained by the original segmentation The line Q is on the same side of the mirror axis, and the mirror Hausdorff distance of the two contour lines P′ and Q on the same side of the mirror axis is calculated (the mirror Hausdorff distance is expressed as MHD (Mirror Hausdorff Distance) value, which is used to measure the value obtained by dividing by the PCA axis The symmetry of the two parts of the contour line P and Q), respectively calculate the two PCA axis divisions to obtain the mirror Hausdorff distance of the contour lines on both sides, and select the PCA axis corresponding to the smaller mirror Hausdorff distance (that is, the axis with more contour symmetry) As the initial clothing outline symmetry axis l ₀ , the initial clothing outline symmetry axis l ₀ is iteratively adjusted to obtain the real clothing outline symmetry axis l;

Step 4. Match the clothing skeleton branch:

Use the symmetry axis l of the real clothing outline to classify the clothing skeleton branches left and right according to the position of the center of gravity, and calculate the mirror Hausdorff distance between the left and right clothing skeleton branches; select the two left and right corresponding to the minimum value of the mirror Hausdorff distance in turn The clothing skeleton branch is used as a matching pair, and the two left and right clothing skeleton branches that have been selected as a matching pair are not used as objects for the next selection of matching pairs until the clothing skeleton branch on one side has been selected, and multiple sets of matching pairs are obtained;

Arrange the mirrored Hausdorff distances from small to large, and then use the Otsu Otsu method. Each mirrored Hausdorff distance is used as a split value in turn. Use the split value to divide all the mirrored Hausdorff distances into two categories according to the size, and calculate the respective intra-class values of the two categories. Variance and inter-class variance, take the segmentation value corresponding to the smallest intra-class variance and the largest inter-class variance of the two classes as the optimal threshold, and remove the matching pairs of clothing skeleton branches corresponding to the mirror Hausdorff distance greater than the optimal threshold; greater than the optimal threshold The clothing skeleton branch matching pairs corresponding to the mirror image Hausdorff distance of are usually some matching pairs with low symmetry and matching errors, so these clothing skeleton branch matching pairs are finally removed.

Step 5, utilize the improved web page sorting method (PageRank method) to calculate the importance value of skeleton key point:

Each skeleton key point is given the same initial value of importance as the page node of the webpage sorting method, and each clothing skeleton branch represents a link between two skeleton key points (two endpoints of the clothing skeleton branch), including the chain and in-chain;

All the page nodes and their mutual links constitute a network graph model, and the importance value of each page node in the network graph model can be calculated by using the PageRank method. The PageRank method first assigns the same importance value to each webpage, and updates the PageRank score (defined as PR value) of each page node through the iterative calculation of the relationship between outbound and inbound links until the score is stable. In the present invention, the improved PageRank method adds the path distances of outgoing links and incoming links (that is, the length of clothing skeleton branches) to the method.

The length of the clothing skeleton branch is used as the distribution factor of the importance value in the web page sorting method, and the web page sorting method is used to iteratively calculate and update through the relationship between the outgoing link and the incoming link until the value is stable and the importance value of each page node is finally obtained. Specifically The following formula is used:

$\mathrm{<span\; class="notranslate">\; PR</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\frac{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; PR</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$

Among them, PR(E) represents the importance value of page node E, m represents the number of page nodes E _i linked by page node E, N(E _i ) represents the number of page nodes linked by page node E i, and n represents the number of page nodes E _i linked to. The number of page nodes E _j linked by node E _i , L(E _i , E _j ) represents the path length between two interlinked page nodes E _i and E _j , and i represents the page node E linked by page node E The ordinal number of _i , j represents the ordinal number of the page node E _j linked by the page node E _i ;

Step 6. Use the iterative cycle to calculate the importance value of the clothing skeleton, and establish a multi-level detailed skeleton model of the clothing image:

The only clothing skeleton branch connected to the clothing skeleton terminal point is defined as the skeleton terminal branch, and the importance value of the key points of the clothing skeleton is repeatedly calculated and the clothing skeleton is simplified to obtain the final multi-level detail skeleton model of the clothing image.

Starting from the initial clothing skeleton, the above process can obtain a series of sequentially simplified clothing skeletons, which together form a series of clothing skeletons from complex to simple. For the convenience of application, the arrangement of these clothing skeletons is rearranged in reverse order to obtain a series of clothing skeletons from simple to complex, that is, the multi-level detail skeleton model of clothing images.

In the described step one, the triangles obtained by splitting are divided into three categories according to the quantity of internal sides and boundary sides: the triangles with one internal side and two boundary sides are Class I triangles, and the triangles with two internal sides and one boundary side are class II triangles, and those with three interior sides are class III triangles.

In said step 1, the internal skeleton line segment includes the midpoint of the internal edge in the type I triangle and the connection line of the triangle vertex to which the internal edge corresponds, the connection line between the midpoints of the two internal edges in the type II triangle, and the line between the midpoints of the internal edge in the type III triangle. The respective lines connecting the midpoints of the three interior sides of the triangle to the Voronoi points of the triangle.

In the second step, the clothing skeleton branch in the clothing skeleton is extracted as follows: according to step one, the points forming the clothing skeleton have the following three types: the skeleton intersection point connecting three triangle internal skeleton line segments, and the point connecting two triangle internal skeleton line segments. Skeleton connection points and skeleton end points that only connect a skeleton line segment inside a triangle; take the skeleton intersection point and the skeleton end point as the skeleton key point, if there is no skeleton intersection point on the skeleton line segment path between the two skeleton key points, then the The path of the skeleton line segment is the branch of the clothing skeleton, and the two key points of the skeleton are the two endpoints of the branch of the clothing skeleton.

In the third step, the initial clothing outline symmetry axis l _k is iteratively adjusted in the following manner to approach the real clothing outline symmetry axis:

3.1) Clothing contour symmetry axis l _k divides the clothing contour into contour lines P _k and Q _k on both sides, k is the number of adjustments to the clothing contour symmetry axis, and maps a part of the contour lines P _k to the clothing contour symmetry axis l _k to the other side, to obtain the mirrored partial profile P′ _k ;

3.2) Take the points on the contour P′ _k of the mirrored part and the contour line Q _k on the other side as two point sets, use the ICP (Iterative Closest Points) method to continuously calculate and update to obtain the contour P″ _k of the mirrored part, and the updated mirror image Partial profile P″ _k and its initial unmapped partial profile P _k carry out one-to-one correspondence of points, and obtain corresponding pairs of matching points;

3.3) Use the least squares method to fit the midpoint of each pair of matching point connecting line segments as a point set to obtain a new clothing outline symmetry axis l _k+1 ;

3.4) Repeat the above 3.1), 3.2), 3.3) steps to iteratively adjust until the deviation angle between the symmetry axis l n of the clothing outline obtained in the current step and the symmetry axis l _{n -1} of the clothing outline obtained in the previous step is less than 2°, and stop the calculation , the clothing outline symmetry axis l _n obtained from the last calculation is taken as the real clothing outline symmetry axis l, and n is the number of adjustments to obtain the real clothing outline symmetry axis.

In the fourth step, the mirrored Hausdorff distances between the left and right clothing skeleton branches include the mirrored Hausdorff distances between any clothing skeleton branch on one side and each clothing skeleton branch on the other side.

Said step 6 specifically includes: 6.1) The importance value of each skeleton key point calculated in step 5 is matched with each clothing skeleton branch matching pair obtained in step 4, and the importance of two skeleton terminal points in the matching pair of skeleton terminal branches is calculated. the average of the values;

6.2) Select the matching pair of skeleton terminal points whose average value is less than the initial value of importance, and remove the matching pair of corresponding skeleton terminal branches; The end branch of the skeleton of the value generates a new clothing skeleton;

6.3) Iteratively repeat steps 6.1) and 6.2) until the number of branches of the remaining clothing skeleton is 1 or 0, and the importance value of the new simplified clothing skeleton cannot be calculated, then the establishment of the multi-level detail skeleton of the clothing image is completed Model.

The beneficial effects that the present invention has are:

The invention applies the modified PageRank method to the skeleton of the clothing image, evaluates the importance level of the skeleton branches, and performs different levels of classification to automatically establish a multi-level detail model. The invention sequentially simplifies the clothing skeleton according to the degree of importance, and has wide application prospects in clothing identification and clothing matching. The method of the invention has high efficiency and good certainty, and can satisfy the multiple and rapid application of the skeleton in practical applications.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

Fig. 2 is the clothing outline extracted by the embodiment of the present invention.

Fig. 3 is the clothing skeleton extracted by the embodiment of the present invention.

Fig. 4 is the clothing skeleton branch of the embodiment of the present invention.

Figure 5 is a smooth garment skeleton according to an embodiment of the present invention.

Fig. 6 is the calculation process and results of the symmetry axis of the garment outline according to the embodiment of the present invention.

Fig. 7 is a matching pair of clothing skeleton branches according to an embodiment of the present invention.

Fig. 8 is the process of simplifying the clothing skeleton and the key points of the skeleton according to the embodiment of the present invention.

In the figure: 1. Clothing, 2. Clothing outline, 3. Internal edge, 4. Boundary edge, 5. Clothing skeleton, 6. Skeleton intersection point, 7. Skeleton connection point, 8. Skeleton end point, 9. Skeleton key point , 10, clothing skeleton branch, 11, clothing outline symmetry axis, 12, PCA axis, 13, matching pair, 14, multi-level detailed skeleton model, 15, skeleton terminal branch.

Detailed ways

The present invention will be further described below in conjunction with drawings and embodiments.

Specific embodiments of the present invention are as follows:

Step 1: As shown in Figure 2, extract the clothing outline 2 corresponding to clothing 1, and use the polygons that make up the clothing outline 2 as constraints to perform Delaunay triangulation.

As shown in FIG. 3( a ), the divided triangle has internal sides 3 and boundary sides 4 . The divided triangles are divided into the following three categories according to the number of interior sides 3 and boundary sides 4: the triangle with one interior side 3 and two boundary sides 4 is a type I triangle; the triangle with two interior sides 3 and one A triangle with boundary side 4 is a type II triangle; a triangle with three interior sides 3 is a type III triangle.

Extract the internal skeleton line segment of each triangle, that is, the connection line between the midpoint of the internal side in the type I triangle and the triangle vertex corresponding to the internal side, the connection line between the midpoints of the two internal sides in the type II triangle, and the connection line between the midpoints of the two internal sides in the type III triangle. The respective lines connecting the midpoints of the three interior sides and the Voronoi points of the triangle. Connect all the triangle internal skeleton line end points end to end to get the clothing skeleton 5, as shown in Figure 3(b) and Figure 3(c).

Step 2: Smooth the extracted clothing skeleton 5 .

As shown in Figure 4(b), there are three types of points that make up the clothing skeleton 5: the skeleton intersection point 6 that connects three triangle interior skeleton line segments, the skeleton connection point 7 that connects two triangle interior skeleton line segments, and only one triangle interior skeleton line segment Skeleton end point 8 of the line segment.

Define the skeleton intersection point 6 and the skeleton end point 8 as the skeleton key point 9, if there is no skeleton intersection point 6 on the skeleton line segment path between the two skeleton key points 9, then define this skeleton line segment path as the clothing skeleton branch 10. The two skeleton key points 9 are the two endpoints of the clothing skeleton branch 10, as shown in Figure 4(c). The two endpoints of the clothing skeleton branch 10 are used as the beginning and end points of the Bezier curve, and the middle skeleton connection point is used as the control point of the Bezier curve. Each clothing skeleton branch 10 is fitted with a Bezier curve, and the result is shown in Figure 5(b) .

Step 3: Extract the symmetry axis 11 of the clothing outline.

Using the PCA (Principal Component Analysis) method to reduce the dimensionality of the equidistant sampling points of the clothing outline to obtain two eigenvectors, the two eigenvectors are respectively used as the main direction and the secondary direction. As shown in Figure 6, the center of gravity of the clothing outline 2 is taken as the passing point, and the main direction and the secondary direction are respectively used as the straight line directions to form two PCA axes 12. Each PCA axis 12 divides the clothing outline 2 into two parts. Lines P and Q. Taking one of the PCA axes 12 as the mirror axis, mirroring a part of the contour line _Ej to the other side to obtain the mirrored part of the contour line P', the mirrored part of the contour line P' and the other part of the contour line Q obtained by the original segmentation are in The same side of the mirror axis, and calculate the mirror Hausdorff distance of the two contour lines P' and Q on the same side of the mirror axis, which is called the MHD value. Calculate the MHD values of the two parts of the contour line obtained by dividing the two PCA axes 12 respectively, and select the PCA axis 12 corresponding to the smaller value as the initial clothing contour symmetry axis l _k (k is the number of clothing contour symmetry axis adjustments).

Make further iterative adjustments to the initial clothing outline symmetry axis l _k to approach the real clothing outline symmetry axis 11:

3.1) The initial clothing outline symmetry axis l _k divides the clothing outline into two parts of the contour line P _k and Q _k , and maps one part of the contour line l _k to the other side based on the initial clothing outline symmetry axis l _k to obtain the mirror part Profile P′ _k , as shown in Fig. 6(a);

3.2) Take the points on the mirrored part contour P′ _k and the segmented part of the contour line Q _k as two point sets, use the ICP method to continuously calculate and update the mirrored part contour P″ _k , and the updated mirrored part contour P ″ _k and its initial unmapped partial profile P _k carry out point-to-point correspondence, as shown in Figure 6(b);

3.3) Use the least squares method to fit the midpoint of each pair of corresponding points as a point set to obtain a new clothing outline symmetry axis l _k+1 , as shown in Figure 6(c);

3.4) Repeat the above steps 3.1), 3.2), and 3.3) until the obtained clothing outline symmetry axis l _n (n is the number of adjustments to obtain the real clothing outline symmetry axis) and the clothing outline symmetry axis l _n- If the deviation angle of ₁ is less than 2°, the calculation is stopped, and l _{n is} taken as the real clothing outline symmetry axis 11.

Step 4: Utilize the clothing outline symmetry axis 11 to classify the clothing skeleton branches 10 left and right, as shown in Figure 7 (a), calculate the pairwise MHD values of the left and right clothing skeleton branches 10, the pairwise MHD values in the present embodiment are as follows As shown in Table 1, the two left and right clothing skeleton branches corresponding to the minimum value of MHD value are sequentially selected as matching pairs, and the two left and right clothing skeleton branches that have been selected as matching pairs are not used as the objects for the next selection of matching pairs until one side The clothing skeleton branch of has been selected, and multiple matching pairs are obtained:

Table 1

The results obtained according to Table 1 are shown in Fig. 7(b) to Fig. 7(m). Using the Otsu Otsu method, the MHD values are arranged in ascending order, and each MHD value is selected as a threshold value. The threshold value divides the matching pair 13 into two categories according to the corresponding MHD value. The variance within the two categories is the smallest and the category When the variance between is the largest, the optimal threshold of this embodiment is 20.082. The clothing skeleton branch matching 13 pairs corresponding to the MHD value greater than the optimal threshold is usually some matching pairs with low symmetry and matching errors. According to the results in Table 1, they are (l) and (m) in Figure 7 , and finally remove these matching pairs 13.

Step 5: Use the improved PageRank method to calculate the importance value of skeleton key point 9, which is the PR value.

Using the skeleton key point 9 as the page node of the PageRank method, each clothing skeleton branch 10 represents a link between two skeleton key points 9 (two endpoints of the clothing skeleton branch 10). All page nodes and their interlinkages constitute a network graph model. The improved PageRank method formula is:

$\mathrm{<span\; class="notranslate">\; PR</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; PR</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$

PR(E) represents the PR value of page node E, m represents the number of page nodes E _i linked by page node E, N(E _i ) represents the number of page nodes linked by page node E _i , and n represents page node E _i The number of linked page nodes E _j , L(E _i , E _j ) represents the path length between two mutually linked page nodes E _i and E _j .

In this embodiment, the same PR initial value of 1.0 is assigned to each skeleton key point 9, and the improved PageRank formula is used to iteratively calculate until the value is stable, and the network PR value of the clothing skeleton is obtained.

Step 6: Use an iterative cycle to calculate the network PR value of the clothing skeleton, and establish a multi-level detailed skeleton model of the clothing image14.

Define the only clothing skeleton branch 10 connected to the clothing skeleton terminal point 8 as the skeleton terminal branch 15, repeatedly calculate the PR value of the clothing skeleton key point 9 and simplify the clothing skeleton 5, and obtain the final multi-level detail skeleton model 14 of the clothing image :

6.1) Calculate the PR value of the skeleton key point 9 of the clothing according to step five, and match each clothing skeleton branch 10 according to step four, and calculate the PR average value of the corresponding skeleton end point 8 in the matching pair of skeleton end branches 15;

6.2) Select 8 matching pairs of skeleton terminal points whose average PR value is less than the initial value of 1.0 for the 15 matching pairs of the skeleton terminal branches, and cut out the corresponding skeleton terminal branch 15 matching pairs; for the unmatched single skeleton terminal branch 15, directly Remove the skeleton terminal branch 15 whose PR value is less than 1.0 at the skeleton terminal point 8, and generate a new simplified clothing skeleton.

6.3) Repeat steps 6.1) and 6.2) for the newly obtained simplified clothing skeleton until the last clothing skeleton branch has 10 remaining 1 or 0. The PR values of each step in this embodiment are shown in Table 2, and the network diagram model is shown in Figure 8 As shown, in Table 2, the results of one calculation correspond to Figure 8(a), the results of two calculations correspond to Figure 8(b), and the results of three calculations correspond to Figure 8(c).

Table 2

Starting from the initial clothing skeleton 5, the above process can obtain a series of sequentially simplified clothing skeletons, that is, obtain the multi-level detail skeleton model 14 of the final clothing image.

It can be seen that the present invention establishes a model of multi-level details, and sequentially simplifies the clothing skeleton according to the degree of importance, which has high efficiency and good certainty, and can be used for clothing identification and clothing matching.

The above specific embodiments are used to explain the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modification and change made to the present invention will fall into the protection scope of the present invention.

Claims (7)

1. A method of setting up a clothing image multi-level detail skeleton model, characterized in that: Step 1. Extract clothing skeleton: The clothing contour is extracted from the clothing image by the OPENCV tool, and the polygon of the clothing contour is obtained by equidistant sampling. The polygons that make up the clothing contour are used as constraints, and the Delaunay triangulation is performed on the interior of the polygon, so that the triangles inside the polygon maintain the attributes of the Delaunay triangle. A triangle with two different types of edges; the two different types of edges are the internal edges existing in the interior of the clothing outline and the boundary edges existing on the clothing outline, and then the divided triangle is obtained according to the internal edge and the boundary edge The quantity is divided into three categories, the internal skeleton line segment of each triangle is extracted, and the end points of the internal skeleton line segments of all triangles are connected end to end to obtain the clothing skeleton; Step 2. Smooth the extracted clothing skeleton: Extract the clothing skeleton branches in the clothing skeleton, perform Bezier curve fitting on each clothing skeleton branch respectively, use the two end points of the clothing skeleton branch as the beginning and end points of the Bezier curve, and the skeleton connection point in the middle of the clothing skeleton branch as the control of the Bezier curve point; Step 3. Extract the symmetry axis of the clothing outline: Using principal component analysis (PCA) method to reduce the dimensionality of the equidistant sampling points of the clothing outline to obtain two eigenvectors, the two eigenvectors are respectively used as the main direction and the secondary direction; The secondary direction is a straight line direction, each of which forms two PCA axes, and each PCA axis divides the clothing contour into contour lines P and Q on both sides; Taking any one of the PCA axes as the mirror axis, mirror the contour line P on one side to the other side to obtain the mirror part contour line P′, and calculate the mirror Hausdorff of the two contour lines P′ and Q on the same side of the mirror axis Distance, respectively calculate the two PCA axis divisions to obtain the mirrored Hausdorff distance of the contour lines on both sides, select the PCA axis corresponding to the smaller mirrored Hausdorff distance as the initial clothing outline symmetry axis l ₀ , and iteratively adjust the initial clothing outline symmetry axis l ₀ to obtain The symmetry axis l of the real clothing outline; Step 4. Match the clothing skeleton branch: Use the symmetry axis l of the real clothing outline to classify the clothing skeleton branches left and right according to the position of the center of gravity, and calculate the mirror Hausdorff distance between the left and right clothing skeleton branches; select the two left and right corresponding to the minimum value of the mirror Hausdorff distance in turn The clothing skeleton branch is used as a matching pair, and the two left and right clothing skeleton branches that have been selected as a matching pair are not used as objects for the next selection of matching pairs until the clothing skeleton branch on one side has been selected, and multiple sets of matching pairs are obtained; Arrange the mirrored Hausdorff distances from small to large, and then use the Otsu Otsu method. Each mirrored Hausdorff distance is used as a split value in turn. Use the split value to divide all the mirrored Hausdorff distances into two categories according to the size, and calculate the respective intra-class values of the two categories. Variance and inter-class variance, take the segmentation value corresponding to the smallest intra-class variance and the largest inter-class variance of the two classes as the optimal threshold, and remove the clothing skeleton branch matching pairs corresponding to the mirror Hausdorff distance greater than the optimal threshold; Step five, using the improved web page sorting method to calculate the importance value of the key points of the skeleton: Each skeleton key point is given the same initial value of importance as the page node of the webpage sorting method, and each clothing skeleton branch represents a link between two skeleton key points, including the out-link and in-link; the clothing skeleton branch The length of is used as the distribution factor of the importance value in the web page sorting method. The web page sorting method is used to iteratively calculate and update through the relationship between outbound and inbound links until the value is stable and the importance value of each page node is finally obtained. Specifically, the following formula is used: $\mathrm{<span\; class="notranslate">\; PR</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\frac{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; PR</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$ Among them, PR(E) represents the importance value of page node E, m represents the number of page nodes E _i linked by page node E, N(E _i ) represents the number of page nodes linked by page node E i, and n represents the number of page nodes E _i linked to. The number of page nodes E _j linked by node E _i , L(E _i , E _j ) represents the path length between two interlinked page nodes E _i and E _j , and i represents the page node E linked by page node E The ordinal number of _i , j represents the ordinal number of the page node E _i linked by the page node E _i ; Step 6. Use the iterative cycle to calculate the importance value of the clothing skeleton, and establish a multi-level detailed skeleton model of the clothing image: The only clothing skeleton branch connected to the clothing skeleton terminal point is defined as the skeleton terminal branch, and the importance value of the key points of the clothing skeleton is repeatedly calculated and the clothing skeleton is simplified to obtain the multi-level detail skeleton model of the final clothing image. 2. a kind of method for setting up the multi-level detail skeleton model of clothing image according to claim 1, is characterized in that: in described step 1, the triangle that subdivides obtains adopts the following according to the quantity of internal side and border side that has. There are three types of triangles: Type I triangles with one interior side and two boundary sides, Type II triangles with two interior sides and one boundary side, and Type III triangles with three interior sides. 3. a kind of method of setting up clothing image multi-level details skeleton model according to claim 1, is characterized in that: in described step 1, interior skeleton line segment comprises the midpoint of interior edge and this interior edge in class I triangle The line connecting the vertices of a pair of triangles, the line between the midpoints of the two interior sides of a type II triangle, the respective lines connecting the midpoints of the three interior sides of a type III triangle and the Voronoi points of the triangle. 4. A method for setting up a multi-level detailed skeleton model of clothing images according to claim 1, characterized in that: in said step 2, extracting the clothing skeleton branches in the clothing skeleton is specifically: forming the clothing skeleton according to step 1 There are three types of points: the skeleton intersection point connecting three triangle interior skeleton line segments, the skeleton connection point connecting two triangle interior skeleton line segments, and the skeleton end point connecting only one triangle interior skeleton line segment; the skeleton intersection point and skeleton end point As a skeleton key point, if there is no skeleton intersection point on the skeleton line segment path between two skeleton key points, the skeleton line segment path is a clothing skeleton branch, and the two skeleton key points are the two endpoints of the clothing skeleton branch. 5. A method for setting up a multi-level detailed skeleton model of clothing images according to claim 1, characterized in that: in said step 3, the initial clothing outline symmetry axis ₁₀ is iteratively adjusted in the following manner to approach real clothing Contour symmetry axis: 3.1) Clothing contour symmetry axis l _k divides the clothing contour into contour lines P _k and Q _k on both sides, k is the number of adjustments to the clothing contour symmetry axis, and maps a part of the contour lines P _k to the clothing contour symmetry axis l _k to the other side, to obtain the mirrored partial profile P′ _k ; 3.2) Take the points on the contour P′ _k of the mirrored part and the contour line Q _k on the other side as two point sets, use the ICP (Iterative Closest Points) method to continuously calculate and update to obtain the contour P″ _k of the mirrored part, and the updated mirror image Partial profile P″ _k and its initial unmapped partial profile P _k carry out one-to-one correspondence of points, and obtain corresponding pairs of matching points; 3.3) Use the least squares method to fit the midpoint of each pair of matching point connecting line segments as a point set to obtain a new clothing outline symmetry axis l _k+1 ; 3.4) Repeat the above 3.1), 3.2), 3.3) steps to iteratively adjust until the deviation angle between the symmetry axis l n of the clothing outline obtained in the current step and the symmetry axis l _{n -1} of the clothing outline obtained in the previous step is less than 2°, and stop the calculation , the clothing outline symmetry axis l _n obtained from the last calculation is taken as the real clothing outline symmetry axis l, and n is the number of adjustments to obtain the real clothing outline symmetry axis. 6. A kind of method for setting up the multi-level detail skeleton model of clothing image according to claim 1, is characterized in that: in described step 4, the mirror image Hausdorff distance between left and right clothing skeleton branches two by two comprises one of them The mirrored Hausdorff distance between any clothing skeleton branch on one side and each clothing skeleton branch on the other side. 7. A method for setting up a multi-level detailed skeleton model of clothing images according to claim 1, wherein said step six specifically includes: 6.1) The importance value of each skeleton key point calculated by step 5 is matched with the matching pair of each clothing skeleton branch obtained in step 4, and the average value of the importance values of the two skeleton terminal points in the matching pair of skeleton terminal branches is calculated; 6.2) Select the matching pair of skeleton terminal points whose average value is less than the initial value of importance, and remove the matching pair of corresponding skeleton terminal branches; The end branch of the skeleton of the value generates a new clothing skeleton; 6.3) Iteratively repeat steps 6.1) and 6.2) until the number of branches of the remaining clothing skeleton is 1 or 0, and the importance value of the new simplified clothing skeleton cannot be calculated, then the establishment of the multi-level detail skeleton of the clothing image is completed Model.