CN104598887B - Recognition methods for non-canonical format handwritten Chinese address - Google Patents

Abstract

The present invention gives a kind of recognition methods for non-canonical format handwritten Chinese address, and establish the address method for expressing of a cannonical format.This method proposes that the structure of word level tree stores Chinese address storehouse, wherein each node stores an address word, paths from root node to leafy node then store the address that a cannonical format is write.Whole Address Recognition includes：Build word level tree；Build character index table；Image over-segmentation；Combination and segmentation block；Character recognition；Generate candidate site word；Cannonical format Address Recognition.The present invention can identify the corresponding address of the address of cache of non-canonical format writing to cannonical format so as to realize.

Description

Recognition method for handwritten Chinese addresses in non-standard format

technical field

The invention belongs to the technical field of handwritten Chinese address recognition, in particular to the recognition of handwritten Chinese addresses in non-standard format.

Background technique

Chinese address recognition plays a key role in the automatic sorting of letters and parcels. In the mail processing center, a large number of letters and parcels are processed and dispatched every day. This requires that the processing of mail not only be fast, but also precise. Although people have made great progress in the research of Chinese address recognition, handwritten address recognition in real letters is still a problem that has not been well solved. For example, there are a large number of Chinese characters and various writing styles, and there may be consecutive strokes between characters in the address. In particular, the variability and irregularity of address writing formats greatly increase the difficulty of handwritten address recognition. Very little work has been done to identify addresses specifically taking this aspect into account.

The main goal of the traditional Chinese handwritten address recognition method is to recognize all the Chinese characters on a given address image. They require a list of addresses to provide context for address identification. Each entry in this list is a complete address, and is usually used one by one to match the recognition result of the input address image. In order to improve the efficiency of address retrieval and reduce the storage space of the address list, a method based on a search tree structure is proposed to store address information. In these tree structures, each node stores a character, so it is also called a word-level tree. However, on the one hand, word-level trees are sensitive to noise because it requires that all characters in the address image must be recognized in order. On the other hand, whether the match between the candidate pattern block and the child nodes of the root node is accurate will have a great impact on the recognition performance. Briefly, address recognition based on a word-level tree structure needs to rely on a pre-defined address list, if the address information in the address list is incomplete, i.e., it does not include all written format variations of the address, or the address list If the address information provided is insufficient, the recognition rate of these address recognition methods will be greatly reduced in practical applications.

Usually, an address is composed of several address words, which are defined as basic administrative units. For example: the standard writing format address "Zhongshan North Road, Putuo District, Shanghai City" shown in Figure 2 (a) contains address words "Shanghai City", "Putuo District", and "Zhongshan North Road". The last word of each address word is defined as a keyword, such as "province", "city", "district", "road", and so on.

But in practical application, the way of writing the address on the envelope is very complicated, and people usually don't write according to the standard format of the address. For example, in Figure 2, Figure 2(a) is the standard writing form of the address, and Figure 2(b-e) shows its various non-standard writing forms, which are considered reasonable in reality.

To sum up, it is almost an impossible task to manually collect all these non-standard address writing forms.

Contents of the invention

The purpose of the present invention is to propose a method based on word-level tree structure at the deficiencies in the prior art, these non-standard handwritten Chinese addresses are finally mapped to the corresponding address of standardized writing, realize its identification; Limitations of Handwritten Chinese Address Recognition.

The purpose of the present invention is achieved like this:

A method for recognizing a handwritten Chinese address in a non-standard format, comprising the following steps:

Constructing a word-level tree, which is used to represent and store addresses in a standardized writing format;

Constructing a character index table, which is used to represent the association between a single character and an address word;

Segmentation-recognition processing, the segmentation-recognition processing method is used to segment the image into characters, merge and perform character recognition on the candidate pattern blocks formed by merging the segmented blocks;

Generate candidate address words, the method for generating candidate address words is used to obtain candidate address words with higher confidence;

Canonical format address recognition, the canonical format address recognition method is used to map the handwritten address to be recognized to its corresponding canonical format writing mode; wherein:

The depth of described construction word-level tree is 5, and the first layer is a root node, stores and represents respectively " province ", " city ", the address word of " district " and " road " name from the second layer to the fifth layer, wherein Each node stores an address word.

The construction of the character index table is used to store all the characters contained in the address words, and associate the characters with all the address words containing the characters.

The segmentation-recognition process also includes:

Image over-segmentation, which divides the image into atomic blocks, which is used to separate the overlapping parts or continuous strokes between handwritten Chinese characters;

Merging segmentation blocks, merging consecutive atoms into candidate pattern blocks, used to restore a single character or the separation of left and right structure characters caused by the over-segmentation process;

Character recognition, used to identify candidate pattern blocks and calculate the confidence of the recognition result;

The image is over-segmented by using connected element analysis, normalized overlap calculation and projection analysis technology to over-segment the image and finally obtain a series of atomic segmentation blocks;

The method for merging segmentation blocks is to merge continuous atomic segmentation blocks one by one to form candidate pattern blocks;

The character recognition also includes:

A handwritten character classifier for classifying candidate pattern blocks;

Confidence conversion, used to calculate the confidence of the recognition result;

The generation of candidate address words is obtained by pruning the word-level tree in combination with the candidate pattern recognition results, the character index table and the address words stored in the word-level tree.

The canonical format address recognition is to combine the candidate address words with the word-level tree, use the bottom-up search method to combine the candidate address words on the word-level tree, and finally generate the candidate address. The candidate address with the highest confidence is taken as the final address recognition result.

The present invention overcomes the limitations of traditional methods for identifying non-standard handwritten Chinese addresses, and proposes a method based on a word-level tree structure, which can map addresses written in non-standard formats to corresponding addresses in standard formats, thereby realizing writing in non-standard formats Address identification.

Description of drawings

Fig. 1 is a flowchart of the present invention;

Figure 2 is an example diagram of different writing methods for the address "Zhongshan North Road, Putuo District, Shanghai";

Fig. 3 is the schematic diagram of the word-level tree of standard writing address format;

Fig. 4 is an example diagram of an address row image over-segmentation result;

Fig. 5 is an example diagram of a candidate mode block diagram;

Fig. 6 is the generation schematic diagram of candidate address words;

Fig. 7 is the example diagram of the corresponding position of candidate address words in the candidate mode block diagram;

Fig. 8 is a flow chart of word-level tree path search;

Fig. 9 is the instance diagram of searching and generating the candidate address in the word-level tree;

Fig. 10 is an example diagram of the recognition result of a handwritten Chinese address in a non-standard format.

detailed description

As shown in Figure 1, it is a flowchart of an embodiment of the present invention, and the method specifically includes:

A word-level tree is constructed to represent and store addresses in a canonical writing format.

The address-administrative relationship in China is a top-down hierarchical structure. The number of layers is generally 4. These 4 layers correspond to the names of "province", "city", "district" and "road" respectively. Define a tree from this structure, with depth 5. The root node is empty, and the address words representing the names of "province", "city", "district" and "road" are respectively stored from the second layer to the fifth layer, and each node stores an address word. In the word-level tree, a path from the root node to the leaf node corresponds to a written address in a normalized format.

In order to deal with the situation of omitting keywords in address writing, the last word of each address word (except the word "road") is defined as an optional item. The constructed word-level tree is shown in Figure 3, and the word representations in brackets are optional.

In this word-level tree, once a certain leaf node (namely road name) is identified, all candidate addresses containing this road name can be obtained. For example, if the address word "Zhongshan North Road" is identified, the address words "Shanghai City", "Putuo District", "Zhejiang Province", "Hangzhou City", " Downtown", and so on. Then the relevant candidate addresses "Zhongshan North Road, Zhejiang Province, Putuo District, Shanghai City" and "Zhongshan North Road, Xiacheng District, Hangzhou City, Zhejiang Province", etc., can be obtained. Further, if the address word "Putuo District" or "Shanghai City" is also recognized, then the candidate address "Zhongshan North Road, Putuo District, Shanghai City, Zhejiang Province" is more likely to be used as the recognition result, especially when "Putuo District " and "Shanghai" are both identified.

Construct a character index table to represent the association between individual characters and address words.

As shown in Table 1, the character index table is divided into 3 columns, the second column is all characters appearing in the address words, and the first column is the GB2312-80 code corresponding to the characters in the second column. Column 3 is all relevant address words including a certain character. When a character is recognized, all address words containing this character can be obtained to generate the final candidate address words.

Table 1

Image over-segmentation, which is used to separate the overlapping parts or continuous strokes between handwritten Chinese characters.

First, the connected element analysis is performed on the image, and then the normalized overlap calculation is performed on adjacent connected elements to determine whether to merge these connected elements, because some of them may be different parts of the same character. Finally, through projection analysis, it is judged whether the connected elements contain continuous strokes, and if so, they are segmented. Split the overlapping parts of different characters or the connecting strokes between them as much as possible, and finally obtain a series of atomic segmentation blocks. The segmentation results of Figure 2(d) are shown in Figure 4. In Fig. 4, the atomic blocks are arranged in order from left to right, and the atomic blocks are numbered sequentially above the atomic blocks.

Merge segmentation blocks, which are used to restore a single character or the separation of left and right structure characters caused by the over-segmentation process.

After the image is over-segmented, consecutive atomic blocks are combined and candidate pattern blocks are generated, as shown in Figure 5. Define all candidate pattern blocks as a set P={p _(1,1) , p _(1,2) , p _(1,3) , p _(2,1) , p _(2,2) , p _{(2, 3)} , ..., p _{(m, n)} , ..., p _{(l, q)} }, where (m, n) is the number of the atomic block (1≤m≤l, 1≤n≤q ), l is the total number of atomic blocks, q is the maximum number of atomic blocks contained in a candidate pattern block, in this embodiment, q is set to 3.

Character recognition, which is used to identify candidate pattern blocks and calculate the confidence of the recognition results.

In the candidate pattern block diagram, a character classifier is used to identify each candidate pattern block to generate a series of candidate characters. For the recognition of handwritten Chinese characters with a large number of categories and no constraints, the MQDF method is currently the most practical method. But its character feature storage is relatively large. The invention combines the method of MQDF discriminant learning and shared distribution subspace, and reduces the space occupied by character storage features without reducing the recognition rate.

Regarding the confidence of character recognition, that is, the posterior probability p(w|x), (w is the recognized character, x is the image feature vector), it is very important for the recognition of character strings, but it cannot be directly obtained from the MQDF classifier The output is obtained. Therefore, it is necessary to use the method of confidence conversion to convert the output of the classifier into the posterior probability. In the present invention, the sigmoidal function is applied to the confidence conversion, and the posterior probability of the character recognition result can be expressed as

Among them, M is the total number of categories of characters, d _j (x) is the output score of the classifier for category w _j , α and β are the confidence parameters to be optimized, which can be minimized by minimizing the cross entropy loss function (CE) Optimize it. Through the theoretical proof of Dempster-Shafer (DS), the calculation of character confidence can be expressed as:

Finally, for each pattern block, the top 20 candidate characters with the highest confidence are taken as the recognition results, and they are arranged in descending order of confidence. The recognition results of the candidate pattern blocks in Fig. 5 are shown in Table 2. The recognition results of some pattern blocks are empty, because they can be directly judged whether they are a reasonable character by their shape, that is, the aspect ratio, such as pattern blocks p _{(10, 1)} , p ₍₁₅ , 1) , p _{( 17, 1)} , p _{(19, 1)} are not reasonable characters.

Table 2

The candidate address words are generated, and the word-level tree is pruned by combining the candidate pattern recognition results, the character index table and the address words stored in the word-level tree.

An address list (AW_O) storing all address words is implicitly represented in the word-level tree, as shown in Figure 6(a). Starting from this table, candidate address words are generated through a series of processing. These processes include three steps: First, through the association of identified candidate characters and address words, the table AW_O is pruned, and the associated address words generate a new address word list AW_R. Then, through the matching of the identified candidate characters in the address words and the position restriction relationship of the candidate pattern blocks, the table AW_R is further pruned to obtain the address list AW_P. Finally, by calculating the score of the address words in AW_P, the address words whose score is greater than a preset threshold are stored in the list AW_C, and the address words stored in AW_C are the final candidate address words. The specific instructions are as follows:

(1), Generation of AW_R: Through the recognition result of the candidate pattern, the table AW_O does not satisfy All address words are deleted (nr is the number of characters that have been identified in a certain address word, and nl is the number of characters contained in this address word). Finally, the remaining candidate address words form the AW_R list (as shown in Figure 6(b)).

(2) Generation of AW_P: It is necessary to consider the position restriction relationship between the recognized characters in the candidate address words and the corresponding pattern blocks in the image. If the position in the image of a pattern block matched by a recognized character in a certain candidate address word in the table AW_R does not meet the position restriction relationship, the candidate address word will be deleted. Finally, the remaining candidate address words form the AW_P list (as shown in Figure 6(c)).

(3), generation of AW_C: the present invention proposes a method for calculating the address word score, which is used to calculate the address word score in AW_P, and the specific calculation method is introduced below. If the score of a candidate address word in AW_P is less than the predefined experience threshold, the candidate address word is deleted. The last remaining address words form the table AW_C. Each address word in the table AW_C is defined as the final candidate address word (as shown in Figure 6(d)).

Address words are calculated by the following formula:

This formula takes two cases into account: one is the proportion of recognized characters in the address word to all characters contained in the address word, and the other is the confidence of the cut block. in, is the word confidence, which can be calculated by formula (2). The calculation of nr/nl takes into account the proportion of recognized characters in the address word. Relatively, if all the characters in the address word are recognized, and the relative position of the match with the pattern block is reasonable, then increase the confidence of the address word, and the increased score is represented by a constant v (1≤v≤4) . SC is the confidence of the cutting block, defined as

Among them, m is the number of continuous atomic blocks forming a pattern block, and the combination of these atomic blocks does not contain continuous strokes. pw/ph The aspect ratio of this pattern block.

In order to reduce the error rate of recognition, all address words whose score is lower than a threshold ε in the table AW_P will be deleted. ε is defined as

Among them, nl is the number of characters contained in the candidate address word, is an empirical threshold, after many tests, Taking 2.5 can make the recognition system get the best performance.

Through this step, the corresponding positions of the generated candidate address words in the candidate mode block diagram are shown in FIG. 7 .

Canonical format address recognition, which is used to map the handwritten address to be recognized to the corresponding canonical format writing method.

When recognizing a Chinese handwritten address, all its non-standard formats can be mapped to a certain path in the word-level tree. After the candidate address words are generated, the tree can be searched in combination with their node relationships in the word-level tree to generate candidate addresses written in a canonical format. Using the bottom-up search method, search from the leaf node of the tree (corresponding to the road name) to the root node. In this step, several candidate addresses can be obtained, and the score of each candidate address is equivalent to the accumulation of the scores of the identified candidate address words contained in it. Finally, take the candidate address with the largest score as the recognition result. The specific flow of this step is shown in Figure 8.

In the present invention, the address words representing "province", "city", "district" and "road" name are stored respectively with 4 lists, and these 4 lists are represented by PR, CI, DI, RO respectively. In addition, a triplet set TN={CN, PN, AS} is used to represent a node in the search space. Among them, CN points to the current node of the word-level tree, PN points to the parent node of CN, and AS is the accumulated address word score during the search process. For a word candidate W, its leftmost pattern block (lp(W)) and rightmost pattern block (rp(W)) correspond to its first matched character and last matched character, respectively. Judging whether the position restriction relationship between the address word corresponding to the parent node and the address word corresponding to the child node in the pattern diagram is reasonable, is based on whether the rp of the parent node is smaller than the lp of the child node. The location and size of address words in the text are sorted in ascending order from left to right.

Before searching, check if the list RO is empty. If RO is empty, that is, no road name has been recognized, at this time AS=0, stop this search, and the recognition result is rejection. Otherwise, the search will be performed one by one starting from the address words stored in the list RO. First, CN points to an address word in RO, AS is initially the score corresponding to the address word, and PN points to the parent node of CN. The next search is divided into two cases: PN∈DI or if It indicates that the candidate address word pointed to by PN is not recognized, in this case, PN directly points to the parent node of the node pointed to by PN, and then the search continues. If PN ∈ DI, it means that the candidate address word pointed to by PN has been identified. If the address words pointed to by PN and CN satisfy the positional relationship rp(PN)<lp(CN), then AS is equal to the accumulation of the scores of these two address words. Then CN points to the word-level tree node corresponding to PN, and PN points to the parent node of CN. Otherwise, if the two words do not satisfy the positional relationship, PN directly points to the parent node of the node pointed to by PN, and then continues searching. When PN points to the root node of the tree, it means that this search is over. Finally, the canonical address obtained from the reverse search from the leaf node to the root node is used as the result of the candidate address, and AS is its corresponding score. A binary set RS={ξ, AS} is used to store the search results, where ξ stores the canonical candidate addresses obtained from the current search.

Figure 9 illustrates the search process in the word-level tree. For example, start searching from the leaf node "Zhongshan North Road", AS is equal to the score of this address word 20.19. The address word "Putuo District" pointed to by PN has been identified as a candidate address word, and rp("Putuo District")<lp("Zhongshan North Road"), then AS is equal to 34.85 (=20.19+14.66). Finally, the search result for this path obtains the candidate address "Zhongshan North Road, Putuo District, Shanghai", and its corresponding score is 48.41 (=20.19+14.66+13.56), which is the highest score of all candidate addresses, so it is used as the final identification result. Through route search, some unrecognized address words are also included in candidate addresses as recognition results, but their scores are not accumulated.

The positions of some candidate address words in the segmentation pattern frame may overlap (as shown in Figure 7). If the address words of the same level overlap, it will not affect the search of the tree, because there is no relationship between them, and there is no relationship between the parent node and the child node in the tree, so different paths can be obtained in the pattern diagram, such as : "Shanghai City" and "Shanghai", "Putuo District" and "Putuo". On the contrary, if the two address words are of different levels, they may correspond to the relationship between the parent node and the child node of the same path, for example: "Putuo District" and "Putuo Road". In this case, the low-priority word will be skipped during the search process, and the path will not accumulate the score of this low-priority address word. In the present invention, the priority of the address word increases with the increase of the number of node layers, so the priority of the address word representing the road name is the highest.

After all address words in the RO are searched in the tree, several candidate addresses are generated. Finally, only the candidate address with the highest score is taken as the recognition result. The recognition result is denoted by S, which is defined as

S=arg maxξ(AS _i |i=1, 2,..., n) (6)

n is the total number of generated candidate addresses, in FIG. 7, n=5. Obviously, AS _i gets the maximum score of 48.41 when i=3, so its corresponding normalized writing address "Zhongshan North Road, Putuo District, Shanghai" is the final recognition result in Figure 2(d).

Figure 10 shows the recognition results of the Chinese handwritten address line image in Figure 2. It can be seen from FIG. 10 that the addresses of these three types of non-standard writing formats can all be recognized as the standard writing address "Zhongshan North Road, Putuo District, Shanghai City" through the present invention.

Claims (3)

1. A recognition method for a handwritten Chinese address in an irregular format, characterized in that the method comprises the following steps:

step 1: constructing a word-level tree for representing and storing addresses in a standard writing format; the depth of the constructed word-level tree is 5, the 1 st layer is a root node, address words representing names of 'province', 'city', 'district' and 'road' are stored from the 2 nd layer to the 5 th layer respectively, and each node stores one address word;

step 2: constructing a character index table for representing the association between a single character and an address word; the constructed character index table is used for storing all characters contained in the address words and associating the characters with all the address words containing the characters;

and step 3: segmentation-recognition processing for performing character segmentation and merging on the image and performing character recognition on candidate mode blocks into which the segmentation blocks are merged; the method specifically comprises the following steps:

over-segmenting the image, namely segmenting the image into original sub-blocks for segmenting the overlapped part or the continuous stroke part among the handwritten Chinese characters;

merging the segmentation blocks, merging the continuous atom segmentation blocks one by one to form a candidate mode block, wherein the candidate mode block is used for recovering the condition that a single character or characters of left and right structures are separated caused by an over-segmentation process;

character recognition, which is used for recognizing the candidate mode block and calculating the confidence coefficient of the recognition result;

and 4, step 4: generating candidate address words for obtaining the candidate address words with higher confidence; the generated candidate address words are obtained by pruning a word-level tree by combining a candidate pattern recognition result, a character index table and address words stored in the word-level tree;

and 5: standard format address recognition, which is used for mapping the handwriting address to be recognized to a corresponding standard format writing mode; the standard format address identification is to combine the candidate address words with a word-level tree, combine the candidate address words by a bottom-up searching method for the word-level tree, and finally generate a candidate address; and taking the candidate address with the highest confidence as a final address recognition result.

2. The identification method of claim 1 wherein the image over-segmentation is performed by applying connected component analysis, normalized overlap computation and projection analysis to the image and finally obtain a series of atomic segmentation blocks.

3. The recognition method of claim 1, wherein the character recognition further comprises:

the handwritten character classifier is used for classifying the candidate mode blocks;

and the confidence conversion is used for calculating the confidence of the recognition result.