WO2007022880A1

WO2007022880A1 - Method for identifying mailings that are to be sorted

Info

Publication number: WO2007022880A1
Application number: PCT/EP2006/007946
Authority: WO
Inventors: Svetlozar Delianski; Katja Worm
Original assignee: Siemens Aktiengesellschaft
Priority date: 2005-08-26
Filing date: 2006-08-11
Publication date: 2007-03-01
Also published as: DE102005040662A1

Abstract

Disclosed is a method for identifying mailings that are to be sorted. In the inventive method, the mailings are grouped after extracting a feature that is conditional upon a minimum recognition effort. Several features such as global mailing features and/or local features on the mailing can be used as feature data sets. Said features are selected strategically for identification purposes also regarding the discriminatory capacity thereof such that a mailing can be identified or rejected faster during different sorting phases as the mailing is easy to recognize. If several features are used, features that are easy to recognize are used first, whereupon more complex features are used as required. Using a plurality of features also increases the quality of the identification process.

Description

description

Method for identifying items to be sorted

The invention relates to a method for identifying items to be sorted according to the preamble of claim 1.

In individual steps of a mail sorting, a very wide variety of information, in particular delivery information, is extracted from the items to be processed, which are required for their processing. These are in particular information about the delivery location of the shipment. Some of this information is not only needed once, but several times at different sorting levels. To be able to process shipments efficiently, this information is stored in a database after the first capture. In order to be able to reassign this information later to a shipment, it is printed with a unique ID code (ID = identification), such as a barcode. Using this ID code, stored data can be reassigned to a shipment at any time. A prerequisite for this is e.g. a barcode reader that determines the ID code of a shipment and uses the code to retrieve the associated information from the database.

The use of bar code readers causes high one-time but also high running costs in a mail sorting machine. On the one hand, barcode printers and readers are needed, which have to be regularly maintained and filled with ink. On the other hand, packages packed in plastic must be detected with special detectors and provided with labein in order to enable the application of barcodes.

The reading of delivery addresses in different sorting levels for processing all shipments is very time consuming and because of the high amount of mail to be sorted in practice unthinkable.

Typically, the volume of mail is very high, resulting in a large amount of delivery information to be stored and thus different shipments. Secure identification within e.g. a national search area is almost hopeless. In particular, broadcasts often occur which differ only slightly in their structure and are thus easily confusable.

In EP 1 222 037 B1 a method is known which effects a first restriction of the search space for items and thus facilitates further identification of items to be sorted by means of an image-based method. A physical reduction of the search space is used in a post-sorting process.

The invention has for its object to provide a fast and secure method for the identification of shipments.

The process should also ensure the highest possible quality sorting properties.

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

In this case, a method for the identification of consignments to be sorted is proposed, in which according to the invention the consignments are grouped according to an extraction of a feature which is conditional / characterized by a minimum recognition effort and by a high discrimination capability with respect to the current search space.

In other words, a feature is adaptively selected in such a way that, compared with other possible useable features and their recognition effort, the recognition of this feature is tailored to This results in an associated search space restriction.

In addition, instead of a feature, several features of the broadcast can be extracted, which are characterized by a minimum detection effort and a high discrimination capability. As a result, the quality of the identification is significantly increased or secured during sorting. Due to the low cost of recognizing these features, the method of the invention remains rapid despite the variety of features that are readily utilized.

In the search space limitation of the programs, the identifiable features are hierarchically arranged depending on the recognition effort and its discriminating capability, such that first a first feature with low detection effort and high discrimination capability and then a second feature with higher detection effort and high discrimination capability with respect to the limited search space for Identification of the shipment is used.

The first feature may include at least one global broadcast feature, preferably information on color, shape, size, texture, texture of the broadcast and / or other broadcast-dependent characteristics. This will e.g. a first group-wise search space restriction in the sorting of

Deliveries are quickly guaranteed, as well as a first allocation and registration of these groups set. Furthermore, a single feature such as a barcode or an address for delivery in different sorting levels is initially not used. This will be one or more

Features that additionally enable rapid recognition of mail items with simultaneous search space restriction. Furthermore, the second feature cited above may comprise at least one feature arranged locally on the shipment, which now refers to a region of the shipment, preferably a stamp, a barcode, a stamp imprint, a text field, a logo, a symbol, an advertising printout , a picture, etc. The second feature is the final identification of the broadcast or its rejection.

In particular, the second feature is used to singulate transmissions sorted in groups by means of the first feature. By the term "a first and second

Feature "are also to be understood as a" first and second group of features ". For example, first the color and the form of the shipment as a first feature and then the stamp and a printout are recognized as a second feature, since these features strategically lead to the optimal identification of grouped or individual shipments or reject shipments.

As a first and / or second feature also a feature data set (or vector as described below) can be used.

For storing and assigning a recognizable first feature used, a simple feature vector is used in which e.g. the assignment is made as a function of the recognition effort and the ability to discriminate (rising / falling).

For this, the recognition effort and the discriminability must also be determined as a function of an adaptive strategy protocol in which features relevant to a search space are used and features that are not or are poorly discriminated are excluded for this search space. Such strategy logs are determined for the current search space and retrieved and used in the identification of mailings in later sorting phases.

In particular, e.g. the second features are entered in a marked and structured graph, which additionally reflects the position of the features as well as possibly further discriminating features.

Therefore, a new technology can be used, which allows the assignment of a shipment to the associated stored information without the use of a specially printed ID code. For this purpose, unique broadcast and / or pictorial features are extracted from the program. These serve as keys for the associated database entry in the sorting process and are thus intended to guarantee the assignment of the shipment and shipment data. In the following, this key is also called a feature data record.

In the first sorting step, pictorial features are extracted for each shipment and stored in the database in parallel to further extracted shipment data. If the characteristic data sets of all programs to be processed are recorded, they are analyzed with regard to their recognition effort and their discrimination expenditure. In the process, an adaptive strategy protocol is determined for the entire search space, which specifies the use of the determined characteristics. If the extracted shipment data is to be reassigned to a shipment, then the feature data record must be calculated for the current shipment and the corresponding feature data record determined in the database. In contrast to an ID code, such a feature data record is subject to certain fluctuations. The feature data sets of different images but of one and the same shipment need not necessarily match exactly, but is ensured by the use of special distance metrics secure identification. In view of the high requirements of postal sorting, further advantages of the method are also ensured:

Robustness due to the use of several possibly redundant recognizable features,

high identification rates with extremely low detection rates,

- possible effective rejections of shipments,

Real-time capability, i. the presence of the identification result within a defined time of a few milliseconds,

- Use of features that do not exceed a certain storage capacity, as a prerequisite for the use of such a method in the mail sorting.

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

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

Show

FlG 1 feature data record of a program consisting of two types of recognizable characteristics,

FIG. 2 Grouping of consignments on the basis of selected features,

FIG. 3 Hierarchical structure of the identification method for the stepwise reduction of the search space.

FlG 1 shows a feature dataset of a broadcast consisting of two types of recognizable features: the global features featurel, feature2, ... featureN (eg texture, color). be, size, etc.) and the local features roiO_featureO, ..., roiO_featureM, roil_featureO, ..., roil_featureM, roi2_featureO, ..., roi2_featureM, roi3_featureO, ..., roi3_featureM in here four strategically chosen areas roiO, roil, roi2, roi3 of the broadcast. In FIG. 1, one stamp may be defined as one of the regions roi0, roil, roi2, roi3, for example, since this region is very easy to locate and recognize. The global broadcast characteristics and the local features form a characteristic, broadcast-describing feature data set. Between the local features there are the relations relation_i_j (i, j = 0, 1, 2, 3 with i ≠ j), over which the identification can be refined or better secured via linear combination or assignment of feature vectors of the local features.

In order to be able to identify large quantities of shipments safely and efficiently pictorially, the search space must be gradually restricted. The early exclusion of consignments from the processing process simplifies and accelerates the identification of a consignment and allows the use of complex methods for programs that are difficult to distinguish. In order to reduce the search space, generally easy-to-determine and robust features are initially used. Then the structure of a shipment can be analyzed more precisely.

For this reason, as shown in FIG. 1, the two types of features are distinguished:

one or more global broadcast features that contain general information about the color, shape, size, texture and / or texture of a broadcast and are easy to extract, and

- one or more local features related to specific regions of a program, describing them and their relations to one another. FIG 2 shows a founding of programs on the basis of selected features. By means of the global broadcast features feature4 and feature7, a quick reduction of the search space is to take place. For this purpose, programs that are similar to the selected features feature4 and feature7 are grouped into clusters clusterl and cluster2. Shipments within such a cluster can then be analyzed using complex methods. Of crucial importance for the identification of a program is, in addition to the selection of suitable features, the representation of the characteristics. Global features are represented by a feature vector. This representation forms the basis for methods for clustering or grouping and for establishing a hierarchy of identification methods.

The local features more accurately describe regions of interest of a shipment. Regions of interest can be of different types. In general, it is structurally or content-related areas of a broadcast image. These can be typical postal objects such as stamps, barcodes, stamp imprints, text boxes of any kind, logos, symbols, advertising imprints or else any other e.g. act pictorial objects. These regions may be represented by general pictorial features or by type

Characteristics are described. The aim is a type-oriented description for better differentiation of broadcasts. For example, text regions may use different characteristics than regions containing a symbol. In addition to this description, the location of the regions within the consignment and each other is an important feature. In summary, the structure of a program in a specific context can be described very abstractly but also in great detail with the aid of local features. An optimal representation of this type of feature provides a structural ruled and marked graph, which has already been clarified in FIG.

The coupling of both feature representations with the associated global and local features or feature vectors represents the feature data record of a program that describes it abstracted. For each shipment to be processed, such a feature data record must be determined and stored in a database.

With the help of the described features, a hierarchy of identification methods can be determined. First, the calculated feature data records of all shipments are considered. By means of special methods, the features or groups of features are examined for their ability to discriminate and sorted according to their meaning. By special methods, in particular cluster algorithms, a hierarchy of identification methods adapted to the current search space is generated. FIG. 2 shows by way of example the result of such a clustering.

This hierarchy is represented in FIG. 3 by a decision tree. Such a tree consists of marked nodes, edges and leaves with the following properties:

Node:

Each node method4, method7 in the tree represents a cluster according to FIG. 2 with identification of the global transmission features feature4, feature7, which is progressively refined or reduced as it passes through the tree. Furthermore, each node is coupled with an identification method by means of which the path to the next level of the tree can be determined.

Leaf: ^, ToHgS leaf cluster! + COItI ¹ ^ XSX IP.ethod, ClilS tSr2 + COITlpl2X ITϊSthod then represents the final search space and contains all shipments that are similar to the corresponding global features.

Each leaf is either linked to a complex identification strategy or represents a rejection re- ject.

edge:

Each node has at least one marked output intervalOO, intervalOl, intervalO2, ie. an edge that leads to the next refined search space or rejection re- ject.

In general, therefore, the identification of a shipment is attributed to a search in the calculated hierarchy. Such a hierarchy is always adapted to the initial search space. The result of this identification is either a reduced search space or a rejection of the current broadcast.

The final reduced search space now uses the complex identification strategy. For this purpose, the graph representation is used. The graph of the shipments to be processed must be compared with those of the eligible shipments and examined for their similarity. For this purpose, individual regions must be examined for their similarity, taking into account the storage relations. The similarity can then be determined with the help of special distance metrics. Such a complex strategy, which should identify a broadcast in the final search space, is not necessarily limited to the described method. A wide variety of features and identification strategies can be used, which are very detailed and thus describe a shipment very precisely. Finally, the calculated distances must be analyzed and

the.

In general, this approach allows use for the current search space of relevant features and the exclusion of non-discriminatory features. These can be different in different search spaces. There is thus always an adaptation to the current search space, which leads to a temporal and qualitative optimization of identification.

Claims

claims

1. A method for the identification of items to be sorted, which means that the items are grouped after an extraction of a feature, which is caused by a minimum recognition effort.

2. Method according to claim 1, characterized in that a plurality of features of the transmission are extracted, which are caused by as minimal a detection effort as possible.

3. The method according to claim 2, characterized in that the search space restriction of broadcasts, the features are arranged hierarchically depending on the recognition effort and their discrimination ability, so that first a first feature with low detection effort and high discrimination capability is used and then a second feature is used with a higher detection effort and the highest possible discrimination capability with respect to the search space reduced by the first feature.

4. The method of claim 3, wherein: the first feature comprises at least one global broadcast feature, preferably color information,

Shape, size, structure and / or texture of the shipment.

5. The method according to any one of claims 3 or 4, characterized in that the second feature comprises at least one on the shipment locally arranged feature, which is based on a Be rich of the shipment, preferably a stamp, a barcode, a stamp imprint, a text box, a logo, a symbol, an advertisement, a picture.

6. The method as claimed in claim 5, wherein the second features are entered in a marked and structured graph which reproduces the position of the features.

7. Method according to one of claims 3 to 6, characterized in that the second feature is used for the singling of group-sorted items or for the rejection.

8. Method according to one of the preceding claims, characterized in that the recognizable feature is stored and arranged in a feature vector whose arrangement is dependent on the detection effort and the ability to discriminate.

9. Method according to one of the preceding claims, characterized in that the recognition effort and the discriminating ability are determined as a function of an adaptive strategy protocol in which features relevant to a search space are used and features that are not discriminating for this search space are excluded.