CN106570475A - Purple clay teapot seal retrieval method - Google Patents

Abstract

The invention discloses a search method for a purple sand teapot seal. Firstly, crop and zoom the Zisha teapot seal image obtained by taking pictures with the mobile phone, and then change the brightness and contrast of the image by adjusting the mean value and mean square error of the image to obtain a standardized image. Extract the edge features and shape region features of the purple sand pot seal from the standardized image, and after feature fusion, use the trained SVM classifier to roughly divide the seal into four categories: circle, triangle, rectangle, and irregular shape. Extract the SIFT feature of the Zisha teapot seal image, introduce geometric constraints, and use an improved matching method based on belief propagation to match the image in the template library; at the same time, use the clustering method to optimize the time required for retrieval to complete the seal retrieval. The invention has good robustness to uneven illumination and affine transformations such as rotation, scaling, and side view.

Description

A search method for purple sand teapot seal

technical field

The invention belongs to the technical field of digital image processing and pattern recognition, and in particular relates to a search method for a purple sand teapot seal.

Background technique

Seal retrieval Due to the diversity of fonts and the instability of the printed carrier, the identification of Zisha teapot seals is still mainly based on visual observation. Through manual methods such as comparing the details of the seals, overlapping comparisons, and splicing comparisons, Zisha teapots can be identified. the producer. In recent years, with the vigorous development of digital image processing and pattern recognition technology, a reliable theoretical basis has been provided for computers to replace manual retrieval of purple sand pot seals.

In 2008, Ni Qi and others proposed a seal recognition algorithm based on gradient segmentation and feature matching (Ni Qi, Wang Xinsai, Li Jian, et al. Research on seal recognition algorithm based on gradient segmentation and feature matching[J].Computer Engineering and Design, 2008,29(20):5400-5402.), firstly stretch the difference between the gray level of the stamp and the gray level of the background through the equalization of the image histogram, and then use the principle of sudden change of the edge of the gray gradient to roughly pick out the edge of the stamp , and finally the edge of the stamp is detected through template matching to complete the recognition. This method is simple and fast, but it is limited by the shortcomings of template matching, and it does not have rotation invariance and scale invariance.

In 2010, Lang Haitao and others proposed a method for identifying counterfeit seals based on randomly generated feature lines from matching feature points (Lang Haitao, Lei Lan Yifei. A method for identifying counterfeit seals based on randomly generated feature lines from matching feature points: CN, CN101894260A[P ].2010.), the specific implementation method is to extract the feature points of the image to be detected, and build a database, including the position of each feature point and other descriptor information. Based on the matching feature points, the recognizable image feature lines of the seal to be tested and the reference seal are randomly generated; according to the seals obtained from the same seal in different environments, there are principles of consistent image information and similar distribution of image feature points to determine whether the image to be detected is Forged stamp image. This method is simple and efficient, but limited by illumination transformation and scaling, the lack of some feature points may cause positive samples to be misjudged as fake seal images.

In 2007, Zhang Yi and others proposed a seal identification control method (Zhang Yi, Lv Jiancheng, Zhang Quanfang, etc. Seal identification system and its control method: CN, CN 101008985A[P].2007.), first use binary The printed text to be recognized is extracted through four operation steps of transformation, skeleton extraction, frame extraction and printed text extraction. Next, in the stamp registration step, the rough registration is performed first, and the print to be recognized and the template print are adjusted to approximately the same position and direction, and the fine registration further adjusts the two prints to almost the same position and orientation. Direction; In the seal identification step, a multi-level identification strategy and a multi-feature classification fusion decision-making method are used to identify the seal to be identified and the template print. This method can well realize the identification control of the complete seal image. But when the image is missing, the effect on image registration will drop sharply.

Through the analysis of the existing seal recognition technology, it is found that there are many current methods: the global matching method is limited by the bottleneck of image matching, and does not have the ability to resist affine transformation; Applied to seal retrieval; at present, there is no systematic framework for the retrieval of Zisha teapot seals.

Contents of the invention

Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a search method for the seal of the purple clay pot, which can reliably and quickly complete the search for the seal of the purple clay pot.

Technical scheme: in order to achieve the above object, the technical scheme adopted in the present invention is:

Step 1, preprocessing the photographed image of the purple clay teapot seal to obtain a normalized image of the detection area;

Step 2, adjusting the brightness and contrast of the standardized image, and fixing the brightness and contrast of the image at a specific value by adjusting the mean value and mean square error of the image;

Step 3, initially dividing the shape of the seal in the image;

Step 4: Extract SIFT features from the image to be retrieved to form a 128-dimensional SIFT descriptor; and use the improved BP-SIFT algorithm to perform feature matching on the template image and the image to be retrieved, and determine the name in the template library by calculating the matching rate of the matching points category, complete the seal name identification.

Further, step 1 is implemented as follows: edge detection and salient area extraction are performed on the captured image to obtain the seal part; and scaling is performed to obtain a 100*100 size detection area normalization after leaving at least 5 pixels blank on the edge image.

Further, the specific method of adjusting the image in step 2 is:

Step 2-1, adjust the mean value of the image to 0, g ₁ =gu ₁ , during the adjustment process, the mean square error remains unchanged; where, g ₁ is the pixel value of the image obtained in the first step, g is the original pixel value, u ₁ is the mean value of the original image;

Step 2-2, adjust the mean square error of the image to d ₀ , g ₂ =g ₁ *d ₀ /d ₁ , during the adjustment process, the mean value remains unchanged and is still 0; where, d ₀ is the mean square error of the new target image , g ₂ is the pixel value of the image obtained in the second step, g ₁ is the pixel value of the image obtained in the first step, and d ₁ is the mean square error of the original image;

Step 2-3, adjust the mean value of the image to u ₀ , g ₃ =g ₂ +u ₀ , where g ₃ is the pixel value of the target image, g ₂ is the pixel value of the image obtained in the second step, and u ₀ is the target image mean.

Further, in step 3, the stamp images to be retrieved are divided into four categories: circular stamp images, rectangular stamp images, triangular stamp images, and irregular stamp images.

Further, the specific method for preliminary division of the shape of the seal in the image in step 3 is:

Step 3-1, the edge information of the image is obtained by detecting the canny operator, sampling is performed at intervals on the obtained edge contour, and the distance from the centroid of the image to the edge point is calculated as the boundary descriptor;

Step 3-2, obtain the largest circumscribed circle through the centroid of the image; on this basis, obtain the regional shape characteristics of the target by sampling;

Step 3-3, normalize edge features and region shape features, linearly fuse features as classifier input;

Step 3-4, use the DAG rule in multi-classification, use N (N-1)/2 SVM classifiers, the classifier forms a directed acyclic graph, when a sample comes, refine the sample from top to bottom Divide until the leaf node is reached, and initially divide the shape of the image, where N is the number of divided categories.

Further, the sampling interval in step 3 is Among them, L is the profile length.

Further, the SVM classifier in step 3 uses the RBF kernel function, which is defined as a monotone function of the Euclidean distance from point x to point y in the space, and y is the center of the kernel function:

Wherein, the influencing factors of RBF kernel SVM performance in space are the penalty parameter C and the kernel parameter σ ² ; use the grid search method to select the parameters: for Take N and M values respectively, where the range of C is set to scope set to The step distance is 0.1; all N×M values are combined to calculate the SVM generalization ability, and the combination that makes the SVM generalization performance optimal is used as the selection parameter.

Furthermore, the improved BP-SIFT algorithm in step 4 includes the introduction of two-point spatial distance ratio constraints, and the clustering of adjacent points before searching to reduce time complexity.

Further, the specific method of the improved BP-SIFT algorithm in step 4 is:

Step 4-1, using the classic SIFT algorithm to extract SIFT features from the image to be retrieved to form a 128-dimensional SIFT feature descriptor;

Step 4-2, use KMEANS clustering to cluster all feature points in the template image into C categories according to their coordinate positions; for all feature points in the template image, calculate which category they belong to, and find the nearest K neighbors in this category point;

Step 4-3, initialize all confidence levels as a constant, set n=1; for the combination of the current feature point in the template image and the current feature point in the image to be retrieved, update its matching confidence level iteratively;

Step 4-4, estimate and calculate the matching probability by combining the geometric distance constraints of the feature point and its K adjacent points and the linear function combination of the Euclidean distance between the feature points of the two images;

Step 4-5, n←n+1, then go to step (4-4), until the matching probability does not change or the number of iterations n reaches the maximum;

In steps 4-6, if the matching degree is less than the preset threshold, the two points are successfully matched; otherwise, they are not a matching pair.

Steps 4-7, repeat the above steps for all template images, calculate the matching rate between the image to be retrieved and each template image, and classify them into the category with the highest matching degree.

Beneficial effects: Compared with the current existing technology, the Zisha teapot seal retrieval method provided by the present invention has the following beneficial effects: the retrieval accuracy is higher, and machine learning and image registration technology are combined to preliminarily divide into categories before feature matching, which can Reduce the time for subsequent template matching and screen out some misretrievals; the combination of spatial information and local information has a good ability to resist affine transformations such as rotation and scaling; use clustering before searching for adjacent points to avoid large-scale global searches , greatly reducing the retrieval time.

Description of drawings

Fig. 1 is the search flowchart of purple sand teapot seal of the present invention.

Figure 2(a), Figure 2(b), and Figure 2(c) are schematic diagrams of the results of adjusting the mean and mean square error of grayscale images.

Figure 3(a), Figure 3(b), and Figure 3(c) are schematic diagrams of extracting edge features.

Figure 4(a) and Figure 4(b) are schematic diagrams of successful retrieval. The black line segment connects the two matching points of the image to be matched and the template image. Among them, 4(a) contains the rotation transformation.

Fig. 5(a) and Fig. 5(b) are schematic diagrams that such stamps are not retrieved in the template image.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings.

The invention relates to a search method for a purple sand teapot seal. Firstly, crop and zoom the Zisha teapot seal image obtained by taking pictures with the mobile phone, and then change the brightness and contrast of the image by adjusting the mean value and mean square error of the image to obtain a standardized image. Extract the edge features and shape region features of the purple sand pot seal from the standardized image, and after feature fusion, use the trained SVM classifier to roughly divide the seal into four categories: circle, triangle, rectangle, and irregular shape. Extract the SIFT feature of the Zisha teapot seal image, introduce geometric constraints, and use an improved matching method based on belief propagation to match the image in the template library; at the same time, use the clustering method to optimize the time required for retrieval to complete the seal retrieval. The invention has good robustness to uneven illumination and affine transformations such as rotation, scaling, and side view.

In the present invention, after standardizing the image of the purple clay teapot seal obtained from photography or the Internet, the mean value and mean square error are adjusted for preprocessing. Extract the edge feature, extract the regional shape feature, and use it as the input of the SVM classifier after fusion, and divide the seal image into four categories according to the geometric shape. Then extract the SIFT feature, and use the improved BP-SIFT algorithm to register with the template image to complete the seal retrieval. The process is shown in Figure 1.

Step 1, normalize the images obtained from shooting or network sources, including effective area extraction, edge cropping and image resizing.

Step 2, adjust the brightness and contrast of the normalized image. By adjusting the mean value and mean square error of the image, the brightness and contrast of the image can be fixed at a specific value, and the robustness of the stamp image to the influence of light can be enhanced. As shown in Figure 2(a), Figure 2(b), and Figure 2(c).

(1) Adjust the mean value of the image to 0. g ₁ =gu ₁ . During the adjustment process, the mean square error does not change. Among them, g ₁ is the pixel value of the image obtained in the first step, g is the original pixel value, and u ₁ is the mean value of the original image.

(2) Adjust the mean square error of the image to d ₀ , g ₂ =g ₁ *d ₀ /d ₁ . During the adjustment process, the mean does not change and remains at 0. Among them, d ₀ is the mean square error of the new target image, g ₂ is the pixel value of the image obtained in the second step, g ₁ is the pixel value of the image obtained in the first step, and d ₁ is the mean square error of the original image.

(3) Adjust the mean value of the image to u ₀ . g ₃ =g ₂ +u ₀ . In the present invention, d ₀ =50, u ₀ =127. Among them, g ₃ is the pixel value of the target image, g ₂ is the pixel value of the image obtained in the second step, and u ₀ is the mean value of the target image.

Step 3: Extract edge features and region shape features from the preprocessed image, input them into the SVM classifier, and complete the initial classification of the seal shape based on the prior knowledge of the seal. As shown in Figure 3(a), Figure 3(b), and Figure 3(c).

(3-1) The edge information of the image is obtained by detecting the canny operator, and then, the obtained edge contour is sampled at intervals, a total of 20 samples are taken, and the distance from the image centroid to the edge point is calculated as the boundary descriptor f ₁ =[b ₁ b ₂ . . . b ₂₀ ]. It includes four steps: first, find the convolution of the Gaussian filter and the image, then use the finite difference of the first-order partial derivative to calculate the direction and magnitude of the gradient; then, perform non-maximum suppression on the gradient magnitude (only keep point with the largest local variation in magnitude); finally, a double-thresholding algorithm is used to reduce the number of false edge segments.

(3-2) Find the maximum circumscribed circle by finding the centroid of the binarized image. On this basis, the regional shape features of the target are obtained by sampling. First use the formula

Find the centroid of the image. Among them, (x _c , y _c ) is the coordinates of the centroid of the image, and m is the number of collection points.

With the centroid as the center and the pixel distance from the centroid as the radius, the maximum circumscribed circle of the target is formed. Divide the radius into 5 equal distances to form 5 concentric circles. Rotate from 0 to 360 degrees with an interval of sampling, record the gray value of the pixel at the intersection of the rotation axis and the concentric circle, and form a shape descriptor with a size of 5×20 where L is the circumference of the circle.

(3-3) Normalize the edge features and region shape features, and linearly fuse the features as the input of the classifier.

Normalized fusion of edge features and shape matrix features: f _inew =(f _i -f _min )/(f _max -f _min ).

Among them, f _inew is the i-th normalized feature, and f _min and f _max are the maximum and minimum feature quantities, respectively. Form the final 6×20-dimensional feature descriptor

(3-4) Utilize the DAG rule in multi-classification, with N (N-1)/2, the present invention N=4, namely 6 SVM classifiers, classifier forms a directed acyclic graph, when coming a When sampling, the sample is gradually finely divided from top to bottom until it reaches the leaf node, and the image is finally divided into four categories: rectangle, circle, triangle, and irregular shape.

Select a classifier based on SVM and use the RBF kernel function, which is defined as a monotone function of the Euclidean distance from point x to point y in the space, and y is the center of the kernel function.

It can map samples to a high-dimensional space, and the linear kernel function is a special case of the RBF kernel function, so there is no need to consider the linear kernel function. Moreover, the RBF kernel function has fewer parameters than the polynomial kernel function, and the calculation complexity is low.

Wherein, the influencing factors of RBF kernel SVM performance in space are the penalty parameter C and the kernel parameter σ ² ; use the grid search method to select the parameters: for Take N and M values respectively, where the range of C is set to The range is set to [2 ^-10 ,2 ³ ] with a step size of 0.1. Combining all N×M values to calculate the SVM generalization ability, the combination that makes the SVM generalization performance optimal is used as the selection parameter.

Step 4: Extract SIFT features from the image to be retrieved to form a 128-dimensional SIFT descriptor. Use the improved BP-SIFT algorithm to match the template image and the image to be retrieved. By calculating the matching rate of the matching points, the category in the template library to which the name belongs is determined, and the identification of the seal name is completed. The idea of the BP-SIFT algorithm is as follows: due to the lack of spatial information constraints in traditional SIFT matching, BP-SIFT judges the similarity of geometric distances by the difference between the spatial geometric distances of feature points and their adjacent points in the two images. Improve the traditional SIFT matching using BP-SIFT algorithm to calculate the matching degree between feature points. But there are two shortcomings: first, under the affine transformation such as zooming and miscutting of the image, the distance difference between the corresponding two points in the two images is not equal; second, the time complexity of finding the adjacent point is too large. The present invention adopts a more general criterion, uses the distance ratio as a new geometric constraint condition; uses clustering before searching for adjacent points globally, and greatly reduces the time required for searching.

(1) Use the classic SIFT algorithm to extract SIFT features from the image to be retrieved to form a 128-dimensional SIFT feature descriptor;

(2) Use KMEANS clustering to cluster all feature points in the template image into C classes according to their coordinate positions; for all feature points in the template image, calculate which class they belong to, and find the nearest K adjacent points in this class;

(3) Initialize all confidence levels as a constant, and set n=1. For the combined pair of the current feature point in the template image and the current feature point in the image to be retrieved, iteratively update its matching confidence;

(4) estimate and calculate the matching probability through the linear function combination estimation of the geometric distance constraint of the feature point and its K adjacent points and the Euclidean distance between the feature points of the two images;

(5) n←n+1, then go to step (4), until the matching probability does not change or the number of iterations n reaches the maximum;

(6) If the matching degree is less than the preset threshold, the two points match successfully; otherwise, they are not a matching pair. As shown in Figure 4(a) and Figure 4(b), the schematic diagram of successful retrieval, the black line segment connects the two matching points between the image to be matched and the template image, and 4(a) contains the rotation transformation. Figure 5(a) and Figure 5(b) are schematic diagrams of unsuccessful retrieval results.

(7) Repeat the above steps for all template images, and calculate the matching rate between the image to be retrieved and each template image. classified into the category with the highest matching degree.

To sum up, the present invention combines SVM shape rough classification and feature point matching to realize the scheme of Zisha teapot seal retrieval:

(1) Adjusting the mean value and mean square deviation of the image can fix the brightness and contrast of the image at a specific value, so that the image has a good ability to resist light interference.

(2) Using SVM to roughly classify the shape features of seals can preliminarily divide the categories of images to be retrieved, reduce the amount of template image data that needs to be matched later, and improve the retrieval time.

(3) The geometric information of image feature points and their surrounding adjacent points is used in matching, which not only preserves local features, but also combines global features, which improves the accuracy of retrieval.

(4) Clustering the adjacent points around the feature points avoids the global search, and improves the retrieval time with little impact on the accuracy.

Through the experiment of 500 samples, the classification accuracy rate of the present invention to the shape of the seal reaches more than 95.6%. The matching degree for craftsmen is over 91.2%. It is more robust to changes in illumination, rotation, scaling, etc. than traditional methods of retrieving stamps with character recognition or image registration.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications are also possible. It should be regarded as the protection scope of the present invention.

Claims (9)

1. a kind of dark-red enameled pottery seal search method, it is characterised in that：Comprise the steps：

(1) pretreatment is carried out to the shooting image of dark-red enameled pottery seal, obtains the normalized images of detection zone；

(2) normalized images are carried out with the adjustment of brightness and contrast, by the average and mean square deviation that adjust image, figure is made The brightness and contrast of picture is separately fixed at a particular value；

(3) by the shape Preliminary division of seal in image；

(4) to image contract SIFT feature to be retrieved, 128 are formed and ties up SIFT description；And to template image and image to be retrieved Characteristic matching is carried out using improved BP-SIFT algorithms, by the matching rate for calculating match point, in judging the affiliated template base of name Classification, complete seal name identification.

2. dark-red enameled pottery seal search method according to claim 1, it is characterised in that：Image semantic classification in step 1 Method is：The shooting image is carried out into rim detection and marking area is extracted, obtain seal part；And zoom in and out, make side Edge retains the normalized images of the detection zone that 100*100 sizes are obtained after at least 5 pixel blank.

3. dark-red enameled pottery seal search method according to claim 1, it is characterised in that：Step (2) the adjustment image Concrete grammar is：

(201) average of image is adjusted to into 0, g₁=g-u₁, it is in course of adjustment, mean square deviation is constant；Wherein, g₁For the first step The pixel value of the image for obtaining, g is original pixel value, u₁For original image average；

(202) mean square deviation of image is adjusted to into d₀, g₂=g₁*d₀/d₁, it is in course of adjustment, average is constant, is still 0；Wherein, d₀For fresh target image mean-squared deviation, g₂The pixel value of the image obtained for second step, g₁The pixel value of image obtained by the first step, d₁For artwork mean square deviation.

(203) average of image is adjusted to into u₀, g₃=g₂+u₀, wherein, g₃For target image pixel value, g₂For second step gained The pixel value of image, u₀For target image average.

4. dark-red enameled pottery seal search method according to claim 1, it is characterised in that：In the step (3), print to be retrieved Chapter image is divided into four big class：Circular stamp image, rectangular seal image, triangle seal image, irregular seal image.

5. the dark-red enameled pottery seal search method according to claim 1 or 4, it is characterised in that：The step (3) is in image The shape of seal carries out the concrete grammar of Preliminary division, comprises the following steps：

(301) marginal information of image is obtained by the detection of canny operators, compartment of terrain is adopted on the edge contour for obtaining Sample, calculates image centroid to the distance of marginal point, as contour description；

(302) by image centroid, maximum circumscribed circle is tried to achieve；Sampling on this basis obtains the region shape feature of target；

(303) normalization edge feature and region shape feature, are input into after linear fusion feature as grader；

(304) using DAG rules in many classification, using N (N-1)/2 SVM classifier, grader forms a directed acyclic Figure, when one sample of arrival, from top to down finely divides sample, until reaching leaf node, the shape of image is entered Row Preliminary division, wherein, N is the classification number for dividing.

6. dark-red enameled pottery seal search method according to claim 5, it is characterised in that：In the step (301), between sampling It is divided intoWherein, L is profile length.

7. dark-red enameled pottery seal search method according to claim 5, it is characterised in that：In the step (304), SVM point Class device uses RBF kernel functions, is defined as the monotonic function of the Euclidean distance of space midpoint x to y points, and y is the center of kernel function：

$K(x,y)=\exp \left(\frac{-\left|\right|x-y|{|}^2}{2{\mathrm{\σ}}^2}\right)$

Wherein, the influence factor of RBF cores SVM performances is punishment parameter C and nuclear parameter σ in space²；Selected using grid data service The parameter：It is rightN and M value are taken respectively, and the wherein scope of C is set to [2^-10,2⁷],Scope is set to [2^-10,2³], step pitch is 0.1；Combine all of N × M value and calculate SVM Generalization Abilities, will wherein cause SVM to promote best performance Combination alternatively parameter.

8. dark-red enameled pottery seal search method according to claim 1, it is characterised in that：Improved BP- in the step (4) SIFT algorithms include introducing the constraint of two space of points distance ratios, and the cluster closed on before point search is reducing time complexity.

9. the dark-red enameled pottery seal search method according to claim 1 or 8, it is characterised in that：It is improved in the step (4) The concrete grammar of BP-SIFT algorithms, comprises the following steps：

(401) using classics SIFT algorithms to image contract SIFT feature to be retrieved, 128 are formed and ties up SIFT feature description；

(402) all characteristic points in template image are gathered for C classes according to coordinate position using KMEANS clusters；For template image In all characteristic points, calculate which class it belongs to, and nearest K point of proximity is found out in the apoplexy due to endogenous wind；

(403) it is a constant to initialize all of confidence level, if n=1；To current signature point and figure to be retrieved in template image The combination of the current signature point as in is right, iteratively updates its matching confidence level；

(404) constrained by the geometric distance of characteristic point and its K point of proximity and the Euclidean distance between two width image characteristic points Linear function combinational estimation calculate matching probability；

(405) n ← n+1, then goes to (404th) step, until matching probability no longer changes or iterationses n reaches maximum；

(406) if matching degree is less than predetermined threshold value, the two Point matchings success；Otherwise it is not that matching is right；

(407) all template images are repeated with above step, the matching rate of image to be retrieved and each template image is calculated, It is classified as the class of matching degree highest one.