CN1025764C - Characters recognition method and system - Google Patents

Abstract

The present invention relates to a character recognition method and a system. Stroke characteristics of a character picture are extracted, and characters are classified, matched and identified by directly making use of the stroke characteristics. Structure meaning of the characters is expressed by adopting a framework form, important influence strokes and stroke connections in a framework are emphasized, strokes with less action is omitted, and a necessary comparison condition for stroke direction of variable permission and similar character discrimination are presented. The present invention is favorable for protruding distinction between the characters, and processes of matching and identification are simplified. Compared with the present universal character identification technology, the present invention has high identification rate and adaptability.

Description

The invention relates to a character recognition method and system, and is especially suitable for recognizing handwritten Chinese characters and multi-font printed Chinese characters.

Several character recognition systems that have been developed at home and abroad mainly use the character recognition method of extracting characteristic parameters from the pixel distribution of character images, and classifying and matching recognition based on these parameters. For example, the character recognition system of CN1003257B on February 8, 1989, and the technology disclosed in CN1010512B on November 21, 1990.

Therefore, conventional techniques have the following problems:

1. It cannot directly reflect the structural features of characters, thus ignoring the stroke structure as the essential feature of character formation.

2. It is difficult to achieve a high recognition rate in the case of a large character set.

3. It is very difficult to distinguish characters with similar shapes or complex stroke structures.

4. In the case of handwritten characters, the writing of glyphs varies greatly, and the extracted feature parameters are highly dispersed, and high-dimensional feature vectors are required.

The purpose of the present invention is to create a character recognition method, strive to accurately extract the stroke features of the character image, fully reflect the structural nature of the character; directly use the stroke structure and meaning of the character to classify and match the character; use knowledge to express the structure of the character Word meaning, to simplify the matching and recognition process of characters, improve the accuracy of recognizing similar characters and the adaptability of recognition methods.

The character recognition method involved in the present invention includes: scanning the page with characters written on it to obtain a character image is the first step; character image binarization, character segmentation and normalization are the second step Steps: extracting the stroke structure feature of character binarized lattice is the third step; obtaining the classification feature code from the structure feature to determine the category to which it belongs is the fourth step; matching the structure feature with the character model of the category to which it belongs and identifying it as The fifth step; converting the recognition result into visible output is the sixth step.

Said third step comprises:

1. The character structure pattern as a whole can be decomposed into three sub-patterns: metacharacter, stroke and stroke meta. Metacharacters are the characters that construct characters. Strokes are decomposed into straight line segments which are stroke elements. The stroke unit is the lowest sub-pattern, used as a structural primitive to describe the character mode, and its structural features include stroke unit center coordinates, length, direction and connection relationship.

2. Do a simple scan of the character lattice, detect the connection between each pixel and the adjacent pixel in 8 directions, distinguish it as the beginning, end, connection area or ordinary stroke elements of the stroke and mark the corresponding The symbol, thus converting the character dot matrix plane (CDP) into the character pixel attribute plane (CAP).

3. Except for the pixels belonging to the connection area, calculate the number of continuous pixels en in the four directions of "|", "-", "/", and "\" for the pixels at the edge points on the CAP, The direction in which en is greatest is taken as the main fiber direction of the edge point. The value of en in the main direction is called the fiber length, and the pixels connected on the fiber length are given the corresponding weight of the main direction. The fibers at each edge point may intersect to form an interweaving area, and the direction weights of the pixels in the interweaving area are accumulated. After all the edge points complete the above calculations, the character fiber structure map (CFP) can be obtained.

4. Compared with the directional characteristics of the CAP connection area, the noise fibers in the CFP were removed, and the fibers belonging to the four directions of "|", "-", "/", and "\" were placed in V, h, s, and b respectively. In a plane, the center coordinates, length and direction of each stroke element can be obtained.

5. The connection relationship of stroke elements can be calculated by using the features of endpoints and connection areas of CAP, combined with the obtained center coordinates, length and direction of stroke elements.

Said fourth step comprises:

1. Using the four-corner and four-edge features of the character's peripheral structure as the classification feature of the character, the description and classification of the peripheral structure are carried out on two levels. A pre-classification dictionary is established from the four-corner features and four-edge features of known characters.

2. On the stroke plane (CSP) of the character, take the four corners of the plane as the center, and search for the stroke element closest to the four corners.

3. Determine the stroke element direction attribute of the nearest corner point, and divide it into six types: horizontal, vertical, left, right, corner, and cross, and assign corresponding codes, which are called corner codes. The code string composed of four corner codes constitutes the first classification feature of the character.

4. On the CSP, the ray is drawn from the center and scanned clockwise to obtain the polygon composed of the ray and the outermost stroke element of the character as the outline of the character, extract the convex points exceeding a certain threshold, and count the number of convex points on each side Obtain the second classification feature of the characters formed by the code strings of the four sides.

5. Find the same kind of character codes in the pre-classification dictionary as the four-corner code and the four-side code of the character to be recognized, and complete the fourth step.

The fifth step described:

1. The word meaning of the character structure adopts the knowledge expression in the form of a frame, and each character pattern is expressed by the character frame. In the framework, all the stroke elements constituting a character are grouped and sorted on the four planes of h, v, s, and b respectively, and the necessary stroke connection relations and the discrimination conditions of stroke element features between similar characters are listed. Each stroke element that participates in the grouping sort in the character frame is described by the stroke element frame. The stroke element frame expresses the normal direction, center position and length of the stroke element. Additionally, the stroke's weight and allowed distortion directions are given. The necessary connections in the character frame and the weights in the stroke element frame belong to the use of knowledge expression, emphasizing the stroke elements and their connections that have an important impact on the recognition results and ignoring those redundant or insignificant components. The similar character discrimination conditions and the allowed distortion direction make the recognition process take into account the It can recognize the nuances of the stroke structure between different characters in a complex and large number of character sets, and has good adaptability to the ever-changing glyphs.

2. Take out the pre-classified character models of the same type, search and match with the stroke meta-features of the characters to be recognized in turn, and calculate the attribute distance. If the distance is less than a certain threshold, the match is considered successful, otherwise the match is considered failed. This process is executed sequentially on the four stroke element sub-planes of each model until the end.

3. Calculate the weighted distance of the stroke meta attribute according to the weight specified by the stroke meta frame. The stroke elements that play a key role in the character structure have the highest weight, which is convenient for distinguishing the subtle differences of strokes between characters, and the stroke elements that have little influence have smaller weights, so as to achieve the purpose of ignoring redundant strokes.

4. If there is a sample sub-plane in the unmatched stroke element that allows the distortion direction and turns to the corresponding direction, search and match.

5. Check the necessary connection relationship, and exit the matching candidate column if this requirement is not met.

6. Detect stroke element comparison and similar character discrimination conditions, and exit the matching candidate list if the requirements are not met.

7. Match all characters whose total distance is within the threshold range, sort by distance from small to large, and take out the smallest few as recognition candidates. If there is no recognition candidate, it will be rejected.

The unique advantages that the present invention has can be summarized as follows:

Accurately extract stroke structure features to fully reflect the essential characteristics of characters. Directly use stroke features to describe the structural skeleton of characters, and use stroke attribute vectors to adapt to various changes in character shapes, and realize character classification and matching recognition. The knowledge expression in the form of a frame is used for the structure and semantic model of characters, which not only facilitates emphasizing important strokes or stroke connection relationships, but also ignores strokes that have little influence on character recognition, which is very conducive to highlighting the differences between characters and simplifying the matching and recognition process. The analysis conditions of similar characters are expressed in the framework, so that the subtle differences between characters can be identified Stroke differences become possible, for example: Feng, Feng; Shi, Tu; Lan, Qiao..., thus greatly improving the recognition rate of characters. The direction of allowable distortion is also given in the stroke frame, so that the flexibility and adaptability of recognition are significantly improved. Compared with the existing technology, it avoids the difficulty of feature selection and pattern separability in the statistical method due to the use of high-dimensional features, which limits the improvement of the recognition rate. It also avoids the defect that the structural method is difficult to adapt to the changeable character form.

The embodiment of the present invention is made up of graphic scanner, microcomputer mainframe, display, printer, magnetic tape drive and relevant interface board. Various types of scanners are available, including handheld scanners. The DOS operating system is the most common for microcomputer mainframes. If the tape drive is not necessary, it can be freely selected as the expansion or backup of the host memory. The working principle of the system is explained step by step with the following drawings.

Description of drawings:

Fig. 1 is the block diagram of the embodiment of the present invention

Figure 2 is a workflow diagram of structural feature extraction

Figure 3 is an example of structural feature extraction

Figure 4 is a description of the stroke element connection relationship

Figure 5 is a flow chart of the pre-classification workflow

Figure 6 is a four-corner feature code table

Figure 7 is the character frame

Figure 8 is a structural diagram of stroke element conditional sorting

Fig. 9 is a flow chart of stroke element conditional sorting

Figure 10 is the stroke meta frame

Figure 11 is a flow chart of matching recognition using knowledge guidance

Fig. 12 is a flow chart of sub-plane h stroke element matching.

Fig. 1 is the system block diagram of the embodiment, the characters written on the paper scan the page with a graphic scanner, and each page scans to obtain an image file, which is converted into binary by the selected gray scale threshold. Value (0, 1) dot matrix, stored in the computer through the interface board. The page segmentation program module searches for the starting line of the dot matrix, the total number of lines, the prefix and the number of words to automatically complete the word segmentation, and obtain the dot matrix of each character after normalization processing (for example, 32×32 or 64×64 characters Dot matrix), extract the stroke features of each character dot matrix, classify, match and then recognize the character until the character dot matrix stored in the machine is fully recognized, and the recognition result is represented by the machine internal code. Finally, display or print out the recognition results of all the characters written on the sample in standard fonts, or continue to make necessary edits.

Figure 2 is a flow chart of structural feature extraction. The normalized character dot matrix (CDP) is used as the starting point of the process, and the rows and columns of the CDP are scanned to detect continuous images with a value of 1 in the two directions of the row and column. As for the number X, record the most frequently occurring X as the stroke width wi, and when the continuous pixel elements measured by the stroke width in the row and column directions are less than wi, mark the pixel with "|" and "-" respectively. Check whether it is 0 on both sides of the "-" pixel. If the left side is 0, it belongs to the left endpoint, and it is marked as "W". If the 0 on the right side belongs to the right endpoint, it is marked as "E". Check whether it is 0 at the upper and lower sides of the "|" pixel. If the upper part is 0, it belongs to the upper endpoint and is marked as "N", and if the lower part is 0, it belongs to the lower endpoint and is marked as "S". All pixels that are neither "-" nor "|" in CDP are marked with lowercase English characters according to the coordinate order of their area. The area marked by the English characters is the connection area of the stroke, and the features of the connection area are calculated. Each pixel of the CDP is marked by the specified symbol according to the above requirements, and then endowed with different attributes such as the beginning, end, connection area or common pixel of the stroke, which is called the character pixel attribute plane (CAP). Figure 3 shows the writing character " Structural feature extraction example of "" character. Among them, the upper left is the CAP diagram, and the lower part is the connection area feature table. The first column is the serial number, the second column is the connection area code, the third and fourth columns are the starting and ending column coordinates, and the fifth and sixth columns are respectively The abscissa of the start and end. The last column is the connection feature of the connection area, and the connection feature is shown in Figure 4 in code. For each edge point of the CAP, except for the pixels in the connection area, calculate the number of continuous non-zero pixels in the four directions of row, column, left oblique, and right oblique, and choose the direction with the largest number of pixels. To the main direction of the fiber as the edge point, the number of pixels connected in the main direction is the length of the fiber, and each pixel is assigned the corresponding weight of the main direction. The fibers at each edge point may intersect to form an interweaving area, and the direction weights of the pixels in the interweaving area are accumulated. After all the edge points complete the above calculations, the character fiber structure map (CFP) is obtained. Remove the noise fibers in the interweaving area, and place the fibers belonging to the four directions of row, column, left oblique, and right oblique respectively in the four planes h, v, s, and b to obtain the center coordinates and length of each stroke element and direction, and then use the endpoints and connection features of CAP to obtain the connection relationship of stroke elements, so as to obtain all the structural features of characters. The upper right figure of Fig. 3 has shown the example of the structural feature of " adjacent " character.

Figure 5 is a flow chart of the pre-classification workflow. On the stroke plane of the character, with the four corners of the plane as the center, search for the stroke element closest to the four corners. Judge the direction attribute of this stroke element, divide them into horizontal, vertical, left, right, angle, intersection and empty seven types. Their coding is called angle code as shown in Figure 6, and the code string that is made up of four angle codes constitutes the first classification characteristic of character. Draw a ray from the center on the stroke plane of the character, and scan clockwise to obtain the polygon composed of the ray and the outermost stroke element of the character as the outer contour of the character, extract the convex points exceeding a certain threshold, and calculate the The number of salient points is used as the side code, and the four side codes form a four-side code string, which is the second classification feature of the character. Look up the pre-classified dictionary from the four-sided code and the four-corner code to obtain the same character code.

Fig. 7 is a framework for expressing the word meaning model of character structure, wherein the subscript ε _i represents the i-th stroke element, which is grouped and sorted on the four sub-faces of h, v, s, and b respectively. Fig. 8 is a conditional sorting structure diagram of stroke elements, Refer to Figure 9 for sorting conditions. The necessary connection relationship Ω _mn refers to the connection relationship that must be satisfied between the mth stroke element and the nth stroke element of the character, for example: "Fu", the first horizontal stroke and the vertical stroke must be intersecting , but God has no such requirement. The stroke element comparison notch is used to distinguish the difference in stroke length or direction within the character, for example: Earth, Shi; Tian, Yao, and the similar character discrimination condition is to judge the transfer direction of the candidate character when a certain stroke element is missing or exists. , For example: wind, phoenix; beam, beam and so on. Fig. 10 is a stroke element framework expressing the structural features of each stroke element ε _i in Fig. 7 . Quantization of the normal direction including the stroke element is horizontal, vertical, left and right, and the four directions are represented by h, v, s, b respectively; the central coordinates of the stroke element ( _xo , _yo ) _i and the length of the stroke. The framework also gives the direction ε ^′ _i of the allowable distortion of the stroke and the structural weight wi. The former makes the matching process flexible and improves the system's ability to adapt to font changes, while the latter emphasizes the emphasis and simplifies the matching. Figure 7 and Figure 10 constitute the structural model of the system. Fig. 11 shows the flow chart of matching identification using knowledge guidance. Fig. 12 is a flow chart of a certain sub-plane stroke element matching. According to the similar character codes given by the pre-classification, the corresponding character models are taken out one by one from the knowledge base, and the conditional sorting program module shown in Figure 9 sorts the obtained stroke elements. On the four sub-planes of h, v, s, b, the stroke elements in the model and the stroke elements of the character to be recognized are first matched within the group and then between the groups. When the attribute distance between the strokes is less than the specified threshold δ, the matching is considered successful. Otherwise, the match is considered a failure. If the matching fails, search down whether the character stroke element matches, if there is no match possible, take the next model stroke element to match. This process continues until the last model stroke element. If all the model stroke elements are successfully matched or the character stroke elements are matched, the attribute distance of all strokes is calculated according to the specified weight. If the distance is within the threshold △ range, it can be considered as a matching candidate. If there is an allowable distortion direction in the unmatched stroke element in the model, turn to the sample sub-plane search matching in the corresponding direction, and the method is the same. For the character models included in the matching candidates, further check whether the characters to be recognized meet the specified connection relationship Ω _mn , for example: husband, sky; force, knife;..., the intersecting relationship of husband and force is necessary, and this requirement is not met Exit the matching candidate column. If there is a requirement for stroke element comparison in the model framework, check whether the requirement is met, and the exit candidate column that does not meet the comparison condition. Repeat the above matching comparison until all classification models are matched. All characters whose matching total distance is within the threshold range are arranged in ascending order of distance as recognition candidates. The first candidate is the first candidate, which is usually taken as the recognition result. If there is no candidate for recognition, it will be treated as rejection.

Claims (1)

1. A character recognition method, the first step is to scan the page with characters written on it to obtain the character image; the second step is to convert the character image into binarization, character segmentation and normalization: to extract the character binarization lattice The third step is the stroke structural feature; the fourth step is to obtain the classification feature code from the structural feature to determine the category to which it belongs; the fifth step is to match the structural feature with the character model of the category to which it belongs; the fifth step is to make the recognition result visible Output is the sixth step, and the feature of the present invention is: Said third step comprises: (1) The character structure pattern as a whole can be decomposed into three sub-patterns: meta-character, stroke and stroke-meta. A metacharacter is a character for constructing a character, and a stroke is decomposed into a straight line segment, which is a stroke element. The stroke unit is the lowest sub-pattern, used as a structural primitive to describe the character mode, and its structural features include stroke unit center coordinates, length, direction and connection relationship. (2) Do a simple scan on the character lattice, detect the connection between each pixel and the adjacent pixel in 8 directions, and distinguish it as the beginning, end, connection area or common stroke elements of the stroke and mark them Corresponding symbols, thereby converting the character dot matrix plane (CDP) into the character pixel attribute plane (CAP). (3) Except for the pixels belonging to the connection area, for the pixels at the edge points on the CAP, calculate the number of continuous pixels en in the four directions of "|", "一", "/", and "\" , the direction with the largest en is taken as the main fiber direction of the edge point. The value of en in the main direction is called the fiber length, and the pixels connected on the fiber length are given the corresponding weight of the main direction. The fibers at each edge point may intersect to form an interweaving area, and the direction weights of the pixels in the interweaving area are accumulated. After all the edge points complete the above calculations, the character fiber structure map (CFP) can be obtained. (4) Compared with the directional characteristics of the CAP connection area, the noise fibers in the CFP were removed, and the fibers belonging to the four directions of "|", "一", "/", and "\" were placed in V, h, s, and b respectively. In the four planes, the center coordinates, length and direction of each stroke element can be obtained. (5) Using the features of endpoints and connection areas of CAP, combined with the obtained center coordinates, length and direction of stroke elements, the connection relationship of stroke elements can be calculated. Said fourth step comprises: (1) Using the four-corner feature and four-edge feature of the character's peripheral structure as the classification feature of the character, the description and classification of the peripheral structure are carried out on two levels. A pre-classification dictionary is established from the four-corner features and four-edge features of known characters. (2) On the stroke plane (CSP) of the character, with the four corners of the plane as the center, search for the stroke element closest to the four corners. (3) Judge the stroke element direction attribute of the nearest corner point, and divide into six types of horizontal, vertical, left, right, corner, and cross, and give corresponding codes, which are called corner codes. The code string composed of four corner codes constitutes the first classification feature of the character. (4) On the CSP, the ray is drawn from the center and scanned clockwise to obtain the polygon composed of the ray and the outermost stroke element of the character as the outer contour of the character, extract the convex points exceeding a certain threshold, and count the convex points on each side Points are obtained to obtain the second classification feature of the code strings on the four sides. (5) Find the same kind of character codes in the pre-classification dictionary as the four-corner code and the four-side code of the character to be recognized, and complete the fourth step. The fifth step described: (1) The word meaning of the character structure adopts the knowledge expression in the form of a frame, and each character pattern is expressed by the character frame. In the framework, all the stroke elements constituting a character are grouped and sorted on the four planes of h, v, s, and b respectively, and the necessary stroke connection relations and the discrimination conditions of stroke element features between similar characters are listed. Each stroke element that participates in the grouping sort in the character frame is described by the stroke element frame. The stroke element frame expresses the normal direction, center position and length of the stroke element. Additionally, the stroke's weight and allowed distortion directions are given. The necessary connections in the character frame and the weights in the stroke element frame belong to the use of knowledge expression, emphasizing the stroke elements and their connections that have an important impact on the recognition results and ignoring those redundant or insignificant components. The similar character discrimination conditions and the allowed distortion direction enable the recognition process to take into account the subtle differences in the stroke structure of different characters in a complex and large character set, and to have good adaptability to the ever-changing glyphs. (2) Take out the pre-classified character models of the same kind, search and match with the stroke meta-features of the character to be recognized in turn, and calculate the attribute distance. If the distance is less than a certain threshold, the match is considered successful, otherwise the match is considered failed. This process is executed sequentially on the four stroke element sub-planes of each model until the end. (3) Calculate the weighted distance of the stroke meta attribute according to the weight specified by the stroke meta frame. The stroke elements that play a key role in the character structure have the highest weight, which is convenient for distinguishing the subtle differences of strokes between characters, and the stroke elements that have little influence have smaller weights, so as to achieve the purpose of ignoring redundant strokes. (4) If there is a sample sub-plane in the unmatched stroke element that allows the distortion direction and turns to the corresponding direction, search and match. (5) Detect the necessary connection relationship, and exit the matching candidate column if this requirement is not met. (6) Detect stroke element comparison and similar character discrimination conditions, and exit the matching candidate list if the requirements are not met. (7) Match all characters whose total distance is within the threshold range, sort according to the distance from small to large, and take out the smallest few as recognition candidates, if there is no recognition candidate, it will be rejected.

Images