CN106023191A - Optical drawing character edge extraction and edge fitting method based on structure features - Google Patents

Abstract

The invention discloses a method for edge extraction and edge fitting of optically carved characters based on structural features. Marking, using the K nearest neighbor algorithm to remove the false edge of the sample according to the marked feature vector; constructing different connection modes according to the stroke structure characteristics of the character and the distance between the endpoints of the intermittent edge line segment to form a pixel scene; according to each pixel scene The positional relationship of each edge line segment, the gray information of the line segment, and the distance between the line segments determine the connection mode of the end points of the line segment, and perform edge fitting to form a character outline. The edge extracted by the invention is accurate, and the profile of the character obtained by fitting is complete, which brings great convenience to the subsequent character feature extraction.

Description

A Method of Edge Extraction and Edge Fitting Based on Structural Features of Optical Scribing Characters

technical field

The invention relates to a method for edge extraction and edge fitting of optically carved characters based on structural features.

Background technique

Character edge extraction and its fitting can constitute the complete outline of the character, which is an important link in the character recognition process. For optically carved characters, the commonly used edge extraction methods and fitting methods cannot get good results. In the process of generating the character image, the strip light source is used so that the strokes parallel to the light source will produce pixels with high gray value, while the strokes perpendicular to the light source will produce pixels with low gray value, and the pixel value of the background is between Between the above two types of pixels, as shown in FIG. 1 . Therefore, the Canny operator cannot extract the real edge of the character, because the Canny operator is essentially a gradient-based edge extraction method. In the middle of a stroke, two pixels with low gray value and high gray value are generated due to different lighting directions, and a large gray value changes at the junction, which can be captured by the gradient non-maximum suppression algorithm of the Canny operator. Considering these changes, the pixels at the corresponding positions are regarded as edges, but they are not the real edge points that characterize characters.

Even if some method is used to remove false edge points, discontinuous edges with relatively large intermittent intervals will be produced. If you want to use the commonly used edge fitting methods such as thresholding sequential edge connection contour extraction method, multi-threshold selection and edge connection edge detection algorithm and neural network edge fitting will not achieve good results, because these methods are only It is suitable for situations where the distance between discontinuous points is relatively short.

Contents of the invention

In order to solve the above-mentioned problems, the present invention proposes a method for edge extraction and edge fitting of optically carved characters based on structural features. The method first uses the Canny operator to extract edges, and extracts pixel features based on grayscale information and structural information. The nearest neighbor method eliminates the false edge points generated by the Canny operator, and then realizes the fitting of intermittent edges based on the structural characteristics of the characters to form the outline of the character, and the formed outline of the character is complete and accurate.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for edge extraction and edge fitting of optically carved characters based on structural features, comprising the following steps:

(1) Use the Canny operator to extract the edge, convert the pixel gray matrix corresponding to the true and false edge template into a feature vector, and mark it, and use the K nearest neighbor algorithm to remove the false edge of the sample according to the marked feature vector;

(2) Construct different connection patterns according to the stroke structure characteristics of characters and the distance between the endpoints of intermittent edge segments to form pixel scenes;

(3) According to the positional relationship of each edge line segment in each pixel scene, the gray information of the line segment and the distance between the line segments, determine the connection mode of the end points of the line segment, perform edge fitting, and form a character outline.

In described step (1), concrete steps include:

(1-1) Convert the pixel grayscale matrix corresponding to the true and false edge template into a feature vector, and assign corresponding classification labels at the same time;

(1-2) Select the k closest samples in the feature space according to the Euclidean distance;

(1-3) count the number of occurrences of each classification label in the K-nearest neighbor sample;

(1-4) Select the class label with the highest frequency of occurrence as the class label of the edge points to be classified.

Preferably, in the step (1-1), the pixel grayscale matrix corresponding to the true and false edge templates is converted into a feature vector with a length of 9.

In described step (2), multiple connection modes are summarized by comprehensively considering the distance between the stroke structure characteristics of characters and the endpoints of intermittent edge segments, and each connection mode is used as a pixel scene, and different connection modes take different Edge fitting method.

In the step (2), different pixel scenes are determined according to the position of the edge line segment to be fitted in the stroke, the arrangement structure between each other, the size relationship and the position relationship of the pixel gray value.

In described step (3), the end point of the line segment of each pixel scene is used as element, forms linked list, carries out edge fitting, starts from the first end point of the first edge line segment of linked list, determines each end point and next by linked list The distance from an endpoint and the number of line segments within the specified distance around the endpoint.

In described step (3), start to search in linked list whether there is the end point with the distance of first end point less than setting value, if exist, then two end points are connected, end point coordinate adopts polynomial fitting of first order, deletes linked list The connecting line segment of the two endpoints in , while the length of the linked list is reduced by 1.

In the described step (3), in linked list, search distance first endpoint less than the line segment of setting distance range, if there is above-mentioned line segment, then record the end point of line segment, count the number of line segments, according to the number of line segments, gray level line segment position Corresponding to the pixel scene, determine the pixel scene it belongs to.

In the step (3), the connection of the end points adopts the method of coordinate fitting.

Preferably, the set distance range is 2-3 times of the stroke width of the character.

The beneficial effects of the present invention are:

(1) The edge extracted by the present invention is accurate, and the character outline obtained by fitting is complete, which brings great convenience for subsequent character feature extraction.

(2) The present invention is based on the structural characteristics of characters to determine the way of intermittent edge line segments, which actually simulates human vision and cognitive principles; no matter which line segment the first line segment is, and no matter how much the distance between the end points of these line segments is , since they are all on one or two adjacent strokes of a character, correct connection can be realized.

Description of drawings

Fig. 1 is a schematic diagram of a character image of the present invention;

Fig. 2 is a schematic diagram of a set of false edge templates of the present invention;

Fig. 3 is a schematic diagram of a set of true edge templates of the present invention;

FIG. 4 is a schematic diagram of a set of scene templates of the present invention;

Fig. 5 is a schematic diagram of connection mode 3 of the present invention;

Figure 6 is the edge map extracted by the traditional method Canny operator, with false edge points;

Fig. 7 is a schematic diagram of the result of removing false edges in the present invention;

Fig. 8 is the fitting process and its result figure;

Fig. 9 is a flowchart of the present invention.

detailed description:

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

The edge extraction and edge fitting method of optically carved characters based on structural features includes the following steps:

Step 1, remove false edges based on the k-nearest neighbor algorithm.

Reasonable selection of feature space is the most critical element of correct classification, not only the accuracy of feature description, but also the balance of samples must be considered. The disadvantage of the K-nearest neighbor algorithm is that when the samples are unbalanced, it may cause that when a new sample is input, the samples of the large-capacity class among the k neighbors of the sample account for the majority.

By analyzing the gray distribution of surrounding pixels of edge points extracted by canny operator, it is found that the gray distribution of 8 neighborhoods of real edge points and false edge points is significantly different. False edge points are always at the junction of high grayscale pixels and low grayscale pixels in a stroke, and true edge points are either at the junction of high grayscale pixels and grayscale pixels in the background, or at the junction of low grayscale pixels and grayscale pixels in the background the junction. For this reason, the 8 neighborhoods of pixels are used as its features. For the convenience of description, the 8 neighborhoods of pixels are called edge templates, the 8 neighborhoods of true edge pixels are called true edge templates, and the 8 neighborhoods of false edge pixels are called false edges. Templates, true and false edge templates constitute the feature space using the k-nearest neighbor algorithm.

Since the pseudo-edge points are always at the junction of two gray-level pixel areas in a stroke, the pseudo-edge templates are divided into the following 8 categories according to the gray values of the pixels around the pseudo-edge points, with a total of 32 structures . The false edge points in the center of the first 16 templates have low gray values, and the false edge points in the latter 16 templates have high gray values, as shown in Figure 2.

The true edge point is either at the junction of the high-grayscale pixel area and the background, or at the junction of the low-grayscale area and the background. Therefore, the true edge templates are divided into the following six categories and 24 structures according to the gray values of the upper, lower, left, and right surrounding pixels of the edge point, as shown in Figure 3.

It should be noted that the values of the three gray levels in the template are represented by the gray mean values of the high, middle and low gray level pixels through histogram analysis.

Algorithm steps:

Convert the pixel grayscale matrix corresponding to the true and false edge template into a feature vector with a length of 9, and assign the corresponding classification label at the same time;

Select the k nearest samples in the feature space according to the Euclidean distance;

Count the number of occurrences of each classification label in the K-nearest neighbor sample;

Select the class label with the highest frequency as the class label of the edge point to be classified.

Step two, intermittent edge fitting based on structural features.

1. Characteristic description

Considering the stroke structure characteristics of characters and the distance between the endpoints of intermittent edge segments, 16 connection modes are summarized, involving 16 actual pixel scenes. Different connection modes adopt different edge fitting methods. Specifically, in different connection modes, the correct connection relationship of edge endpoints is selected according to three elements: the positional relationship of each edge line segment, the gray information of the line segment, and the distance between the line segments.

The 16 pixel scenes are shown in Fig. 4, and their stroke features are described as follows.

Scenario 1, the three edge segments to be fitted are at the end of a stroke, the two parallel edge segments have the same gray level, the pixel gray level of the other edge segment is higher than the former, and there are two other edges near the end of one edge , you need to choose one of them for edge fitting.

The four edge line segments to be fitted in scene 2 are in the middle of a stroke, and they are arranged symmetrically. The relative position of a pair of line segments with high pixel gray value is at the upper right, and the relative position of a pair of line segments with low pixel gray value is at the lower left square. There are three other edge endpoints near one edge endpoint, and one of them needs to be selected for edge fitting.

Scenes 3 to 5 are similar to scene 2, the difference is that in scene 3, the relative position of a pair of line segments with high pixel gray value is at the upper left, and the relative position of a pair of line segments with low pixel gray value is at the lower right; The relative position of a pair of line segments with high gray value is at the lower right, and the relative position of a pair of line segments with low pixel gray value is at the upper left; The relative position of a pair of line segments with a low value is at the upper right;

Scene 6 to Scene 8 are several situations derived from Scene 15 after edge fitting. Scenes 6 and 5 are similar, the difference is that the relative position of a pair of line segments with low pixel gray value is at the bottom right and the other is at the top right. Scenes 8 and 3 are similar, and the relative positions of a pair of line segments with high pixel gray values are in the upper left and the other in the upper right. In scene 7, the relative position of a pair of line segments with high pixel gray value is a left-right relationship, and the relative position of a pair of line segments with high pixel gray value is an up-down relationship.

Scenario 9 to Scenario 12 are several scenarios derived from subsequent scenarios 13 and 14 after edge fitting. Among the four edge line segments to be fitted, three line segments have low grayscale pixel values and the other one has high grayscale value. The three line segments with low pixel gray value are all vertical or nearly vertical line segments, in a regular arrangement of parallel or consistent up and down. A line segment with a high pixel gray value is a horizontal line segment. There are three other edge endpoints near one edge endpoint, and one of them needs to be selected for edge fitting.

Scene 13, the six edge line segments to be fitted are located at the junction of horizontal strokes and vertical strokes, and they are arranged symmetrically. The relative position of the pair of line segments with high pixel gray value is on the right side, and the two pairs of line segments with low pixel gray value The relative positions are upper left and lower left respectively. There are 5 other edge endpoints near one edge endpoint, and one of them needs to be selected for edge fitting. Scene 14 is similar to Scene 13, and the two are in a horizontal scene relationship.

Scene 15, the six edge line segments to be fitted are in a star-shaped symmetrical distribution relationship, and are divided into three groups of line segments according to the slope. A pair of line segments with a slope of zero have higher pixel gray value, and a pair of line segments with a negative slope A line segment has a lower pixel gray value, and another pair of line segments has one with a higher pixel gray value and the other with a lower pixel gray value.

In scene 16, the eight edge line segments to be fitted are symmetrically distributed in a star shape, and are divided into four pairs of line segments according to the slope and coordinate position, which are respectively located in the upper left, lower left, upper right, and lower right. There are 7 other edge endpoints near one edge endpoint, and one of them needs to be selected for edge fitting.

2. Specific description:

Use Ci(i=1,2....n) to represent any edge segment in the linked list, and Ci(1) and Ci(2) to represent its first end point and second end point. When fitting the edge, it always starts from C1(1) of the first edge segment in the linked list, and we care about the endpoint Ci(1) between C1(1) and other edge segments Ci(i=2....n). The distances D1i(1) and D1i(2) from the endpoint Ci(2). Edge fitting also cares about the number N of line segments within a certain distance range Dn around the endpoint C1(1), so as to determine which connection mode the current connection belongs to. Use the projection method to detect the width w of the stroke of the character. Observations and experiments have found that Dn=2.5W is appropriate.

First, search for endpoints Cj(k) where D1i(1) or D1i(2) is less than 0.3*w in {Ci(1), Ci(2)i=2,3...} (j=2,3. ..k=1,2), if Cj(k) exists, C1(1) is directly connected to Cj(k), the endpoint coordinates are fitted by a first-degree polynomial, Cj in the linked list is deleted, and the length of the linked list is reduced by 1. This is connection mode 1.

Secondly, search for line segments Cj whose min(D1i(1), D1i(2)) is less than 2.5*w in {Ci i=2,3...}, if Cj exists, record these endpoints Cj(k)k=1 ,2. According to the number of line segments, it can be divided into the following ways:

(1) If Number=1, and it is not the last intermittent edge line segment of the character outline, C1(1) is directly connected to Cj(k). This is connection mode 1.

(2) If Number=2, corresponding to scene 1, use grayscale and slope information to select a connection path. This is connection mode 2.

(3) If Number=3, corresponding to scene 2 to scene 12, the gray level information of the line segment and the positional relationship of the line segment are used to distinguish, and different endpoint connection methods are selected. Correspondingly referred to as connection modes 3, 6...13.

(4) If Number=4, 5, remove the line segment with the farthest distance, and process according to Number=3.

(5) If Number=6 corresponds to scenes 13, 14, and 15, use the grayscale information, slope information, and positional relationship of the line segments to distinguish, and select different connection methods. Called connection modes 14, 15, 16.

(6) If Number>7, it usually occurs when the length of the intermittent edge line segment is relatively short, and the intermittent edge line segment is relatively dense within the range of 2.5W, such as the middle part of the character "B". Wherein when Number=8, it can be judged according to the slope of the line segment that if it corresponds to scene 16 (two pairs of parallel line segments), it is connected according to the positional relationship of the line segment, which is called connection mode 16; otherwise, it is the same as other Number>7 situations. The same process as follows, reduce the search range, take Dn=1.5W, re-search the number N of line segments within the range of Dn around the endpoint C1(1), so that N<=6, and then determine the connection mode according to the above method.

3. Edge fitting method

In fact, scene recognition is used to determine the connection mode. The connection mode 3 is taken as an example below. When Number=3, there are ten scenes from scene 2 to scene 12 in total. There are 3 line segments around C1(1), and it is necessary to determine which end point of which line segment C1(1) is connected to the other three line segments. The eleven kinds of scenarios such as scene 2 to scene 12 generate eleven connection modes such as connection mode 3 to connection mode 13 .

The method of set decomposition is used for scene recognition, and the gray information, distance information and slope information of each line segment are needed for recognition. The line segment C1 increases with the length of the line segment connection, take 10 points from C1(1) to find its average coordinate XY1(x, y) and average gray level Bgray1; the average coordinates and average gray level of the other three line segments are respectively used by XY2 (x,y), XY3(x,y), XY4(x,y) and Bgray1, Bgray2, Bgray3, Bgray4 represent.

Eleven scenarios constitute set B. First, the set B is divided into subsets B1 and B2 according to the gray information of the three line segments. B1 contains a total of seven scenes from scene 2 to scene 8. Among the average gray levels of the four line segments Bgray1, Bgray2, Bgray3, and Bgray4, the pixel gray value of one pair of line segments is higher, while the pixel gray value of the other pair of line segments is lower. ; B2 includes scenes 9 to 12, one of the four line segments has a higher pixel gray value, and the other three line segments have a lower pixel gray value.

Second, divide B1 into subsets B11 and B12 according to the distance information of the set of line segments. Use the average coordinates XY1(x,y) to calculate the distance between two line segments with equal (near) gray levels, the distance between a pair of line segments with high gray value is represented by Hgd, and the distance between a pair of line segments with low gray value The distance is Lgd. B11 includes scenes 3, 4, 5, and 6. Both Hgd and Lgd are less than 1.5W; B12 includes scenes 6, 7, and 8. Only one of Hgd and Lgd is less than 1.5W.

Sets B11, B12, and B2 are the smallest sets that cannot be further divided. The four scenes of set B11 correspond to connection modes 3, 4, 5, 6; the three scenes of set B12 correspond to connection modes 7, 8, 9; the four scenes of set B2 correspond to connection modes 10, 11, 12, 13.

In the set B11, four scenarios are distinguished according to the relative positional relationship of the line segments. The row and column average values of a pair of line segment pairs with low gray value are represented by LL and LH, and the row and column average values of a pair of line segment pairs with high gray value are represented by HL and HH, if LL>HL and LH<HH It is scene 2, that is, connection mode 3; if LL>HL and LH>HH, it is scene 3, that is, connection mode 4; if LL<HL and LH<HH, it is scene 4, that is, connection mode 5; if LL<HL and LH>HH is scenario 5, that is, connection mode 6.

In the set B12, the slope information of the line segment is used to distinguish three scenarios. The slopes of the high gray line segment pairs are represented by HS1 and HS2, respectively, and the slopes of the low gray line segment pairs are represented by LS1 and LS2, respectively. If HS1=HS2, it is scene 6, that is, connection mode 7; if LS1=LS2, it is scene 8, that is, connection mode 9; otherwise, it is scene 7, that is, connection mode 8.

In set B2, it is necessary to exclude a pair of low-gray-scale line segments arranged in parallel from the left and right from the three low-gray-scale line segments according to the gray-scale information and column position information, and calculate the row average LL1 value and column Average value LH1, calculate the row average HL1 value and column average value HH1 of each pixel in the high gray line segment at the same time, use the position information LL1, LH1, HL1, HH1 to determine the positional relationship between the two to distinguish scene 9 to scene 12, that is Connection mode 10 to 13.

After determining the connection mode, it is necessary to further determine the specific position of each line segment in its scene in order to determine the specific connection relationship between the endpoints, so as to achieve edge fitting. Still take connection mode 3 as an example for illustration.

There are four line segments {C1 Ci Cj Ck}, ij k=2,3,4...N, the first line segment in the set is always C1. The grayscale sorting results in the above feature recognition are represented by {Ca Cb Cc Cd}, where Ca and Cb are a pair of line segments with a lower gray value of a pair of pixels, and Cc and Cd are a pair of line segments with a higher gray value of a pair of pixels. on the line segment.

Compare the column average values of Ca and Cb, assign the label ④ to the column average value that is small, and assign label ③ to the column average value that is large; compare the row average values of Cc and Cd, assign the label ① to the row average value that is small The one with the larger value is given the label ②. If the label of the line segment C1 is ①, select the line segment labeled ④ as the connecting line segment; if the label of the line segment C1 is ④, select the line segment labeled ① as the connecting line segment; The line segment of ③ is used as the connecting line segment, if the label of the line segment C1 is ③, the line segment labeled as ② is selected as the connecting line segment. The connection endpoint of the line segment C1 is always C1(1), and the connection endpoint of the connected line segment has been determined in the above feature recognition.

The method of coordinate fitting is used for the connection of the end points. If the distance between the end points is less than 0.8W, the first degree polynomial is used for fitting, otherwise the second degree polynomial is used for fitting.

Simulation results:

A typical character G is taken as an example for illustration. As shown in FIG. 6 , there are false edge points in the existing edge extracted simply by using the Canny operator. As shown in Figure 7, using the result of step 1 to remove false edges, combined with the fitting process and results of step 2 in Figure 8, it can be seen that the present invention can remove the problem of discontinuity well, and there is no false edge, and the identification The accuracy is high, and the contour of the character obtained by fitting is complete.

Although the specific implementation of the present invention has been described above in conjunction with the accompanying drawings, it does not limit the protection scope of the present invention. Those skilled in the art should understand that on the basis of the technical solution of the present invention, those skilled in the art do not need to pay creative work Various modifications or variations that can be made are still within the protection scope of the present invention.

Claims (10)

1. a kind of edge extraction and edge fitting method based on the optical scribe character of structural feature, it is characterized in that: comprise the following steps: (1) Use the Canny operator to extract the edge, convert the pixel gray matrix corresponding to the true and false edge template into a feature vector, and mark it, and use the K nearest neighbor algorithm to remove the false edge of the sample according to the marked feature vector; (2) Construct different connection patterns according to the stroke structure characteristics of characters and the distance between the endpoints of intermittent edge segments to form pixel scenes; (3) According to the positional relationship of each edge line segment in each pixel scene, the gray information of the line segment and the distance between the line segments, determine the connection mode of the end points of the line segment, perform edge fitting, and form a character outline. 2. a kind of optical scribe character edge extraction and edge fitting method based on structural feature as claimed in claim 1, is characterized in that: in described step (1), concrete steps comprise: (1-1) Convert the pixel grayscale matrix corresponding to the true and false edge template into a feature vector, and assign corresponding classification labels at the same time; (1-2) Select the k closest samples in the feature space according to the Euclidean distance; (1-3) count the number of occurrences of each classification label in the K-nearest neighbor sample; (1-4) Select the class label with the highest frequency of occurrence as the class label of the edge point to be classified. 3. a kind of optical character edge extraction and edge fitting method based on structural features as claimed in claim 2, is characterized in that: in described step (1-1), graying out the pixel corresponding to true and false edge template The degree matrix is converted into an eigenvector of length 9. 4. a kind of optical character edge extraction and edge fitting method based on structural feature as claimed in claim 1, is characterized in that: in described step (2), consider the stroke structural feature and intermittent edge of character comprehensively The distance between the endpoints of the line segment summarizes a variety of connection modes, and each connection mode is regarded as a pixel scene, and different connection modes adopt different edge fitting methods. 5. a kind of optical character edge extraction and edge fitting method based on structural features as claimed in claim 1, is characterized in that: in described step (2), be positioned at the position of stroke according to the edge segment to be fitted , the arrangement structure between each other, the size relationship and position relationship of the pixel gray value determine different pixel scenes. 6. a kind of optical scribe character edge extraction and edge fitting method based on structural feature as claimed in claim 1, is characterized in that: in described step (3), with the endpoint of the line segment of each pixel scene as element, Form a linked list and perform edge fitting, starting from the first end point of the first edge line segment of the linked list, determine the distance between each end point and the next end point and the number of line segments within the set distance range around the end point through the linked list. 7. a kind of optically carved character edge extraction and edge fitting method based on structural features as claimed in claim 1, is characterized in that: in described step (3), start to search in linked list whether to exist and first end The distance between the points is less than the endpoint of the set value, if it exists, connect the two endpoints, the endpoint coordinates are fitted by a polynomial, delete the connecting line segment between the two endpoints in the linked list, and the length of the linked list is reduced by 1. 8. a kind of optical character edge extraction and edge fitting method based on structural feature as claimed in claim 1, is characterized in that: in described step (3), search distance first end point in linked list less than setting For a line segment with a fixed distance range, if the above line segment exists, record the endpoint of the line segment, count the number of line segments, and determine the pixel scene it belongs to according to the number of line segments and the position of the gray level line segment corresponding to the pixel scene. 9. A method for extracting and edge fitting of optically carved characters based on structural features as claimed in claim 1, characterized in that: in the step (3), the connection of the endpoints adopts the method of coordinate fitting. 10. A method for edge extraction and edge fitting of optically carved characters based on structural features as claimed in claim 6, characterized in that: the set distance range is 2-3 times the stroke width of the character.