EP1799359B1 - Method for sorting addressed mailings according to the destination address - Google Patents

Abstract

In the multistage sorting method, the mail items read in the first step are re-identified in the subsequent steps on the basis of characteristic features determined. If unambiguous assignment to a plurality of candidates is not possible, reading steps are additionally carried out in the subsequent sorting steps and the partial reading results of the candidates are compared with the corresponding partial reading results of the first steps until unambiguous assignment is produced.

Description

The invention relates to a method according to the preamble of claim 1, such as for example FR-A-2841673 known.

The automatic sorting of postal consignments based on the destination address takes place in several processing steps.

The first step is to exclude a picture of the front of the program. An automatic reading system determines the position of the address block in the image, reads it using OCR techniques and determines the required distribution information. The pictures of programs that could not be read automatically are connected to a video coding station and manually coded there.

In many sorting systems, the distribution information is either printed directly on the item or it can be assigned from a database via a unique identification that is printed on the item. In both cases, the shipments are marked by machine-readable bar codes. Thus, the once determined distribution information can be accessed in the subsequent processing steps.

The application and reading of these bar codes requires some hardware and is maintenance-intensive. In some cases, the printing of barcodes is not possible or not desired.

In order to recognize the broadcasts even without applied identifications, a method has been disclosed which makes it possible to use a program based on characteristic features (so-called fingerprints) ( DE 40 00 603 C2 ) to identify. To simplify the application of this identification technique in practice, a method has been developed which significantly limits the search space for identifying consignments ( EP 1 222 037 B1 ).

The invention has for its object to provide a method for sorting addressed broadcasts in several processing steps, in which even with hardly distinguishable programs their identification without applied identifiers with the help of characteristic features to reduce the necessary reading effort is possible.

According to the invention the object is achieved by the features of claim 1.

In this case, in the first processing step, together with the characteristic features assigned to them, the respective reading intermediate results are stored in the database in addition to the read destination addresses / distribution information.

In the recognition process of a further processing step, if there is no unambiguous agreement in defined credibility of the characteristic features with a plurality of stored feature sets of found candidates from the first processing step, the automatic address reader for the respective further processing step is started for reading the address of the current program to be recognized.

Then, the current interim reading results or reading results of the item to be identified are compared with the corresponding stored reading intermediate results or reading results of the items of the determined feature candidates.

When found match with defined credibility, the feature associated with this feature candidate with the stored distribution information of the destination address represents the recognized shipment and the sorting is based on this distribution information.

This makes it possible, in the case of several possible feature candidates, to which no clear selection of a single candidate is only possible due to the feature comparisons and therefore had to be manually re-coded so far, to process the vast majority automatically.

Advantageous embodiments of the invention are set forth in the subclaims.

It is advantageous to carry out the reading and comparing process in the recognition process of a further processing step in stages according to the subsequent reading steps, the subsequent reading and comparing step takes place only if the previous does not give a clear candidate assignment. As a result, the reading and comparison costs are only to the necessary extent.

For this it is advantageous to proceed in the following order:
After the start of the automatic address reader and after the determination of the address area of the destination address, this is compared with the address areas to the feature candidates. If the required credibility matches exactly one of the address areas, the associated program is identified and the read process is terminated. If the assignment to only one feature set is not possible, the reading of the individual characters and words and a comparison with the individual characters and words stored for the remaining feature candidates takes place. If the required credibility matches the individual characters and words of only one feature candidate, the assigned program is identified and the reading process is terminated. If the assignment to only one feature set is still not possible, the address interpretation is performed, the results of which are then compared with those of the remaining feature candidates. If the required credibility matches the results of the address interpretation to only one feature candidate, the assigned program is identified and the reading process is terminated.

In this context, it is advantageous to carry out the process described in the previous paragraph with existing primary and secondary readers in the address readers, which are started sequentially in the recognition process with the or the respective secondary readers, if reached with the primary reader yet no automatic identification has been.

In order to limit the required memory requirement, it is advantageous to erase the characteristic features stored in the database with the reading intermediate results and reading results / distribution codes when a certain expiry time has been exceeded or when the last processing step has been completed.

It is also advantageous, if the automatic identification has not been successfully completed, to encode only the images or address alternatives of the broadcast candidates manually to video encoding by means of selection and thus keep the coding effort low.

Subsequently, the invention will be explained in an embodiment with reference to the drawing.

FIG 1

eine schematische Strukturdarstellung für den ersten Verarbeitungsschritt,

FIG 2

eine schematische Strukturdarstellung für einen weiteren Verarbeitungsschritt,

FIG 3

ein Flussbild des Verfahrensablaufes.

Show

FIG. 1

a schematic structural representation for the first processing step,

FIG. 2

a schematic structural representation for a further processing step,

FIG. 3

a flow chart of the procedure.

In the first processing step after the image acquisition with a camera 1 and a preprocessing 10 for the creation of defined image relationships, the fingerprint method includes the steps of feature extraction 2 (determination of the characteristic Characteristics 3), determination of the distribution information in a reading process and storage of these data 3, 9 including the intermediate results of the reading process in a database 4.

During preprocessing, the in FIG. 3 is shown, various image normalization steps 10 are performed. This includes a rotational orientation, which eliminates minor deviations in the rotational position of the programs, and a standardization to compensate for specific camera characteristics. For gray images this could be a brightness normalization. This area also includes an optional binarization.

Then, in addition to global characteristics such as the size of the broadcast, significant features are extracted from the preprocessed image 2. Structural features, histogram-based features derived from transformations (Fourier, Cosine Transform, Wavelets) can be used. Structural features used are, in particular, features derived from the franking and the address block, since these are already being processed by machine. For the address block e.g. As stable as possible features such as the number of lines and the decomposition of lines in words and characters are used.

In a concrete application, the exact composition of the feature vectors used will be matched to a significant sample of broadcast images to achieve optimal recognition results.

The destination address is then read by a primary reader 5. The area of the destination address (ROI) is first determined 5.1, then the characters of the destination address are read with the OCR part 5.2 and the last step is the address interpretation using an address dictionary 5.3 , The result is then either a read verified and unique address or there was no unique result and it will be a second read attempt in a secondary reader 7 started with another reading software. The same processing steps take place here, ie determining the ROI 7.1, reading the destination address 7.2 and carrying out the address interpretation 7.3. If there is still no clear result, a video coding is performed. 8. After the storage in the database 4, the characteristic feature set / feature vector 3 of the respective mail item is then located and the reading results and reading intermediate results 9 are assigned to it.

In the recognition process, the preprocessing 20 of the image data, the feature extraction 12.1 and the storage of the feature vector also takes place after image acquisition by means of a camera 11 for fingerprint recognition 12. Subsequently, this is compared with the feature vectors contained in the database 4 12.2, 12.3. In order to be able to identify a shipment efficiently by means of a determined feature vector / fingerprint, space partitioning techniques are used for large data volumes to determine the nearest neighbors. These make it possible to quickly find both the candidates adjacent to the feature space and to efficiently manage the set of currently considered candidates. This aspect is of particular importance in the fingerprint method, since the reference quantity is subject to constant changes due to newly added items and completely processed items.

Was found in the comparison / database query 12.2,12.3 a clear match with one of the stored feature vectors, the associated distribution information 9 associated with the relevant program and the sorting can be done.

If no candidate from the stored feature vectors could be determined at all, a completely normal and known automatic reading process must be carried out, if necessary with two readers 15, 17, ie with the steps: ROI search 15.1, 7.1, OCR character reading 15.2, 17.2 and address interpretation 15.3,17.3. If the destination address can not be uniquely identified, the program picture is manually encoded at a video coding station 18 so that the program can be encoded.

In the event that a plurality of candidate candidates 13 of feature vectors / fingerprints 3 were determined, additional features are used by the address readers 15 or 17. Of particular interest are the results of the Region of Interest (ROI) search 15.1,17.1, if this information is not yet used in the fingerprint itself, the single character alternatives found in the character reading process 15.2,17.2 and of course the results of the address interpretation 15.3,17.3. It suffices, as in the flowchart ( FIG. 3 ) when only the reading steps that are just needed to select the correct fingerprint candidate are successfully completed. In this case, the results obtained by / from the address readers 15, 17 are compared with the information stored on the fingerprint candidates 13 in the database 4, and irrelevant candidates are discarded until only one candidate remains. The sequence including the comparison processes is determined in a so-called voter 16.

For example, only the outgoing line, i. Postal code and location automatically recognized by the reader. With this information, however, the voter 16 can in many cases already clearly select a program from the fingerprint candidates 13. The video coding 18 of the still missing address parts can then be omitted.

In addition, the voter 16 decides, based on the credibility of the result of the fingerprint determination and on the results used and partial results of the reading process, whether the result has been determined sufficiently reliably. 22. If not, the complete destination address must be automatically read or video-coded.

In FIG. 3 was shown because of the clarity of the process with only one primary reader. If there are still secondary readers, the processes are also carried out accordingly for these readers.

Claims (6)

1. Method for sorting addressed mail items by destination address in a plurality of processing steps, wherein in the first processing step an image of the mail item is recorded and the destination address is read therefrom and the distribution information and also characteristic features of the image (3) are determined (2) and stored, assigned to the mail item, in a database (4), and wherein in the further processing steps the images of the mail items are likewise recorded, the characteristic features (13) extracted therefrom (12), compared with the features (3) determined in the first processing step and, in the event of a match of defined plausibility, the associated item is identified using the stored distribution information,

   - in the first processing step the relevant intermediate reading results are stored (9) in the database (4) together with the characteristic features (3), assigned thereto, as well as the read destination addresses/distribution information,

   - the current intermediate reading results or reading results of the mail item to be identified are then compared with the corresponding stored intermediate reading results or reading results of the mail items of the feature candidates determined and if a match of defined plausibility is found, the mail item assigned to this feature candidate with the stored distribution information of the destination address constitutes the re-identified mail item and sorting takes place on the basis of this distribution information
   characterised in that

   - in the re-identification process of a further processing step, if there is no clear match of defined plausibility of the characteristic features determined in said processing step with a plurality of stored features sets of found candidates from the first processing step, the automatic reading process for said further processing step is activated to read the address of the mail item currently to be re-identified.
2. Method according to claim 1,
   characterised in that
   the reading and comparison operation is carried out stepwise in the re-identification process of a further processing step according to the sequential reading steps (15.1, 15.2, 15.3, 17.1, 17.2, 17.3), the subsequent reading and comparison step only taking place if the preceding step yields no unambiguous candidate assignment.
3. Method according to claim 2,
   characterised in that
   after initiation of the automatic reading process and after determination (15.1) of the address region of the destination address, this region is compared with the address regions for the feature candidates, in the event of an unambiguous match of required plausibility (22) with one of the address regions the associated mail item is identified and the reading process is terminated, if assignment to a signal feature set is not possible, reading (15.2) of the individual characters and words and comparison with the individual characters and words stored for the remaining feature candidate take place and in the event of a match of defined plausibility (22) with the individual characters and words of a single feature candidate, the associated mail item is identified and the reading process is terminated, and if assignment to a single feature set is not possible, address interpretation (15.3) is performed, the results of which are then compared with the remaining feature candidates, in the event of a match of defined plausibility (22) with the address interpretation results for a single feature candidate, the associated mail item is identified and the reading process is terminated.
4. Method according to claim 3,
   characterised in that
   the address readers possess primary readers (5, 15) and secondary readers (7, 17) which are activated consecutively as required and that during the re-identification process the sequence described in claim 3 is also carried out using the secondary reader or readers (17) if automatic identification cannot yet be achieved with the primary reader (15).
5. Method according to claim 1,
   characterised in that
   the characteristic features stored in the database (4) with the intermediate reading results and reading results/distribution codes are deleted if a certain decay time has been exceeded or the last processing step has been completed.
6. Method according to claim 1,
   characterised in that
   if automatic identification has not been completed successfully, only the images or address alternatives of the mail item candidates (13) are manually coded at video coding stations by means of selection (18).

Images