DE19533585C1 - Character segmentation method - Google Patents

Description

The invention relates to a method for segmentation of characters and is applicable especially to Supplement to OCR (Optical Character Recognition). A preferred field of application of the invention is automatic attribution of documents in archives systems.

It is known, texts, images in paper form and graphics using so-called OCR programs Capture scanners and store them in the memory of computers lay.

Essential criterion for the applicability in the daily practice is the recognition rate of the OCR.

In operation or when setting up operations a variety of technical or commercial arises Documents within a set  Classification system can be registered and filed. The All of these documents contain all data and serves as the basis of the company.

This is an important requirement for the documentation find documents quickly. For this it is necessary that the documents when fed into the Document system attributes, such as account number, Title, document number or similar in one appropriate systematics must be assigned. That I these attributes are located on the documents it is from great advantage, directly when digitizing the Paper documents the corresponding attributes automatically filter out, edit, with the help to interpret a text recognition program and into the Integrate documentation.

This process for automatically storing paper documents in the document system is called automatic attribution and is carried out as follows:
After digitizing / scanning the documents, they are first aligned. This means that any misalignments caused by scanning or copying are corrected.

Since there are several forms of the same form type, template for each form type all the necessary information that one Form is automatically assigned to the same form type becomes. Only then can the corresponding automatically Cut attributes from the form and the Documentation system be made accessible.

There is a form for which there is no template yet is created, this is as described above aligned and created another template.  

Now the recognition process of the attributes takes place with the Target a very high detection rate. The so far known methods of image processing lead to Detection rates necessary for effective, automatic Attribution are not sufficient.

US 4,635,290 describes a method which only a segmentation of characters for Non-proportional fonts, d. H. only for fonts with constant character width enabled.

No contour information flows for the determination of the segmentation positions. Alone from the Projection distribution function out and only using the values zero and greater than zero the positions (separation points) for segmentation determined.

Complete histories, i.e. H. if by getting lost areas enclosed in the ink are completely lost go, such as B. at the characters 'o', 'e' or '8', are not considered.

The simple value arithmetic can be particularly useful for longer words lead to large deviations because the Intervals increase rapidly with each operation. The constants used are also very different Font dependent, and they would therefore have to be for everyone segmenting character string can be determined. This limits the procedure to strings with a corresponding minimum length.

Another disadvantage is that no font enhancement is made, e.g. B. if the acute angle in Characters 'N' have run to roundings.

In the publication Proc. of the IEEE, Vol. 80, No. July 7, 1992, pp. 1079-1092 becomes a procedure described with which areas on pictures how e.g. B. buildings, people or objects can be recognized should be, a combination of threshold and Quadtree segmentation is used.

This procedure is for segmenting characters unsuitable in a string, in particular it can not to separate converged characters (touching characters) can be used.

The process is based on grayscale and color images and not aimed at binary images.

In the publication IEEE Trans. On PAMI, Vol. 16, No. 7, July 1994, pp. 689-700 becomes a procedure described, which mainly on manuscripts is specialized and a simplified Font thickness analysis used to separate characters. There are only vertical bonds between individual Character tries to resolve, d. H. horizontal Bonding between individual characters, e.g. B. two lines on top of each other, and gluing within characters, e.g. B. at the characters 'i', 'a' or 'e', are not resolved.

Complete histories, i.e. H. if by getting lost areas enclosed in the ink are completely lost go, such as B. at the characters 'o', 'e' or '8' are not considered.

The procedure is also not for proportional fonts designed because the font width is included as a constant and no characters can be separated in which a double bond, such as. B. at the characters '88' possible or a very thick bond, such as. B. at the characters 'HH' possible.

No font enhancement is made, e.g. B. if the acute angles in the character 'N' lead to rounding have passed.

The invention is therefore based on the object Creating methods of segmenting characters which can be realized with simple means, the The causes of errors in OCR are minimized and one enables effective, automatic attribution.

This object is achieved by the Features in the characterizing part of claim 1 in Connection with the features in the generic term.

Advantageous embodiments of the invention are in the Subclaims included.

A particular advantage of the invention is that Bonding within characters and / or between adjacent characters as one of the most common Causes of errors in optical character recognition entirely or can at least partially be resolved by

• - dark ink pixels with a high temperature and light paper pixels with a low one Temperature can be equated and
• - The spread of heat distribution as a criterion computationally for the leakage of the printing ink is processed, whereby
• - Conclusions about previously existing temperatures can be drawn from the existing temperature distribution
  and or
• - Craters in the form of funnel-shaped depressions when gluing within characters and / or between adjacent characters localized with image processing methods and saved and subsequently
• - the contours of the craters by melting be changed by the bottom of the crater and / or the walls of the craters are expanded.

The invention is based on Embodiments are explained in more detail.

Show it:

Figure 1 shows an example of ink leakage as a function of time.

Fig. 2 is a text / numerical example for melting of craters;

Fig. 3 shows another text example for melting craters.

The biggest problems with font recognition are generated by the sticking that occurs. These arise from the fact that the Press color runs out and thereby close adjacent text elements are connected. The process of phasing out is now going backwards in the Considered time to allow the gluing to to reduce.

With this procedure, the physical Analogy application that you have a black pixel with a high and a white pixel with a low Temperature identified. Ink leakage then corresponds to the spread of the heat distribution.  

This can now go back and forth in time considered and from a given temperature distribution inferred from the previously existing temperatures will.

A simulation of these processes is possible using the Heat conduction equation

u _t = c Δu

or by means of suitable modified diffusions equation where weight factors are dependent on the size of the gradient z. B.

u _t = exp (- | ∇u | ² / 2σ) Δu

can be inserted. At a time t <0 the ink runs out and at a time t <0 simulates the dissolution of the bonds.

These equations can be created by inserting Weight factors further modified and the respective Application to be adjusted.

As can be seen from FIG. 1, the thermal development equation can be used to simulate the temporal development of the runout of the printing ink.

A shows the phase-out model at a time t <0
B shows the original
and
in C and D is the result of the reduction at
t = -1 ZE (= time unit) and t = -2 ZE are shown.

Due to the backward development in time there are bonds solved completely or at least partially.

However, as can be seen, gluing can only then be completely solved if at the same time significant changes to the letters in purchase be taken.

A further increase in the detection rate by Dissolution of bonds, which are next to the leak the ink also through the contrast setting and Interpolation can be caused by the scanner, is done by recognizing and editing so-called craters.

If you take a closer look at the bonds, you can see that they are not present evenly, but increasingly on acute-angled contours of characters such. B. at the letters "N" and "W" according to FIG. 3 and between closely adjacent written elements within one or different letters such as. B. at "i" in Fig. 2 or at "H", "N", "90" and "-N" in Fig. 3rd

However, it is striking that almost always one funnel-shaped depression remains. this will referred to as a crater. To eliminate the for example from the bleeding of the printing ink resulting errors must be traced back. It it should be noted that in some places Bonding thicknesses that have to be cut even exceed the font thickness, d. H. one takes one Melting off from the contour before, some must Make deep and on up to twice the font thickness other parts of the contour are not melted. So a procedure is necessary that is strong and precise can make localized interventions.  

Because the crater is strong in shape and size distinguish is a variable representation of the Crater necessary. This can be done by:

• - Reference pixel, pixel of the crater for localization of the crater in the big picture
• - Height, vertical extension from the crater floor to upper edge of the crater in pixels
• - Distances, horizontal for each crater height Distance of the left and right rim of the crater from Reference pixel
• - Contour information, for example which the steeper side of the crater

Starting from the reference pixel, it is examined whether there is a crater floor (closed floor and left and right a wall), then height by height examines whether left and right - within the framework of set parameters - there is a crater wall.

If there is no crater wall or a maximum of investigating height is reached, is canceled. The Representation of a crater is saved.

If you compare the craters with each other, you can see that the more acute the angle of the crater walls is, the deeper the bottom of the crater must be melted in order to restore the initial state. The extreme case occurs with closely parallel writing elements, as in the character string "HN", between the letters, as can be seen in FIG. 3.

Accordingly, depending on the recorded Representation of the crater (e.g. height) and one of the parameter set to be adapted to the respective font Melted crater floor. Then the The walls of the crater melted out. When entering as well  Melting takes into account the contour of the crater, for example on the steeper crater wall side the procedure is different from that of the opposite one Page. The melting processes are as follows Simulated equations:

Crater condition: | f _* D _a ϕ _K | <T,
Melting process: f new = (ff _{χ U (K)} ) + T _x ϕ _{K ′} , with ϕ _{K ′} = A (K, f _{χ W (K)} ) (D _a ϕ _K ).

Here, f denotes the typeface, D _a the operator who enlarges the crater by a factor (a <1) or reduces it (a <1) and ϕ _K the crater model function.

The crater condition is met if a crater of size a and shape ϕ _{K is} found in the typeface function f (pattern matching). The condition is already met if the crater match is greater than T.

The new typeface f _new arises from the fact that the found crater in the typeface is replaced by the adapted crater K '. The adaptation operator a depends on the crater and the surroundings U (K) in the typeface.

Because most commercial OCR programs don't Provide feedback; d. H. they give no percentage likelihood that they will have recognized individual characters, it is generally not possible, only on characters with insufficient Detection probability to make changes that  leave other characters unchanged. With everyone Therefore, not only the number of procedures characters recognized correctly due to the editing be considered, but also the number of characters, the side effect that can never be completely excluded are recognized incorrectly. The relationship between these two Figures provide information about the overall improvement of the Detection rate.

Claims (6)

1. A method for segmenting characters, wherein paper documents are digitized by scanning and subsequently saved, characterized in that adhesions within characters and / or between adjacent characters are completely or partially resolved by

• - dark ink pixels with a high temperature and light paper pixels with a low temperature are equated and
• - The spread of the heat distribution is processed as a criterion for the leakage of the printing ink, whereby
• - conclusions about previously existing temperatures can be drawn from the existing temperature distribution and / or
• - Craters in the form of funnel-shaped depressions in the gluing within characters and / or between adjacent characters using methods of image processing localized and stored and subsequently
• - The contours of the craters are changed by melting, by expanding the bottom of the craters and / or the walls of the craters.

2. The method according to claim 1, characterized in that by means of the heat conduction equation u _t = c Δu or by means of a suitable modified diffusion equation, in which weight factors as a function of the size of the gradients z. Bu _t = exp (- | ∇u | ² / 2σ) Δue are inserted, the modification causing large gradients (= essential edges) to be retained,
for t <0 the ink runs out and
simulates the dissolution of bonds for t <0
becomes. 3. The method according to claim 1, characterized, that as a parameter the crater in each case Reference pixel, the height in pixels as a vertical extension from the crater floor to the upper crater rim and the horizontal distance for each height of the left and right crater rim from Reference pixels are captured. 4. The method according to claim 3, characterized, that additional contour information is recorded.   5. The method according to claim 4, characterized, that the contour information the steepness of the Characterize crater walls. 6. The method according to claim 1, characterized in that the melting of the craters takes place by melting the crater floor and / or melting the crater walls, the melting processes being simulated by the following equations: Crater condition: | f _* D _a ϕ _K | <T ,
Melting process: f _new = (ff _{x U (K)} ) + T _x ϕ _{K ′} , with ϕ _{K ′} = A (K, f _{χ U (K)} ) (D _a ϕ _K ), whereby
f the original typeface,
f _new the new typeface,
D _a the operator,
ϕ _K the crater model function,
K the crater,
K 'is the adapted crater and U (K) is the surrounding area.