DE19517178C1 - Image restoration using background cleaning process - Google Patents

Abstract

An electronic image obtained from a document is subjected to a process in which the information is filtered in such a way that only the foreground information is extracted. The generated information serves as an output image that is used for image restoration. The information is used as a reference image that is compared with the remaining image data in a process operating on adjacent image points. Typically the heavily filtered image has a number of line points (10) and many of these are eliminated in the comparison process.

Description

Field of the Invention

The invention relates to the separation of foreground information from background information in a template.

State of the art

For image or character recognition (Optical Character Recognition) in an electronically scanned and To be able to carry out digitized documents often has to a first preprocessing step, the Information of the foreground from the information of the Background. In doing so, provide the information of the foreground often the actual ones for which Character recognition image data to be viewed while the Information from the background can often be disruptive and also lead to misinterpretation of the foreground information can. Also, the data of the background of the picture often carry to reduce the recognition performance of the Foreground information at. The information of the background should therefore preferably before image or character recognition be removed. Foreground dates the one font should represent in many applications by the subsequent character recognition, preferably in ASCII data, being transformed.

The background data in an image mostly consist of the the following groups: Disruptions in digitization (like noise) and components that cannot be optically filtered out of the digitized document (e.g. dirt, Writings, stamps, etc.). To be the best possible Detection rate in a subsequent image processing To make the image data accessible after a Image processing preferably only contain the data that to be processed later by character recognition. Each  the better the image preparation before recognition, the more The recognition results can also be expected to be more favorable.

Many image processing units expect an input to the Grayscale reduced image, e.g. B. a binarized black and white Picture. To generate an electronically scanned image of readers (scanners) are often used for a template, those with a predetermined sampling frequency (e.g. 200 to 300 Pixels per square inch) images with a variety of Provide grayscale (e.g. 16 or 256 grayscale).

Due to a number of possible errors in the electronic reading of a template, such as Scanner tolerances or inadequate comparison of individual reading elements of a scanner, a number of Influencing factors occur that adversely affect image recognition can affect. Also the type and quality of the template itself can decisively shape image recognition. So go with a conversion of a color image into a gray image existing color contrasts lost. With handwritten, however the also depends on machine entries on a template Readability and recognizability, especially of theirs Contrast caused, for example, by more or less strong pressure when writing, from.

For a number of applications, there is a requirement in addition to the character recognition a so-called archive picture generate, for example to reduce paperwork. It So two images have to be created, which if possible should be done without loss of throughput. Of the Applications currently available for this are black and white Images, both for the recognition unit and for Archiving purposes. Grayscale images must be and have to be computationally compressed and decompressed still requires more memory than black and white images.

By electronically scanning a template, one obtains a electronic image of the template, in which each pixel  an image value, for example a gray value, a black or White value or a color value, is assigned. By using it these assignments can be influenced by filters. At For example, a threshold filter can be applied the image values above or below a given Filter threshold values from the electronic image. So can z. B. those not desired for image processing Eliminate background information or interference.

However, the filtered out image values are not available afterwards more available for image recognition. Perhaps unintentionally filtered foreground information can falsify the result of a subsequent image recognition. Each more or less go according to the value of the threshold Information lost. The choice of the threshold value comes one eminent importance for the quality of image recognition.

To generate binary images from grayscale images often static or dynamic conversion algorithms used. Color filters are often used for background cleaning applied or the template is with appropriately colored light illuminated. Classic filter operations (such as morphological) are used to remove selected structures (such as straight horizontal or vertical lines) are used to certain events in a certain environment (Document type) tailor-made.

From the book by Peter Haberäcker, "Digital Image Processing", Carl Hanser Verlag Munich Vienna, ISBN 3-446-14442-0, 1985, a number of image processing methods are known which can also be used for image preprocessing. In particular, procedures for the smoothing of gray values (section 8.1) and the use of difference operators (section 8.2) as well as processes for dynamic threshold value determination (section 12.3) are presented there as operations in the local area (section 8 ).

An overview of the best-known methods for determining a threshold is in P.K.Sahoo, S.Soltani and A.K.C.Wong  "A Survey of Thresholding Techniques", Computer Vision, Graphics and Image Processing 41, 233-260, 1988. It is there between histogram transformations methods that the form of a histogram to determine the threshold value change a gray value distribution and algorithms for Differentiated threshold value calculation.

Further processes are known from US-A 4,590,606 and US-A 5,038,381 for image processing, such as B. foreground filter Runtime behavior, known.

EP-A 0.505.729 describes a system for the binarization of Images that are reading documents with a normal Print quality enables.

J. M. White and G. D. Rohrer: "Image Thresholding for Optical Character Recognition and Other Applications Requiring Character Image Extraction ", IBM J. Res. Develop., Vol. 27, No. 4, July 1983, pages 400 to 411, describes effective filter algorithms for the Use in the optical character recognition of machine or handwritten documents.

The patent DE 38 26 285 C2, filing date July 30, 1988, describes a method for the detection of abnormal ones anatomical areas in a digital x-ray image, being from a stored digital image using amplification or Suppression filtering produces two filtered images and then by subtracting these two filtered images a difference image is generated, which the Basis for the final evaluation using represents predefined search profiles.

The patent DE 38 26 285 C2 represents the closest State of the art with regard to the present invention represents.  

All procedures with a filtering out of information However, an electronic template has in common that Can also filter information that is relevant for a Image recognition are necessary, can be lost.

Summary of the invention

It is an object of the invention, a method and an apparatus for the treatment of To create templates that separate the best possible Foreground and background information enables in particular a method and a device for pretreating templates to create an image processing system in one electronic image of a template as good as possible Allows recognition of information.

The objects of the invention are achieved by the independent Claims 1 and 16 solved. Further advantageous versions of the Invention can be found in the subclaims.

According to the invention, at least two electronic are first Images from a template with a different one Filtering, e.g. B. a threshold filtering generated. A picture should be filtered so that essentially only  Foreground information, or fragments thereof, is present therein are. This image serves as the starting image for a subsequent one Image restoration. The task of image restoration is, if possible many of those that may have been lost in this original image Get foreground information again. The remaining and serve less filtered images than the original image for this as reference pictures. It should be understood that as Reference image also an unfiltered image of the template or the Template itself can be used.

The image restoration happens in that, starting from Information in the output image, by comparing the Information of the original image with the information at least one of the reference images in the, the information of the Starting picture corresponding environment, the information of the Output image through the information of at least one Reference picture are added. You can do these supplements either as new information in the initial image be included, or, separately, together with the Information of the source image in a new result image be filed.

The selection of the information to be added from the Reference pictures are done with the help of a suitable one Selection schemes. The preferred selection schemes are a series of neighboring point considerations or others Criteria such as B. Apply information density. At a Use information density as a selection scheme for example, only reconstructed those pixels that are in a Environment larger or smaller a given pixel density lie. In this way, individual points of excluded from the restoration. The information density as Selection scheme can also be used in combination with others Selection schemes such as the neighboring point consideration used will.

Individual areas are included for a neighboring point consideration a certain size around such points (information points),  that have an information content in the initial image, both in the Output image as well as in each of the reference images accordingly created. Information contained in the corresponding Area of one of the reference images in addition to the Information points are present as new ones Information either added to the original image or otherwise together with the information of the original image filed. It should be understood that such pixels, their Information by filtering from the original template were filtered out, pixels without information content represent.

The size of the individual areas around the information points can be either the same for all information points or adaptive, e.g. B. according to a respective density of information will. With the choice of size and shape of the areas the degree of restoration affects each run. Will the Size of the areas selected small (e.g. only the immediate ones Neighboring points to an information point) in particular already existing ones, but through the filtering weakly displayed image content immediately supplemented. An in its thickness reduced line is thus reinforced. Will the The size of the areas selected too large may already be filtered out interference points can be obtained again without this is desired.

For certain applications it may be cheaper than Information points either only to consider those in the original image was originally there, or that by adding all new information to the initial picture included pixels to use as information points. The latter procedure can possibly be Restore fragmented lines more advantageously, however can "resurrect" unwanted areas. Added however, new information in an image does not necessarily have to again represent new information points.  

As a template of the electronic pictures can also be taken from the Original image modified images, such as individual color filtered Use pictures. This allows for multi-color images Restore or remove color-coded information individually a comparison with reference images of different colors gain suitable knowledge for the restoration.

The method according to the invention makes it possible to Archive image and an image for character recognition by one To generate a scan. Any scanner can do this can be used as long as there is a slight gray value difference between the background data and the foreground. The Detection rates can be achieved using the method according to the invention considerably compared to a simple, not restored Enhance black and white image. Character-based recognition rates could be increased from 75% to 81% in one embodiment.

The inventive method can preferably be as Program (software) in electronic data processing to implement.

Description of the drawings

To explain the invention in more detail, the following are Embodiments with reference to the drawings described. Functionally identical elements should be the same Wear reference numbers.

FIG. 1a shows a detail of a weakly filtered reference image of an original;

FIG. 1b shows the section corresponding to FIG. 1a of an output image of the original that has been filtered so strongly that only fragments of the foreground information, here of the line, are contained therein;

FIG. 2a-b show the neighboring point restoration with a 3 × 3 pixels containing restoration portion;

Fig. 3a-b show the corresponding adjacent point restoration with a 5 × 3 pixels, and a 4 × 3 pixels containing restoration portion;

Fig. 4a-f show a sequence of successive restoration;

FIG. 5a-d show an exemplary application of the invention.

Detailed description of the invention

For an inventive separation of foreground and Background information is first in one embodiment two electronic images from one template each different information content generated. this happens preferably by separating two black and white images from one grayscale image the template with different threshold values be generated. The first of the black and white pictures is said to be the Background data essentially no longer included while the second black and white image essentially all of the image data should contain, including the background data. Exemplary could be a static first threshold for the first Apply approximately 50% black and white image, d. H. Information with a gray value above 50% of a maximum gray value lost. A static second threshold could correspondingly 90% for the second black and white image. However, dynamic threshold values are advantageous related. It is to be understood that the second picture also without threshold filtering can be used immediately.

When the first black and white image (in but essentially without background data) may also work lost a number of data that belong to the foreground and are required for character recognition. The first Black and white image is now the starting image for one Image restoration used with the aim of as many of the in foreground information lost in this image to obtain. The second black and white image serves as Reference image. It is particularly suitable for image restoration  a neighboring point consideration as described below is.

The neighboring point consideration is based on the fact that, starting from the information in the original picture, which is essentially that represent foreground information or fragments thereof by comparing this output image information with the corresponding information of the reference image, the Information of the original image by the information of the Reference picture using a suitable comparison scheme be supplemented. Those that no longer exist in the original image In this way, information can be restored.

Individual areas are included for a neighboring point consideration a certain size around such dots in the original image have an information content (information points), both created accordingly in the output image as well as in the reference image. Information in the corresponding area of the Reference image, but not in the original image, will appear as new information in either the original image supplemented or otherwise, e.g. B. in a new result picture, filed together with the information of the original image. It should be understood that such pixels, their Information through the (threshold) filtering from the original template were filtered out, pixels without Represent information content.

The size of the individual areas around the information points can be either the same for all information points or adaptive, e.g. B. according to the respective information density will. Also, it can either be for a particular application be cheaper than information points only to those consider that originally present in the initial image were, or also by the additions as new Information included in the original image as pixels Use information points.  

So that the restoration process multiple times for one picture or Tile is repeatable, will be necessary for any second run uses the image just created. To a Restore unwanted background data, which by the described process is also possible to prevent are preferably isolated pixels from the reference image neglected by the size during the restoration the areas are not chosen too large. However, this is only for a first cycle over the picture necessary because isolated Pixels are already eliminated there.

It is to be understood that the invention is not limited to use of black and white images, but that any Images with different information content can be used. Also can instead of two pictures additional images can be used so that any Image property, e.g. B. horizontal or vertical lines, Background patterns, etc., extracted and later to the one you want Result picture can be restored or combined.

An embodiment of a neighboring point treatment is shown below. It should be noted that, to illustrate a possible application of the invention, the pixels B2, C2, B3, C4 and B5 are intended to represent a line course, while the pixels F1 and E3 are to be attributed to accidental contamination of the original. FIG. 1a 10 shows a detail of a weakly filtered reference image from a, not shown, template. FIG. 1b shows the corresponding section 100 of an output image of the original that is filtered so strongly that only fragments of the foreground information, here the line, are contained therein.

The coordinates AF shown indicate the horizontal position and the coordinates 1-5 the vertical position of a pixel. As can be seen from FIG. 1a, the detail 10 of the reference image has the pixels F1, B2, C2, B3, E3, C4 and B5. The corresponding section 100 of the output image in FIG. 1b now only contains the pixels B2 and C4. The pixels F1, C2, B3, E3 and B5 which are missing in FIG. 1b have been lost by appropriate filtering.

A respective restoration area is then created around each of the pixels of the output image FIG. 1b, which contains information, that is to say information points B2 and B5, both in section 10 and in section 100 . In addition, information points that span a restoration area are to be marked with an x in the figures for better recognition. The Fig. 2 show the neighboring point restoration with a 3 × 3 pixels containing restoration area. The Fig. 3 show the corresponding neighbor point restoration with a 5 × 3 pixels, and a 4 × 3 pixels containing restoration area.

It can be seen from FIG. 2a that the restoration area containing 3 × 3 pixels around the information point B2 also includes the pixels C2 and B3. The restoration area around the information point C4, likewise containing 3 × 3 pixels, on the other hand comprises the pixels B3 and B5. These pixels C2, B3 and B5, as newly restored pixels, are now either included in the output image according to FIG. 2b or are stored separately together with the information points B2 and C4. From FIGS. 2a and 2b it can be seen that by selecting the restoration areas containing 3 × 3 pixels, the line B2, C2, B3, C4 and B5 could be completely restored, the more distant pixels F1 and E3 (e.g. the random impurities) were filtered out.

It can be seen from FIG. 3a that the restoration area containing 4 × 3 pixels around the information point B2 also includes the pixels C2 and B3. The restoration area containing 5 × 3 pixels around the information point C4, on the other hand, includes the pixels E3 in addition to the pixels B3 and B5. Thus, by choosing a larger restoration area than in Fig. 2 in addition to the line B2, C2, B3, C4 and B5, the not immediately adjacent pixel E3 is also restored and the newly restored pixels can either be included in the original image according to Fig. 3b or filed separately with information points B2 and C4.

The examples in FIGS. 2 and 3 clearly show that the choice of the size of the restoration area plays an important role. However, it should also be emphasized that individual, random impurities that are not in the immediate vicinity of the main information of a template can be well suppressed by the method according to the invention, even in relatively large restoration areas.

If the newly restored pixels are included in the original image and thus form a new original image for a further restoration process, it is particularly advantageous to restore contiguous lines. A successive restoration can be carried out, for example, in such a way that the restorations around all information points originally present in the starting image are carried out simultaneously in a first step and the newly obtained information is written into the starting image. This supplemented initial image, in which the supplemented image points now represent new information points, can then serve as the initial image for a further restoration step and so on. It should be understood that a new restoration step only has to be carried out with the newly received information points from the previous restoration step. FIGS. 4a-f show successively as a sequence such restoration.

FIG. 4a shows the reference image and Fig. 4b, the corresponding original output image. Starting from the information points from FIG. 4b, as shown in FIG. 4c, a first restoration is carried out with restoration areas containing 3 × 3 image points around the respective information points. The pixels thus obtained are written into the output image and the modified, original output image thus obtained is used as the new output image ( FIG. 4d). A further restoration step, starting from the newly acquired information points, results in further information points ( FIG. 4e), which are rewritten in the original image ( FIG. 4f). The image shown in Fig. 4f is the result of the successive restoration. It should be noted that the continuous line, formed from the pixels F2, E2, D2, D3, C4, D5, ES, E6, D7, C8 and B7, was completely reconstructed, while individual points, such as points A2, H5 , F9 and I9, were not reconstructed and thus eliminated.

Another successive restoration can, for example performed that the restorations on one original information point (e.g. a corner point) begins. The new through the restoration around the first information point Information obtained is immediately in the initial picture written. This supplemented initial picture in which the supplemented Pixels now represent new information points, then serves as a starting picture for the next information point and so on further. For example, a restoration can be done Carry out recursively line by line. It should be understood that this line-by-line process may have to be run through several times got to.

FIGS. 5a-d show an exemplary application of the invention. Fig. 5a shows the template for the inventive restoration process. FIG. 5b shows a binarized black / white image of the template from FIG. 5a filtered with a threshold value of approximately 90%, which is to serve as a reference image for the restoration. FIG. 5c shows a black and white image of the template from FIG. 5a filtered and binarized with a threshold value of approximately 50%, which serves as the starting image for the restoration. Fig. 5d shows the result of the carried out with the aid of the reference image restoration of the original image. It can be seen from this that the method according to the invention enables a suitable separation of foreground and background information, so that subsequent image processing steps can be carried out successfully.

Claims (17)

1. A method for separating foreground information from background information in a template, wherein
in a first step
at least two images of the template are generated with different filtering,
characterized in that
a first of the images is filtered such that essentially only foreground information, or fragments thereof, are present therein,
this first filtered image serves as the starting image for a subsequent image restoration, and
the at least one remaining, second filtered image represents a reference image for the image restoration; and
in a second step
the image restoration is carried out so that as much as possible of the foreground information lost in the original image is restored again, whereby
the image restoration is accomplished such that, starting from information in the output image, by comparing the information of the output image with the information of the at least one reference image, the information of the output image is supplemented by the information of the at least one reference image. 2. The method of claim 1, wherein the selection of to supplementary information from the at least one Reference image in the second step by a Neighbor point consideration is carried out. 3. The method according to claim 2, wherein for the Neighbor point consideration of individual areas with a certain size around such points (information points), that have an information content in the initial image, both in the original image and in the at least one Reference picture are created accordingly, and in the Information in the relevant area of the at least one reference image in addition to the  Information points exist as new information be supplemented. 4. The method of claim 3, wherein the sizes of each Areas around the information points for everyone Information points can be defined identically or adaptively. 5. The method of claim 3, wherein the size of an area through the immediate vicinity of each Information point is set. 6. The method according to any one of the preceding claims, where the selection of the information to be added from the Reference images in the second step through an analysis the density of information is carried out. 7. The method according to any one of the preceding claims, where the additions in the second step as new Information to be included in the original image. 8. The method according to any one of the preceding claims, where the additions in the second step along with the information of the original image in a new one Result picture can be filed. 9. The method according to claim 7 or 8, wherein the supplements are considered as new information points. 10. The method according to claim 9, wherein a successive Complete the initial image by selecting and Filing the information to be added and a reselect and save those to be added Information from the newly received Output image is performed. 11. The method according to any one of the preceding claims, where the reference image is an unfiltered image of the template is.   12. The method according to any one of the preceding claims, where the template is a modified from the original image Image, preferably a color filtered image. 13. The method according to any one of the preceding claims, wherein the filtering is threshold filtering. 14. Use of the method according to one of the above claims for preprocessing the Submission before a subsequent image processing. 15. Use of the method according to one of the preceding claims in a program for a electronic data processing.   16. Device for separating foreground information from background information in a template, showing: first means for generating at least two images of the template, each with different filtering, wherein
a first of the images is filtered such that essentially only foreground information, or fragments thereof, are present therein,
this first filtered image serves as the starting image for a subsequent image restoration, and
the at least one remaining, second filtered image represents a reference image for the image restoration; and
second means for image restoration, whereby
the second means restore as much of the foreground information lost in the original image as possible, and
the second means have a comparator which, based on information in the output image, by comparing the information in the output image with the information in the at least one reference image, the information in the output image is supplemented by the information in the at least one reference image.