CN105938547A - Paper hydrologic yearbook digitalization method - Google Patents

Abstract

The invention relates to a paper hydrologic yearbook digitalization method. A feature fusion method of high complementarity is put forward on the basis of single feature so that the recognition rate is enhanced. The hydrologic process is similar due to the fact that the hydrologic process is influenced by similar seasonal climatic factors and other random factors, i.e. the flow has contextual relevance. In view of the correlation, a later error correction mechanism based on time sequences is also put forward. In other words, error correction processing is performed according to a certain criterion after classifier recognition. The experiment proves that the recognition accuracy can be effectively enhanced by the mechanism and the working efficiency can be guaranteed.

Description

A digital method for paper-based hydrological yearbook

technical field

The invention relates to a digitalization method for a paper-based hydrological yearbook, which belongs to the fields of computer image processing technology and hydrological intersection technology.

Background technique

The paper hydrological yearbook records the most basic hydrological test data, which contains information on the long-term evolution of nature and the impact of human activities, and plays an important role in production, scientific research, and social services. In view of the fact that hydrological yearbooks have been preserved for a long time, are frequently used, and are in poor storage conditions, paper hydrological yearbooks have gradually begun to be damaged, and once they suffer man-made or natural damage, irreparable losses will be brought about. Rescuing these precious historical materials has become an urgent problem. The most effective way to protect the hydrological yearbook is to digitally scan and process the hydrological yearbook to form an electronic file. Based on the above problems, the prior art studies the digitization of the hydrological yearbook, and proposes the intelligent recognition of the hydrological yearbook data. Recognizing the numbers in the hydrological data (ie, digital character recognition) is an important task for the digitization of the hydrological data.

Hydrological data is a kind of data published year by year, expressed in a unified and scientific graphic form. The content is mainly the basic hydrological data that are measured in the previous year and have been strictly compiled and reviewed. The table is characterized by the fact that the horizontal row indicates the specific month, and the vertical row indicates the date of each month. It consists of maximum flow, minimum flow, annual statistics and notes. Therefore, this paper analyzes the layout of the hydrological yearbook and extracts the table lines before identifying the numbers of the hydrological yearbook.

The digital characters of the hydrological yearbook are relatively standardized and the number of strokes is relatively small, which is relatively easier to extract than the Chinese character feature code. However, they have little shape change and too little stroke information, which in a sense increases the difficulty of effective feature vector extraction. For example, the numbers "8" and "6", when their ink is a bit heavier, sometimes the upper half of "6" in Bai Zheng Song typeface becomes a small circle, which is almost similar to "8". The numbers "1" and "3", "2" and "7", when the ink is heavy or the font is too small, it is likely that the numbers "1" and "3", "2" and "7" have the same feature vector. Therefore, in practical applications, using existing technologies to identify hydrological data has the disadvantages of low precision and low efficiency.

Contents of the invention

The technical problem to be solved by the present invention is to provide a digitalization method for paper-based hydrological yearbooks that adopts a new feature fusion design method, can effectively improve the recognition rate, and ensure work efficiency.

The present invention adopts the following technical solutions in order to solve the above-mentioned technical problems: the present invention designs a kind of digital method of paper hydrological yearbook, comprises the following steps:

Step 001. According to the layout design of the paper hydrological yearbook page, determine the pixel position where the hydrological data table is located in the paper hydrological yearbook page, and then enter step 002;

Step 002. According to the pixel position of the hydrological data table in the paper hydrological yearbook page, perform vertical and horizontal projections on the hydrological data table respectively, and analyze the longitudinal projection and horizontal projection of the hydrological data table respectively, and extract the hydrological data table respectively. The abscissa of each vertical line, the ordinate of each horizontal line, and then enter step 003;

Step 003. According to the layout of the hydrological data table, and the abscissa of each vertical line in the hydrological data table, and the vertical coordinate of each horizontal line, for the projected image of the hydrological data table, obtain the values in each numerical cell of the hydrological data table respectively. Data image, then enter step 004; Wherein, the numerical characters in each data image of the hydrological data form are white, and the background color is black;

Step 004. For each data image, perform character segmentation for each numerical character in the data image, obtain each numerical character block in the data image, and then obtain each numerical character block in each data image, and then enter step 005 ;

Step 005. For each numerical character block in each data image, extract the grid feature, Fourier feature, and contour moment feature of the numerical character in the numerical character block, and use them together as the identification feature of the numerical character, and then obtain each data respectively The recognition feature of numerical characters in each numerical character block in the image, then enter step 006;

Step 006. For each numerical character block in each data image, determine whether there is a preset number of black pixels from the top edge of the numerical character block downward, and if so, determine that the numerical character block is a decimal point, otherwise do not do anything Further operation; after completing the judgment for each numerical character block in each data image, then enter step 007;

Step 007. For all the identification features of numerical characters in all data images, perform feature fusion to form numerical identification features corresponding to "0" to "9" in the hydrological data table, and then enter step 008;

Step 008. According to the numerical identification features corresponding to "0" to "9" in the hydrological data table, and the identification features of the numerical characters in each numerical character block in each data image, respectively obtain the The number corresponding to each numerical character block, then enter step 009;

Step 009. According to the number or decimal point corresponding to each numerical character block in each data image, respectively form the numerical value corresponding to the data image in each numerical value cell of the hydrological data table, and then combine various attributes of the hydrological data table layout to obtain the hydrological data Each attribute in the table and its corresponding value are stored.

As a preferred technical solution of the present invention, after step 009, the following steps are also included, after step 009 is executed, enter step 010;

Step 010. For each attribute in the identified and stored hydrological data table and its corresponding value, respectively for the flow value of each month, execute according to the following steps 010-01 to 010-02, and then respectively obtain the flow rate for each month Preliminary identification and judgment of the daily flow value, and then enter step 011;

Step 010-01. Use the flow value on the first day of the current month as the first threshold, and then judge whether the difference between the next day’s flow value and the current day’s flow value is less than the first threshold for the flow values of the first two days of the current month, if yes, then Judging that the identification of the flow value of the current day is correct; otherwise, it is determined that the preliminary identification of the flow value of the current day is wrong; thereby obtaining the preliminary identification and judgment of the flow values of the first two days of the current month, and then proceed to step 010-02;

Step 010-02. For each daily flow value starting from the third day of the current month, determine whether the difference between the next day’s flow value and the current day’s flow value is smaller than the previous day’s flow value, and if so, determine that the current day’s flow value is correctly identified ; Otherwise, it is judged that the preliminary recognition of the flow value of the current day is wrong; thus, the preliminary recognition and judgment of the flow value of each day starting from the third day of the current month are obtained;

Step 011. According to each numerical value in the identified and stored hydrological data table, and the identification features of each number in each numerical value, through the preset trainer, obtain each number in each numerical value in the identified stored hydrological data table, respectively corresponding to "0 " to "9" ten recognition result probabilities, and then enter step 012;

Step 012. Obtain the maximum recognition result probability and the second largest recognition result probability among the ten recognition result probabilities from "0" to "9" corresponding to the numbers for each number in each numerical value in the identified and stored hydrological data table, And obtain the difference between the maximum recognition result probability and the second largest recognition result probability, judge whether the difference is less than the preset recognition result probability threshold, if so, judge that the initial recognition of the number is incorrect; otherwise, judge that the recognition of the number is correct; thus Obtain a preliminary identification judgment for each number in each numerical value in the identified stored hydrological data table, and then enter step 013;

Step 013. For each preliminary misidentified flow value in each month, judge whether there is a preliminary misidentified number in the preliminary misidentified flow value, if yes, judge that the preliminary misidentified flow value is wrong, and issue an alarm; otherwise, judge The preliminary identification error flow value is correct; thus, the verification of each value in the identified stored hydrological data table is realized.

As a preferred technical solution of the present invention: in the step 011, according to each numerical value in the identified stored hydrological data table and the identification features of each number in each numerical value, the identified stored hydrological data is obtained through the support vector machine trainer. Each number in each numerical value in the data table corresponds to ten recognition result probabilities from "0" to "9".

As a preferred technical solution of the present invention: in the step 013, according to the number of initially identified incorrect numbers in the initially identified incorrect flow value, it is judged that the initially identified incorrectly identified flow value is wrong, and at the same time as an alarm, according to The position of the preliminary identification error number in the preliminary identification error flow value is analyzed. If the preliminary identification error number is in the integer part of the preliminary identification error flow value, the previous day of the date corresponding to the preliminary identification error flow value is used. The average value of the daily flow value and the next day's flow value is used to replace the preliminary identification error flow value; if the preliminary identification error number is in the decimal part of the preliminary identification error flow value, the date corresponding to the preliminary identification error flow value is used The average value of the decimals of the traffic values of the previous day and the decimals of the traffic values of the next day is to replace the decimals in the preliminary identification error traffic values.

As a preferred technical solution of the present invention, the step 004, performing character segmentation for each numerical character in the data image, to obtain each numerical character block in the data image, specifically includes the following steps:

Step a01. Detect each white pixel point inside each numerical character in the obtained data image, and the minimum distance between each edge of the data image and the white pixel point on the corresponding numerical character, and then enter step a02;

Step a02. For the previous step, judge each white pixel point obtained in the data image separately, and judge whether the pixels at the upper, lower, left, and right positions of the pixel point are all white pixels, and if so, judge the pixel point is the pixel inside the numeric character; otherwise, it is judged according to the identifier that the pixel is the edge pixel of the character, and the column number of the pixel column where the pixel is located in the data image is obtained; thus for the previous step, the data image Each white pixel obtained in is judged respectively, obtains the column number of the pixel row in the data image where the upper edge pixel of each numerical character in the data image is located, and then enters step a03;

Step a03. According to the column number of the pixel column where the upper edge pixel point of each numerical character in the data image is located in the data image, divide each numerical character in the data image to obtain each numerical character block in the data image.

As a preferred technical solution of the present invention, in the step 005, for each numerical character block in each data image, the grid feature of the numerical character in the numerical character block is extracted, specifically comprising the following steps:

Step b01. Obtain the upper, lower, left, and right boundaries of the numerical character block, and thus obtain the numerical character ontology image, and then enter step b02;

Step b02. Perform center-of-gravity normalization on the numerical character body image, and divide the weight-normalized numerical character body image into a preset number of sub-region images on average, and then enter step b03;

Step b03. Obtain the proportion of white pixels in each sub-region image of the numerical character body image, and together form the grid features of the numerical character in the numerical character block.

As a preferred technical solution of the present invention, in the step 005, for each numerical character block in each data image, extract the Fourier feature of the numerical character in the numerical character block, specifically comprising the following steps:

Step c01. Carry out two-dimensional discrete Fourier transform for the numerical character block, and then enter step c02;

Step c02. Continue to perform central transformation on the numerical character block after two-dimensional discrete Fourier transform, that is, divide the numerical character block into four sub-region images on average, and perform diagonal exchange to obtain the Fourier image spectrum, and then Go to step c03;

Step c03. Analyze the Fourier coefficients of the Fourier image spectrum after the central transformation, and obtain the area where the Fourier coefficients of the numerical character block are concentrated, which are greater than the preset amplitude threshold, to form a large Fourier coefficient. Liye coefficient area, then enter step c04;

Step c04. Extract a preset number of discrete Fourier transform coefficients from the large-scale Fourier coefficient area, and normalize them to form Fourier features of the numerical characters in the numerical character block.

As a preferred technical solution of the present invention: in the step 005, for each numerical character block in each data image, extract the contour moment feature of the numerical character in the numerical character block, specifically comprising the following steps:

Step d01. Perform contour extraction for the numerical characters in the numerical character block, and then enter step d02;

Step d02. Perform moment invariant processing on the contours of the numerical characters in the numerical character block, and extract a preset number of two-dimensional contour invariant moment features to form the contour moment features of the numerical characters in the numerical character block.

As a preferred technical solution of the present invention, the step 007 specifically includes the following steps:

Step e01. According to the permutation and combination, for all the identification features of the numerical characters in all data images, perform any combination of two identification features to form all identification feature combinations, and then enter step e02;

Step e02. All the identification features of the numerical characters in all data images are used to form the sample set S corresponding to the numbers "0" to "9" in the hydrological data table, and then each group of identification features is combined according to the following formula (1):

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&cup;</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&cap;</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Obtain the feature complementary index C _ij,A of the group of recognition feature combinations relative to the standard number "0"-"9"respectively; and then obtain the feature complementary index C of each group of recognition feature combinations respectively relative to the standard number "0"-"9" _{ij, A} ; Then enter step e03; Wherein, S _i and S _j respectively represent the sample set that the sample set S is misclassified by the recognition feature F _i and the recognition feature F _j ; E (S) represents the number of samples in the sample set S ; E(S _i ∪ S _j ) represents the number of samples in the union between the sample set S _i and the sample set S _j ; E(S _i ∩ S _j ) represents the intersection between the sample set S _i and the sample set S _j The number of samples; A={0, 1, ..., 9}, C _{ij, A} represents the feature complementary index of the identification feature combination made of identification feature F _i and identification feature F _j relative to the standard number A;

Step e03. Respectively for each group of identification feature combinations, according to the following formula (2):

${\mathrm{<span\; class="notranslate">\; TC</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 9</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}}{{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; 10</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Obtain the overall complementary index TC _k of each group of identification feature combinations relative to the standard number, and then enter step e04; where, k={1,...,K}, K represents the number of combinations of all identification feature combinations, and TC _k represents the kth The overall complementarity index of group-identifying feature combinations relative to standard numbers;

Step e04. For all the identification feature combinations, sort according to their overall complementary index from large to small to obtain the first two identification feature combinations, and then carry out feature fusion for the two identification feature combinations to form the corresponding "0" in the hydrological data table. ” to “9” to identify the character.

As a preferred technical solution of the present invention, in the step 008, according to the numerical identification features respectively corresponding to "0" to "9" in the hydrological data table, and the identification features of numerical characters in each numerical character block in each data image , through a support vector machine (SVM) classifier, the numbers corresponding to each numerical character block in each data image are respectively obtained.

Compared with the prior art, a paper-based hydrological yearbook digitization method and control method according to the present invention have the following technical effects: the paper-based hydrological yearbook digitization method designed by the present invention proposes a new method based on a single feature. The feature fusion method with strong complementarity has improved the recognition rate. Since the hydrological process is affected by similar seasonal climate factors and other random factors, it shows similarity, that is, its flow has context correlation. Therefore, the present invention considers this At the same time, a later error correction mechanism based on time series is proposed. That is, after the classifier is recognized, it is corrected according to a certain criterion. It is proved by experiments that the mechanism proposed by the present invention effectively improves the recognition accuracy and ensures the work efficiency.

Description of drawings

Fig. 1 is the flow chart of the paper hydrology yearbook digitization method and control method designed by the present invention;

Fig. 2 a is the horizontal projection schematic diagram of hydrological data table in the embodiment;

Fig. 2b is a schematic diagram of the longitudinal projection of the hydrological data table in the embodiment;

Fig. 3 is the schematic diagram of the table formed by each vertical line and each horizontal line extracted respectively in the hydrological data table in the embodiment;

Fig. 4 is a schematic diagram of layout analysis of hydrological yearbook in the embodiment;

Fig. 5 is the schematic diagram that respectively obtains the data image in each numerical value cell of hydrological data table in the embodiment;

Fig. 6 is a schematic diagram of each numerical character block in the data image obtained in the embodiment.

detailed description

The specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings.

In our daily business activities, we use a lot of documents and forms every day. At the same time, form documents are also widely used in various fields. Usually, people need to manually process form documents. For example, customers need to pay taxes, and librarians need to collect data information contained in paper form documents. Due to the development of Optical Character Recognition (OCR) technology, people began to try to extract the data information in the form by using the standard form image of the available data, which can reduce the working time and reduce the workload. In the business field, using OCR technology can improve the quality of work and reduce the amount of time people spend on processing form documents. In many fields where OCR is used, we let the user know the target string printed in the image through the obtained form template. These string information includes many items such as traffic information, text information and mathematical formulas. The existence of tables hinders the extraction of data information, so table line detection is an important task in printed table recognition technology.

In printed documents of hydrological data, tables are an essential part, which can highly gather all document information together and allow readers to accurately understand the meaning of their expressions, which is both concise and standardized. By consulting the discharge tables of major hydrological stations in the Hydrological Yearbook, it can be found that the layout structure of the discharge table in the Hydrological Yearbook is regular. We can use these rules to cut out characters.

The hydrological yearbook is the carrier for the hydrological agency to monitor the water bodies of the rivers in the basin, and to process, arrange and publish the following year to form the carrier of the hydrological monitoring results. Its content includes various reorganization results and summary data explained with charts and necessary words. It is a treasure house of systematic and standardized hydrological data.

In 1958, the Hydrological Bureau of the Ministry of Water Resources divided the country into volumes of hydrological data according to the basin water system, and named the year-by-year data uniformly the "Hydrological Yearbook of the People's Republic of China", which is divided into 10 volumes and 94 volumes nationwide. Its characteristics are as follows.

Color characteristics: Black characters on a yellow background.

Structural features: the paper width is 440mm, the height is 140mm, and the aspect ratio is 3.14. The numbers in the yearbook are approximately 15mm wide, 24mm high, and have an aspect ratio of 0.625. The characters are inside the table.

Texture features: The yearbook contains character-like areas, that is, the horizontal and vertical color chromaticity of numbers presents regular changes in peaks and valleys.

The hydrological yearbook characters are multi-line horizontal and regular characters with relatively stable structure and texture characteristics. The projection-based top-down layout analysis method is the application of this feature. In the character area of the yearbook, the edge information of the characters is very rich. Using certain tools to detect and analyze the edge information of the characters can separate the hydrological data from the background. The pixel values in the hydrological yearbook area will show specific fluctuations, and the frequency of change will also be kept within a certain range. Using these features, the character location of the hydrological yearbook can be realized. According to the fact that the horizontal and vertical features of the yearbook digital area are more abundant than the non-digital area, a character positioning algorithm based on horizontal and vertical projection is proposed. Calculate its jump points, and determine possible character regions according to the number of jump points and the distance between jump points.

About 275 pixels away from the top margin of the page, the page is blank, followed by the name of the watershed and hydrological station in the hydrological yearbook plus the words daily average discharge table. Units of water catchment area and flow rate are marked about 30 pixels away from this word. About 20 pixels from this is where the table starts. The hydrological yearbook tables are composed of 11 horizontal lines and 14 vertical lines. The month information is marked in the middle of the first two horizontal lines, the date of each month is marked between the first two vertical lines, and the area between each two vertical lines and before the third horizontal line is the traffic value of each month . The average flow value of each month, the maximum flow value and the minimum flow value of the date, annual statistics and notes are marked between the subsequent horizontal lines. Our ultimate goal is to identify the flow value, so we must first analyze the layout of the hydrological data, analyze its table structure, and extract the table frame, so as to specifically locate the flow value of each month.

As shown in Figure 1, the present invention has designed a kind of digitalization method of papery hydrological yearbook, first will take a picture of the hydrological data table in the page of papery hydrological yearbook, obtain the image of hydrological data table, and carry out preprocessing operation, wherein include image two Value, gray scale, denoising, rotation and inversion processing; then for the hydrological data table image of the preprocessing operation, the specific steps are as follows:

Step 001. With the in-depth research on document layout analysis algorithms, this paper combines the advantages of two typical algorithms based on the original document layout segmentation algorithms (top-down and bottom-down), that is, using structural features at the same time and texture features to process document layouts in the Hydrological Yearbook. This processing method not only considers the accuracy of segmentation, but also takes into account the time consumption of analysis and processing, so it can quickly and accurately locate the table. According to the layout design of the page of the paper hydrological yearbook, determine the pixel position of the hydrological data table on the page of the paper hydrological yearbook, and then go to step 002.

Step 002. According to the pixel position of the hydrological data table in the paper hydrological yearbook page, vertical and horizontal projections are respectively performed on the hydrological data table. The horizontal projection is shown in Figure 2a, and the vertical projection is shown in Figure 2b. The vertical and horizontal projections were analyzed separately. In Figure 2a, the 11 black dots represent the horizontal lines of the hydrological yearbook table, and the hollow dots after the second black dot represent the upper and lower positions of the flow values in each row. The two sides of each peak represent the upper and lower positions of the flow values of each row from the first day to the 31st day; in Figure 2b, 14 black dots represent the abscissa of the 14 vertical lines in the table, and every two black dots The interval, that is, the left and right coordinates of the monthly flow value on both sides of the peak between every two vertical lines, is marked with a hollow point. Extract the abscissa of each vertical line and the ordinate of each horizontal line in the hydrological data table respectively, as shown in Figure 3 in the practical application embodiment, wherein, the numerical characters in each data image of the hydrological data table are white, and the background color It is black; therefore, the monthly discharge value and table position can be roughly located through Figure 2a and Figure 2b, and the final layout analysis results of the hydrological yearbook are shown in Figure 4, and then enter step 003.

By counting the number of black pixels on the same row or column, direct detection of straight line segments is avoided, the connectivity of table lines is not high, and it has good anti-interference and generalization capabilities. Effective information such as the position and size of the target in the image can be reflected by this method. It provides convenience for the positioning and processing of subsequent hydrological yearbook numbers.

Step 003. According to the layout of the hydrological data table, and the abscissa of each vertical line in the hydrological data table, and the vertical coordinate of each horizontal line, for the projected image of the hydrological data table, obtain the values in each numerical cell of the hydrological data table respectively. The data image, the actual application example is shown in Figure 5, and then enter step 004; wherein, the numerical characters in each data image of the hydrological data table are white, and the background color is black.

In the digitization process of paper water level data, only by segmenting the hydrological data images adaptively can the accuracy of the subsequent extracted feature data be guaranteed. The segmentation of paper water level data images is the basis of the entire digitization process. After digital positioning, the image is still a whole, including the blanks between numbers. For the whole number that has been extracted, character segmentation is required. Separate individual characters from the overall number.

Step 004. For each data image, perform character segmentation for each numerical character in the data image, and obtain each numerical character block in the data image, which specifically includes the following steps:

Step a01. Detect each white pixel point inside each numerical character in the obtained data image, and the minimum distance between each edge of the data image and the white pixel point on the corresponding numerical character, and then enter step a02;

Step a02. For the previous step, judge each white pixel point obtained in the data image separately, and judge whether the pixels at the upper, lower, left, and right positions of the pixel point are all white pixels, and if so, judge the pixel point is the pixel inside the numeric character; otherwise, it is judged according to the identifier that the pixel is the edge pixel of the character, and the column number of the pixel column where the pixel is located in the data image is obtained; thus for the previous step, the data image Each white pixel obtained in is judged respectively, obtains the column number of the pixel row in the data image where the upper edge pixel of each numerical character in the data image is located, and then enters step a03;

Step a03. According to the column number of the pixel column where the upper edge pixel point of each numerical character in the data image is located in the data image, divide each numerical character in the data image to obtain each numerical character block in the data image.

Based on the above-mentioned design process, each numerical character block in each data image is further respectively obtained. In the actual application embodiment, each numerical character block in the data image is obtained, as shown in FIG. 6 ; then go to step 005.

If the preprocessed data is directly used as the input of the classifier, the amount of data in the classification calculation is large, and the purpose of feature extraction is to start with the analysis of the topology of the number, extract some of its structural features, and make the displacement of the number , size change, font distortion and other interference are relatively reduced, that is, the key information that reflects the digital features is provided to the classifier, which is equivalent to indirectly increasing the fault tolerance of the classifier, and the amount of data is greatly reduced after feature extraction. Small; feature extraction plays a key role in recognition, and it should follow the following principles:

(1) easy to extract;

(2) It has a strong classification ability, that is, the feature should show a large difference for different numbers, and should show as small a difference as possible for the same number;

(3) It has high stability and minimizes the impact of stroke breakage or adhesion.

Step 005. For each numerical character block in each data image, extract the grid feature, Fourier feature, and contour moment feature of the numerical character in the numerical character block, and use them together as the identification feature of the numerical character, and then obtain each data respectively The recognition features of numerical characters in each numerical character block in the image, and then enter step 006.

Among them, the grid feature is a group of distribution features that focus on the overall character image, and this feature has a strong ability to suppress noise. The main idea of its extraction method is to divide the digital lattice into several local small areas, and use the lattice density on each small area as a description feature, that is, to count the percentage of image pixels in each small area as feature data; Since the grid feature reflects the local statistical characteristics of the image, it is a relative percentage value, and the digital lattice corresponding to the local deformation or noise of the image is the exchange of the values of "0" and "1" of the local elements, so if the image has Local deformation or noise, compared with the original image without deformation and noise, the relative value of the calculated percentage changes little. That is to say, this relative value is not sensitive to the deformation of local strokes in the digital picture or the influence of isolated noise points. Therefore, the digital recognition with the grid as the feature has better anti-noise ability. For the numbers segmented in this article, I divided them into small areas with a size of 3×3, a total of 9.

In the above step 005, for each numerical character block in each data image, extract the grid features of the numerical character in the numerical character block, specifically including the following steps:

Step b01. Obtain the upper, lower, left, and right boundaries of the numerical character block, and thus obtain the numerical character ontology image, and then enter step b02.

Step b02. Carry out barycenter normalization for the numerical character ontology image, and evenly divide the barycenter-normalized numerical character ontology image into a preset number of sub-region images, and then proceed to step b03.

Step b03. Obtain the proportion of white pixels in each sub-region image of the numerical character body image, and together form the grid features of the numerical character in the numerical character block.

Fourier transform is a two-dimensional orthogonal transform widely used in image processing. After Fourier transform, the average value, that is, the DC term, is proportional to the average value of the gray value of the image, and the low-frequency component indicates the strength and direction of the target edge in the image. Numerical characters can generally be represented by closed contours composed of many line segments, and some discrete quantities obtained through mapping can fully reflect changes in these closed contours. The Fourier coefficient can well describe the image boundary contour, and its value has nothing to do with the translation, rotation, displacement and size of similar glyphs. These features form an obvious data compression during character representation and recognition.

In the above step 005, for each numerical character block in each data image, extract the Fourier feature of the numerical character in the numerical character block, specifically including the following steps:

Step c01. Perform two-dimensional discrete Fourier transform on the numerical character block, and then enter step c02.

Step c02. Continue to perform central transformation on the numerical character block after two-dimensional discrete Fourier transform, that is, divide the numerical character block into four sub-region images on average, and perform diagonal exchange to obtain the Fourier image spectrum, and then Go to step c03.

Step c03. Analyze the Fourier coefficients of the Fourier image spectrum after the central transformation, and obtain the area where the Fourier coefficients of the numerical character block are concentrated, which are greater than the preset amplitude threshold, to form a large Fourier coefficient. Liye coefficient area, and then go to step c04.

Step c04. Extract a preset number of discrete Fourier transform coefficients from the large-scale Fourier coefficient area, and normalize them to form Fourier features of the numerical characters in the numerical character block.

The moment invariant feature is a statistical feature of the image, and it is a mathematical feature in the image that is invariant to translation, scaling, and rotation.

In the above step 005, for each numerical character block in each data image, the contour moment feature of the numerical character in the numerical character block is extracted, which specifically includes the following steps:

Step d01. Perform outline extraction for the numerical characters in the numerical character block, and then proceed to step d02.

Step d02. Perform moment invariant processing on the contours of the numerical characters in the numerical character block, and extract a preset number of two-dimensional contour invariant moment features to form the contour moment features of the numerical characters in the numerical character block.

Step 006. For each numerical character block in each data image, determine whether there is a preset number of black pixels from the top edge of the numerical character block downward, and if so, determine that the numerical character block is a decimal point, otherwise do not do anything Further operation; after completing the judgment for each numerical character block in each data image, then enter step 007;

All the recognition features obtained in the above steps, if they are classified by neural network and support vector machine classifier, the classification effect is not ideal, mainly because it is difficult to find a feature suitable for different numbers, while the predecessors The methods are all feature extraction and fusion in the analysis of specific digital recognition applications. Each number has different characteristics. In order to classify correctly, various features need to be combined. The complementarity of features is to ensure that the extracted features have higher recognition. The key to the rate and generalization ability is the basis of feature fusion; therefore, the problem of feature complementarity measurement must be solved before feature fusion.

Step 007. For all the recognition features of the numerical characters in all data images, perform feature fusion to form numerical recognition features corresponding to "0" to "9" in the hydrological data table, and then enter step 008.

The above step 007 specifically includes the following steps:

Step e01. According to the permutation and combination, for all the recognition features of the numerical characters in all the data images, combine any two recognition features to form all the recognition feature combinations, and then go to step e02.

Step e02. All the identification features of the numerical characters in all data images are used to form the sample set S corresponding to the numbers "0" to "9" in the hydrological data table, and then each group of identification features is combined according to the following formula (1):

${\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&cup;</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&cap;</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Obtain the feature complementary index C _ij,A of the group of recognition feature combinations relative to the standard number "0"-"9"respectively; and then obtain the feature complementary index C of each group of recognition feature combinations respectively relative to the standard number "0"-"9" _{ij, A} ; then enter step e03; wherein, C _ij, the larger the A, the stronger the feature complementarity of the identification feature F _i and the identification feature F _j relative to the standard number A; otherwise, the weaker the feature complementarity; S _i and S _j represent the sample set S misclassified by the recognition feature F _i and the recognition feature F _j respectively; E(S) represents the number of samples in the sample set S; E(S _i ∪ S _j ) represents the sample set S The number of samples in the union between _i and the sample set S _j ; E(S _i ∩ S _j ) represents the number of samples in the intersection between the sample set S _i and the sample set S _j ; A={0, 1,... , 9}, C _ij,A represents the feature complementarity index of the identification feature combination composed of identification feature F _i and identification feature F _j relative to the standard number A.

Step e03. Respectively for each group of identification feature combinations, according to the following formula (2):

${\mathrm{<span\; class="notranslate">\; TC</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; 9</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}}{{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; 10</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Obtain the overall complementary index TC _k of each group of identification feature combinations relative to the standard number, and then enter step e04; where, k={1,...,K}, K represents the number of combinations of all identification feature combinations, and TC _k represents the kth The overall complementarity index of groups identifying feature combinations relative to standard numbers.

Step e04. For all the identification feature combinations, sort according to their overall complementary index from large to small to obtain the first two identification feature combinations, and then carry out feature fusion for the two identification feature combinations to form the corresponding "0" in the hydrological data table. ” to “9” to identify the character.

The above technical solution uses different features for classification in the classifier, analyzes the recognition result of a single feature, calculates the overall complementary index of each feature through the above formula, and then fuses the selected features using a certain linear relationship. Experiments show that the coarse grid feature and Fourier feature are very effective in digital recognition of hydrological yearbook data, and their overall complementarity is strong, so the Fourier feature is connected in series after the coarse grid feature, and the proposed fusion feature is obtained through experiments The recognition rate of the method is 3.8981% higher than that of a single Fourier feature, 1.4033% higher than a grid feature, and 83.1956% higher than a contour moment.

Step 008. According to the numerical identification features respectively corresponding to "0" to "9" in the hydrological data table, and the identification features of numerical characters in each numerical character block in each data image, through a support vector machine (SVM) classifier, respectively obtain The number corresponding to each numeric character block in each data image, and then enter step 009.

Step 009. According to the number or decimal point corresponding to each numerical character block in each data image, respectively form the numerical value corresponding to the data image in each numerical value cell of the hydrological data table, and then combine various attributes of the hydrological data table layout to obtain the hydrological data The attributes in the table and their corresponding values are stored; then go to step 010.

This paper analyzes the law of traffic and proposes a post-debugging mechanism based on time series. It can be seen from the experimental results that the final recognition result of the Hydrological Yearbook is close to 99%, and the error rate is relatively low. A flow value is composed of 4 to 5 numbers. If one of the numbers is wrongly recognized, the result is considered to be wrong. This is the same as The error rate statistics of the recognition results on the previous data sets MNIST and USPS are still slightly different. Observing the recognition results, we can see that there is generally only one digital misrecognition error for a flow value, and the number of misidentified flow values per month is less than 3. In this case, if we can find flow values with low recognition reliability through certain algorithmic ideas, we can also That is to find the identification error of the key position of the number before the decimal point of the flow value, and use the average value method to correct the error by counting the change rule of the monthly flow, which will bring high application efficiency.

Because the flow rate itself is also obtained through instrument measurement, there are certain errors in itself. Therefore, if the flow rate fluctuates within a certain small range, that is, when the number after the decimal point of the flow value is incorrectly identified, in the As long as it does not affect the analysis and application of traffic data, we can tolerate it. That is, it is not considered to be misidentified.

Step 010. For each attribute in the identified and stored hydrological data table and its corresponding value, respectively for the flow value of each month, execute according to the following steps 010-01 to 010-02, and then respectively obtain the flow rate for each month Preliminary identification and judgment of the daily flow value, and then enter step 011.

Step 010-01. Use the flow value on the first day of the current month as the first threshold, and then judge whether the difference between the next day’s flow value and the current day’s flow value is less than the first threshold for the flow values of the first two days of the current month, if yes, then Judging that the identification of the flow value of the current day is correct; otherwise, it is determined that the preliminary identification of the flow value of the current day is wrong; thereby obtaining preliminary identification judgments for the flow values of the first two days of the current month, and then proceed to step 010-02.

Step 010-02. For each daily flow value starting from the third day of the current month, determine whether the difference between the next day’s flow value and the current day’s flow value is smaller than the previous day’s flow value, and if so, determine that the current day’s flow value is correctly identified ; Otherwise, it is judged that the preliminary recognition of the flow value of the current day is wrong; thus, the preliminary recognition judgment of the flow value of each day starting from the third day of the current month is obtained.

Step 011. According to each numerical value in the identified stored hydrological data table, and the identification features of each number in each numerical value, through the support vector machine trainer, obtain each numerical value in the identified stored hydrological data table, corresponding to " Ten recognition result probabilities from 0" to "9", and then go to step 012.

Step 012. Obtain the maximum recognition result probability and the second largest recognition result probability among the ten recognition result probabilities from "0" to "9" corresponding to the numbers for each number in each numerical value in the identified and stored hydrological data table, And obtain the difference between the maximum recognition result probability and the second largest recognition result probability, judge whether the difference is less than the preset recognition result probability threshold of 0.1-0.25, if yes, judge that the initial recognition of the number is incorrect; otherwise, judge that the recognition of the number is correct ; Obtain a preliminary identification judgment for each number in each numerical value in the identified stored hydrological data table, and then enter step 013.

Step 013. For each preliminary misidentified flow value in each month, determine whether there is a preliminary misidentified number in the preliminary misidentified flow value, specifically the following two situations:

If it is, it is judged that the flow value of the initial recognition error is wrong, and an alarm is issued. At the same time, the position of the preliminary recognition error number in the preliminary recognition error flow value is analyzed. Integer part in the initial recognition error flow value, the average value of the previous day’s flow value and the next day’s flow value on the date corresponding to the preliminary identification error flow value is used to replace the preliminary identification error flow value; if the preliminary identification error number is located in the preliminary identification For the decimal part in the wrong flow value, replace the decimal in the preliminary identified wrong flow value with the average of the decimals of the previous day’s flow value and the following day’s flow value on the date corresponding to the preliminary identified wrong flow value;

Otherwise, it is judged that the flow value of the preliminary identification error is correct; thus, the verification of each value in the identified and stored hydrological data table is realized.

Through experimental comparison, it can be found that in the digitalization method of the paper hydrological yearbook designed by the present invention, the feature fusion improves the recognition rate compared with a single feature. The lattice feature has a poor recognition effect on the number 0, but it has a better recognition effect on the numbers 6 and 9, and the contour moment feature has a poor recognition effect on the numbers 0, 6, and 8. The results of the three features for other digital recognition are generally consistent. By calculating the complementarity index between the features, it can be found that the fusion of Fourier and coarse grid features has a good ability to distinguish different numbers; the digital boundary contour and the internal digital The fusion of features can describe the whole number more completely from the inside to the outside, which is enough to represent a number, so a better recognition effect is obtained.

The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments, and can also be made without departing from the gist of the present invention within the scope of knowledge possessed by those of ordinary skill in the art. Variations.

Claims (10)

1. a papery Water Year Book digitizing solution, it is characterised in that comprise the steps:

Step 001., according to the layout of the papery Water Year Book page, determines that hydrological data form is positioned at the papery Water Year Book page In location of pixels, subsequently into step 002；

Step 002., according to the location of pixels of hydrological data form in the papery Water Year Book page, is entered respectively for hydrological data form Row vertical and horizontal project, and longitudinal projection's figure, the transverse projection for hydrological data form is analyzed respectively, carries respectively The abscissa of each bar vertical line, the vertical coordinate of each bar horizontal line in water intaking literary composition data form, subsequently into step 003；

Step 003. is according to the abscissa of bar vertical line each in the format of hydrological data form, and hydrological data form, each bar horizontal line Vertical coordinate, for the projection picture of hydrological data form, obtain the number in each numerical value cell of hydrological data form respectively According to image, subsequently into step 004；Wherein, the numeric character in each data image of hydrological data form is white, the end Color is black；

Step 004. is respectively directed to each data image, carries out character cutting for each numeric character in data image, it is thus achieved that Each numeric character block in this data image, and then obtain each numeric character block in each data image respectively, then Enter step 005；

Each numeric character block that step 005. is respectively directed in each data image, extracts the net of numeric character in numeric character block Lattice feature, Fourier's feature, Contour moment feature, collectively as the identification feature of this numeric character, and then obtain each respectively The identification feature of numeric character in each numeric character block in data image, subsequently into step 006；

Each numeric character block that step 006. is respectively directed in each data image, it may be judged whether exist by numeric character block top margin There is downwards the black pixel point of predetermined number, be, judge in this numeric character block as arithmetic point, the most do not do any enter one Step operation；It is respectively directed in each data image after the judgement of each numeric character block, subsequently into step 007 completing；

Step 007., for all identification features of numeric character in all data images, carries out Feature Fusion, constitutes hydrological data In form, the most corresponding " 0 " arrives the numerical identification feature of " 9 ", subsequently into step 008；

Step 008. arrives the numerical identification feature of " 9 ", and each datagram according to the most corresponding " 0 " in hydrological data form In Xiang, the identification feature of numeric character in each numeric character block, by default grader, obtains in each data image respectively Each numeral corresponding to numeric character block, subsequently into step 009；

Step 009., according to the numeral corresponding to each numeric character block in each data image or arithmetic point, respectively constitutes hydrology money Expect the numerical value corresponding to data image in each numerical value cell of form, in conjunction with every attribute of hydrological data form format, Obtain every attribute in hydrological data form, and corresponding numerical value, and store.

A kind of papery Water Year Book digitizing solution, it is characterised in that after described step 009 Also comprise the steps, after execution of step 009, enter step 010；

Step 010., for being identified every attribute and corresponding numerical value thereof in storage hydrological data form, is respectively directed to each The flow number of the moon, 010-01 performs to step 010-02 as follows, and then obtains for each moon every respectively Daily flow numerical value tentatively identify judgement, subsequently into step 011；

Of that month first daily flow numerical value as first threshold, is then respectively directed to two daily flow numerical value before this month by step 010-01., Judge that the difference between next daily flow numerical value and same day flow number, whether less than first threshold, is then to judge flow number on the same day Value identifies errorless；Otherwise judge that flow number on the same day tentatively identifies mistake；It is derived from being respectively directed to two daily flow numbers before this month Value tentatively identify judgement, subsequently into step 010-02；

Step 010-02. is respectively directed to of that month each daily flow numerical value by the 3rd day, it is judged that next daily flow numerical value and same day Whether the difference between flow number, less than proxima luce (prox. luc) flow number, is then to judge that flow number identification on the same day is errorless；Otherwise sentence The flow number on the same day that breaks tentatively identifies mistake；It is derived from being respectively directed to of that month each daily flow numerical value preliminary by the 3rd day Identify and judge；

Step 011. is according to each numerical value identified in storage hydrological data form, and the identification of each numeral in each numerical value Feature, by default training aids, it is thus achieved that is identified each numeral in each numerical value in storage hydrological data form, the most right " 0 " is answered to arrive ten recognition result probability of " 9 ", subsequently into step 012；

Step 012. is respectively directed to be identified each numeral in each numerical value in storage hydrological data form, it is thus achieved that numeral is corresponding " 0 " arrives the maximum recognition result probability in " 9 " ten recognition result probability, and second largest recognition result probability, and obtains Obtain the difference of this maximum recognition result probability and this second largest recognition result probability, it is judged that whether this difference identifies knot less than presetting Really probability threshold value, is to judge that this numeral tentatively identifies mistake；Otherwise judge that this numeral identifies errorless；It is derived from pin respectively Judgement is tentatively identified, subsequently into step 013 to identified in storage hydrological data form each numeral in each numerical value；

Step 013. is respectively directed to each middle of the month, and each tentatively identifies wrong flow number, it is judged that the preliminary flow number identifying mistake In whether there is the preliminary numeral identifying mistake, be to judge that this tentatively identifies the flow number mistake of mistake, and report to the police； Otherwise judge that this tentatively identifies that error flow numerical quantity is errorless；It is achieved in for being identified each number in storage hydrological data form The inspection of value.

A kind of papery Water Year Book digitizing solution, it is characterised in that in described step 011, According to each numerical value identified in storage hydrological data form, and the identification feature of each numeral in each numerical value, pass through Support vector machine training aids, it is thus achieved that identified each numeral in each numerical value, the most corresponding " 0 " in storage hydrological data form Ten recognition result probability to " 9 ".

A kind of papery Water Year Book digitizing solution, it is characterised in that in described step 013, Described basis tentatively identifies and there is the preliminary numeral identifying mistake in wrong flow number, it is judged that this preliminary stream identifying mistake Numerical quantity mistake, and while reporting to the police, tentatively identify that according to this error number tentatively identifies in error flow numerical quantity at this Position be analyzed, if this tentatively identifies that error number is positioned at this and tentatively identifies the integer part in error flow numerical quantity, then Tentatively identify the proxima luce (prox. luc) flow number on date corresponding to error flow numerical quantity and the meansigma methods of a rear daily flow numerical value with this, replace Change this and tentatively identify error flow numerical quantity；If it is little that this tentatively identifies that error number is positioned in this preliminary identification error flow numerical quantity Fractional part, then tentatively identify the decimal of the proxima luce (prox. luc) flow number on date corresponding to error flow numerical quantity and a rear daily flow with this The meansigma methods of the decimal of numerical value, replaces this and tentatively identifies the decimal in error flow numerical quantity.

5. according to papery Water Year Book digitizing solution a kind of described in any one in Claims 1-4, it is characterised in that institute State step 004, carry out character cutting for each numeric character in data image, it is thus achieved that each number in this data image Value character block, specifically includes following steps:

The detection of step a01. obtains each white pixel point in data image within each numeric character, and each limit of this data image Edge is respectively at a distance of each numeric character minimum range, and the white pixel point on corresponding numeric character, subsequently into step a02；

Step a02. judges by being obtained each white pixel point in this data image respectively for previous step, it is judged that pixel Whether the pixel of upper and lower, left and right each position is white pixel point, is, judges that this pixel is inside numeric character Pixel；Otherwise judge, according to identifier, the edge pixel point that this pixel is character, and obtain this pixel in these data The row number of place pixel column in image；It is thus directed towards previous step and is entered respectively by this data image obtains each white pixel point Row judges, it is thus achieved that the row of place pixel column in each this data image of numeric character top edge pixel place in this data image Number, subsequently into step a03；

Step a03. is according to each numeric character top edge pixel place pixel column in this data image in this data image Row number, divide for each numeric character in this data image, it is thus achieved that each numeric character block in this data image.

6. according to papery Water Year Book digitizing solution a kind of described in any one in Claims 1-4, it is characterised in that institute State in step 005, each numeric character block being respectively directed in each data image, extracts numeric word in numeric character block The grid search-engine of symbol, specifically includes following steps:

Step b01. obtains the border of the upper and lower, left and right of numeric character block, and is derived from numeric character ontology diagram picture, so Rear entrance step b02；

Step b02. carries out center of gravity normalization for this numeric character ontology diagram picture, and will be through center of gravity this numeric character normalized Ontology diagram is slit into predetermined number sub regions image as average mark, subsequently into step b03；

Step b03. obtains in this numeric character ontology diagram picture the proportion of white pixel point in each sub regions image respectively, altogether With constituting the grid search-engine of numeric character in this numeric character block.

7. according to papery Water Year Book digitizing solution a kind of described in any one in Claims 1-4, it is characterised in that institute State in step 005, each numeric character block being respectively directed in each data image, extracts numeric word in numeric character block Fourier's feature of symbol, specifically includes following steps:

Step c01. carries out two dimensional discrete Fourier transform for numeric character block, subsequently into step c02；

Step c02., by this numeric character block through two dimensional discrete Fourier transform, proceeds central transformation, will numeric word Symbol block is averagely divided into four blocks of subregion images, and carries out diagonal angle exchange, it is thus achieved that Fourier's image is composed, subsequently into step c03；

Step c03. is for Fourier's its Fourier coefficient of image analysis of spectrum after central transformation, it is thus achieved that in Fu of this numeric character block In leaf system number, concentrate region more than the Fourier coefficient presetting amplitude thresholds, constitute significantly Fourier coefficient region, so Rear entrance step c04；

Step c04., by significantly Fourier coefficient region, is extracted predetermined number discrete Fourier transform coefficient, and is carried out Normalization, constitutes Fourier's feature of numeric character in this numeric character block.

8. according to papery Water Year Book digitizing solution a kind of described in any one in Claims 1-4, it is characterised in that institute State in step 005, each numeric character block being respectively directed in each data image, extracts numeric word in numeric character block The Contour moment feature of symbol, specifically includes following steps:

Step d01. carries out contours extract for the numeric character in numeric character block, subsequently into step d02；

Step d02. carries out not bending moment for the profile of numeric character in this numeric character block and processes, and extracts predetermined number two dimension wheel Wide invariant moment features, constitutes the Contour moment feature of numeric character in this numeric character block.

9. according to papery Water Year Book digitizing solution a kind of described in any one in Claims 1-4, it is characterised in that described Step 007 specifically includes following steps:

Step e01., according to permutation and combination, for all identification features of numeric character in all data images, carries out any two Identify the combination of feature, constitute the combination of all identification features, subsequently into step e02；

Step e02., by all identification features of numeric character in all data images, constitutes corresponding numeral in hydrological data form " 0 " arrives the sample set S of " 9 ", is then respectively directed to each group of identification feature combination, according to equation below (1):

$C_{ij,A}=\frac{E(S_i\&cup;S_j)-E(S_i\&cap;S_j)}{E\left(S\right)}---\left(1\right)$

Obtain the feature complementary index C of this group identification feature combination relative standard digital " 0 "-" 9 " respectively_ij,A；And then obtain respectively Obtain the feature complementary index C of each group of identification feature combination relative standard digital " 0 "-" 9 " respectively_ij,A；Subsequently into step e03；Wherein, S_iAnd S_jRepresent that sample set S is identified feature F respectively_iWith identification feature F_jThe sample set of wrong point； E (S) represents the number of samples in sample set S；E(S_i∪S_j) represent sample set S_iWith sample set S_jBetween and concentrate Number of samples；E(S_i∩S_j) represent sample set S_iWith sample set S_jBetween occur simultaneously in number of samples； A={0,1 ..., 9}, C_ij,ARepresent by identifying feature F_iWith identification feature F_jConstituted identification feature combination relative standard numeral The feature complementary index of A；

Step e03. is respectively directed to each group of identification feature combination, according to equation below (2):

${\mathrm{TC}}_k=\frac{{\mathrm{\&Sigma;}}_{0,i\&NotEqual;j}^9{C}_{ij}}{A_{10}^2}---\left(2\right)$

Obtain each group of identification feature combination overall complementation index TC relative to standard digital respectively_k, subsequently into step e04；Its In, k={1 ..., K}, K represents the number of combinations that all identification features combine, TC_kRepresent kth group identification feature combination phase Overall complementation index for standard digital；

Step e04. combines for all identification features, sorts from large to small by its overall complementation index, it is thus achieved that sequence the first two is known Other feature combines, and then identifies that feature combination carries out Feature Fusion for these two, constitutes difference correspondence in hydrological data form " 0 " arrives the numerical identification feature of " 9 ".

10. according to papery Water Year Book digitizing solution a kind of described in any one in Claims 1-4, it is characterised in that institute State in step 008, arrive the numerical identification feature of " 9 ", and each number according to the most corresponding " 0 " in hydrological data form According to the identification feature of numeric character in each numeric character block in image, by support vector machine classifier, obtain each respectively Each numeral corresponding to numeric character block in data image.