How to use the IBM Datacap Image Enhancement ruleset

Content

1.    Introduction and goal
The Image Enhancement ruleset contains image processing operations that can be run on images within a Datacap batch.  It is possible to run the same enhancements on all images, or different enhancements on different sets of images.

There can be many reasons why images should be processed with image enhancement actions.  The most common reason is to improve the quality of the text to aid the recognition engine.  If you are experiencing reduced recognition accuracy and available engine settings are not resolving the issue, then image enhancement is the first “go-to” tool that should be used to improve recognition. Some image adjustments can help even if the image already is of good quality.

If performing recognition with Datacap, it is strongly recommended to read the document “Best Practices for optimal text recognition in IBM Datacap” in addition to this guide.

The Image Enhancement ruleset is an interactive way of configuring image enhancements.  This guide not only describes some of the key features in the ruleset, but it also helps explain how to use the ruleset to properly configure enhancements for different page types.

Recognition cannot be guaranteed to be 100% accurate.  By using image enhancement, it can improve accuracy.

2.    Using compiled rulesets
The Image Enhancement ruleset is an interactive “compiled” ruleset that allows viewing of an image so the adjustments can be previewed.  All compiled rulesets have similar behaviors but they are configured much differently from traditional rulesets.

2.1.    The difference between compiled and traditional rulesets
A “compiled” ruleset is different from a traditional ruleset in Datacap Studio.  A “traditional” ruleset allows full configuration of what actions are performed.  The actions called and the flow of the logic is all controlled by the application developer,  such as, creating each rule and function, adding the required actions, and fully controlling the order of the actions. Individual actions can be added and removed by the developer in a traditional ruleset.  A traditional ruleset can only be created and edited in Datacap Studio.  A compiled ruleset can be configured by using Datacap Studio or FastDoc.

In a compiled ruleset, a configuration panel is used to configure the rules.  The selections and settings in a panel are unique for each compiled ruleset.  A user can open the configuration panel and make selections that determine the work that is to be performed by the rules.  After the configuration settings are set, the user can save the configuration.  The configuration then automatically creates the necessary rules to perform the selection when running a Datacap batch.  The user does not create any of the rules, and cannot alter the rules in any way other than changing the selections in the configuration setup panel.

While it is easier to create the rules, they cannot be edited as traditional rules. As an aside, it may not be obvious that when using a compiled ruleset, the generated ruleset is visible in Datacap Studio.  It is possible to configure settings in studio, then copy the ruleset and paste it back as a traditional ruleset that can then be edited as normal.  However, when the ruleset is copied, the ruleset copy is now disconnected from the configuration panel and can now only be changed by manually updating the rules.

2.2.    Displaying a compiled ruleset setup panel
2.2.1.    Datacap Studio
In Datacap Studio, the available rulesets are shown at the bottom of the “Rulesets” panel in the “Global rulesets” section. Right-click a ruleset and select Install In Application.  This copies the ruleset to the rules directory of the application and it can now be configured.  A ruleset can be added only once.  All configuration settings are unique for the application.  If the same ruleset is installed in two applications, the settings are unique for each application.

After the compiled ruleset has been installed, it can be configured by right-clicking the ruleset and selecting Settings.

When right-clicking a ruleset to install it in an application, there is a second selection called “Install As Global”.  This selection makes the ruleset global to all applications.  This means that if a ruleset is installed as global for two or more applications, the settings made applies to all applications.  As an example, if you are configuring the ruleset for application “A”, any ruleset changes are also instantly applied to application “B”.  There may be times when this can be useful, however, it is generally not the recommended way of using the rulesets.  The drawbacks to this approach are that by making changes to application “A”, it changes application “B”, possibly causing it to run in an undesired way.  It is also more difficult to roll out production changes to global rulesets.  If you roll out changes to application “A”, you are also changing application “B” simultaneously, making the process more difficult to stage and test.  It is for these reasons that it is recommended that compiled rulesets are not used globally, but instead installed to the application. When installed in the application, each application has independent settings.  Additionally, global rulesets are not supported by FastDoc.

After a ruleset is installed to an application, lock the application’s profiles in the Task Profiles tab and add the ruleset to the appropriate task in the location from where it should be run when rules are processed by Rulerunner or another Datacap client.

2.2.2.    FastDoc
Compiled rulesets are added to an application in FastDoc by using the workflow panel.  Pick the desired compiled ruleset shown on the right of the screen can be dragged into the appropriate position in the application’s workflow.  This has the same effect as selecting Install In Application from Datacap Studio and attaching it to a profile.  Be aware that FastDoc does not support the Install As Global feature that is available in Datacap Studio.

In FastDoc, when a ruleset is installed in an application, the panel can be edited by selecting the ruleset name in the drop-down list on the configuration panel.  When the ruleset is selected, the configuration panel is displayed on the right with the DCO tree on the left.  As when editing the panel in Datacap Studio, selecting the DCO node on the left controls the settings on the right.  

In both clients, if a DCO node selected in the tree is not appropriate for the settings in the panel, then the settings may appear as gray and unselectable or may be automatically hidden.  The displayed controls and selections in the compiled ruleset panel can change dynamically as different DCO nodes are selected.

2.3.    Configuring a compiled ruleset
A compiled ruleset typically operates on specific DCO levels - Batch, Document, Page, or Fields.  It may configure only one level or multiple levels.  An aspect of ruleset configuration is that the object that is configured is dependent on selecting the correct DCO object in the DCO tree.  In Datacap Studio, while a ruleset panel is displayed, selecting nodes in the DCO hierarchy on the left updates the panel with the selected node.  The DCO tree does not need to be locked for this configuration to be performed.  However, do not add or delete nodes to the DCO tree while the panel is displayed.  In FastDoc, selecting the DCO node on the left side of the configuration panel updates the displayed ruleset panel and the settings are applied to the selected node.

For example, on a recognition ruleset, when a page node is selected in the tree, the panel configures the settings for that page type, and when a field is selected, it configures recognition for that field type.  If a ruleset only works at the batch level, such as virtual scan, then the selected level may be ignored and the settings configure the entire batch regardless of the selected DCO node.  

For the Image Enhancement ruleset, a page DCO must be selected to make configurations and the image enhancement selections apply to that page type.

The compiled Image Enhancement ruleset provides feedback indicating if the selected type is supported.  In the following Datacap Studio screen shot, the “Invoice” document DCO node is selected.  This type cannot have image enhancements configured.  Because of this, options are disabled.  The panel indicates “Invoice” is the selected DCO node and it also states that it is an “Unsupported Type”.

Datacap Studio DCO tree “Invoice Selected”

When a page type is selected, then the settings become enabled.  The panel shows the selected DCO as “Other” and it also says “Configured”, which means there are saved configuration settings.  Exactly how each compiled ruleset displays information can be slightly different but the overall operation is similar.

Datacap Studio DCO tree “Other” page type selected

3.    Configuring the Image Enhancement ruleset
The Image Enhancement ruleset runs a sequence of image adjustment operations on a single page image.  If multi-page images are ingested, they should first be split into single images.

3.1.    User-defined groups of settings
Like other rulesets, the configured options apply to specific DCO page nodes.  The Image Enhancement ruleset also optionally allows grouping of settings into user-defined categories. There is always an initial category called “Default”.  If not changed, all selections are saved in this category. For example, a user-defined category called “Line Removal” can be created, or any other name, then each category has unique image settings.  It is actually these categories that are then assigned to page types, making this an aspect of the ruleset that some may find slightly confusing at first.  To say it another way, the “default” has a set of operations, the “default” set can be assigned to say, type “Main_Page” and “Trailing_Page”.

If all pages are processed with the same settings, then there is no need to create custom rule settings. If different rule settings are required, then it is necessary to create additional settings. For example, if some pages should have line removal, despeckle and close and other pages only get processed with line removal, then multiple rule settings are required.

This means it is possible to create an additional set of operations, give it a unique name and assign that set to other types of pages.

3.1.1.    Applying settings to page types
When initially displaying the panel, the panel indicates if the DCO node is configured. In this example, for the TravelDocs application, the Other page node is selected.  The default rule setting is displayed and the panel shows some configured settings, such as Deskew.  However, the panel also says “Other (Not Configured)”.  That means this node has not been configured and the “default” settings do not take effect until the Save button is clicked.  Until the settings are saved, the “default” settings are not be applied to pages of type “Other”.

Initial selection of a new DCO page appears as “Not Configured”

After you click the Save button, the “default” configuration is now associated with the page type “Other”, meaning that when rules are eventually run by Rulerunner or another client, all pages of type “Other” have the operations configured in the “default” set applied.  Type “Other” is now associated with the “default” settings.

After you click Save, the panel shows type Other is “Configured”

Selecting another new page node that has never been selected again shows the settings as “Not Configured”.  Although, the panel shows “default” and the settings are set for the “default” group, they are not applied to this page type.

 “Rental_Agreement” has not been configured

In a single session, the panel remembers the nodes that have been visited in the DCO tree.  Now, without clicking Save, if “Optional_Insurance” is selected, it says “Optional_Insurance” is “Not Configured”.

“Optional_Insurance” is not configured

Now, if the Save button is clicked, the “default” set of enhancements are associated with “Optional_Insurance” and it also associates “default” with “Rental_Agreement” types since that was also visited in the same session.

After clicking Save, the node is now configured to “default”

However, now, after clicking Save, if “Rental_Agreement” is re-selected, it shows that this page type too is now associated with “default” because it was selected during the same session.  This is an important point that might not be obvious. If a node that has not been previously configured is “visited”, it is remembered and when the Save button is clicked, the node has the “default” selection associated with it.

Clicking Save associates all visited nodes as Configured for “default”

3.1.2.    Using multiple configuration settings based on page type
It is possible to perform different sets of enhancements based on the page type. For example, if some files are ingested from a known location that are known to be clean input pages, such as converted word documents, they can be assigned a specific page type, let’s assume the page type is “Word”.  Then assuming there are other pages ingested from a scanner and these pages can be skewed or speckled, they can initially be assigned a page type of “Scanner”.  Now the “Word” pages and “Scanner” pages can have unique enhancements performed.

This scenario can be configured by first selecting the “Word” page type in the DCO.  In the enhancement panel, update the selections so only the Line Removal feature is enabled with the desired settings. All other settings can then be disabled. It is not necessary to delete all of the settings from the panel.  If the settings are not checked, then they are not enabled. Then click the Save Settings As button and type something like “Remove Lines Only” as the configuration name and then save.  The panel now looks like this:

Page type “Word” configured to use “Remove Lines Only” configuration set

This means the selected DCO page node called “Word” is configured to run with the “Remove Lines Only” set of settings.

Now, select the “Scanner” page node and select the settings desired for the scanner.  This can still enable “Remove Lines”, but you may now also want to also enable deskew or despeckle, to make the scanned documents as clean as possible for recognition.  Click Save Settings As and enter a name such as “Improve Scanned Images”, and click Save and the panel looks like this:

Page type “Scanner” configured to use “Improve Scanned Images” configuration set

Now, the “Scanner” page type is configured to use the “Improve Scanned Images” setting. By clicking back to the “Word” object in the DCO, the panel changes again to show that “Remove Lines Only” is the configured setting for “Word” page types.  The operations also have their selections changed to show the state for the current setting. Now when rules are run, each page type runs with their unique settings.  

Additional sets of configured settings can be created. Different page types can also share settings. For example, if you had a page type called “Bill”, by selecting the DCO node, either the “default”, “Remove Lines Only” or “Improved Scanned Images” selection can be picked to run on this page type.

3.2.    Interactive setting configuration
The compiled Image Enhancement ruleset allows viewing of a sample image while selecting and adjusting values. Select the Open image file button to browse to an image to process. The image should be an actual image that represents the pages that are processed. While a test image can be used, an image that is representative of real data is the best for testing.

When loaded, the panel shows the source image next to the processed image. As changes are made to selections, the image on the right automatically updates with the processed image. The enhancement may take a second or two, so it is recommended to click and leave time for the image to update.

This example performs deskew, line removal, despeckle and boarder removal.  The image is also zoomed in to help see the detail in the image.  The scroll wheel on a mouse zooms in and out of the image.

Image before and after while configuring settings

Looking at the resulting image, the image has been deskewed, lines are removed but not all of the specs are removed.  Small specs are gone but larger ones still appear.  From here, the Despeckle settings can be expanded by clicking the triangle on the left of the check box, and then adjust the settings.  There is interactive help that explains the available settings. They can be adjusted and the results can be viewed on the right. This is a good example of where there can be a trade off to some operations.  The settings can be adjusted to get the larger specs, but at some point, it is likely to start degrading the good text.  If this happens, there are possible paths to take:

• Try additional image enhancements following the operation.
• A different image operation prior to the despeckle operation may improve the despeckling.
• A follow on operation may be able to make additional adjustments.
• Adjust the setting to remove larger specs and use a follow-on setting to rebuild characters.
• Perform an operation twice.
• Leave the larger specks, they may or may not be recognized as garbage characters. Test and see.  The garbage characters may be able to be filtered out later, or may need to be removed in a verify step, or they may be OK to ignore.

This kind of scenario can be used with any enhancement, not just despeckle.  There are often tradeoffs like this and by experimenting with real data, different enhancement operations and different orders, the right balance can be found that gives the best overall improvement for recognition.

When configuration settings appear to be optimal, run a number of real pages in a batch and review the results to determine if further adjustment is needed or if the settings are acceptable.

3.2.1.    A note about keeping original documents
It is a common use case that the original input documents must be preserved. Sometimes, this misleads people into avoiding image enhancement because they must not change the source document.

Using image enhancement does not mean the original images are lost. When image enhancement is performed, the original image is renamed and remains in the batch.  It is the adjusted image that then is used for further processing.  Typically it goes on to have page identification, recognition, verification, anything else that a Datacap application needs to do to process and verify an image.  After the batch is complete, the original images can always be obtained. The application can decide to save only the original images and throw away the transient adjusted images processed in the batch.  The adjusted images can be the ones that are stored, or both the original and adjusted images can be retained.  All of the images are there in the batch and the application developer can decide what to do with the image based on their own image requirements.

Datacap actions provide numerous ways to manage images. Actions can copy images in a batch, build sets of images into a PDF, and upload images and other files to 3rd party repositories.  The original images are never lost but it is common that adjusting images for processing provides better recognition results.  It is up to the application to decide if these adjusted images are to be transient, only for the life of the batch, or not.

3.2.2.    Adding and removing image adjustments
The ruleset provides a number of available enhancements.  Upon opening of the ruleset settings dialog, not all of the available settings may be visible.  The drop-down box noted by “Add Operation” contains a list of all operations that can be performed during image enhancement.  To add one, select the operation in the list and click the Add button.  The operation is then added to the available list of operations.

Similarly, an X button exists on each of the displayed operations.  By selecting the X button, the operation can be removed from the list.  Just unchecking the operation prevents the operation from running, however, some may find it useful to remove unneeded enhancements from the list.  If an operation is removed, it can always be added back again by selecting it from the drop-down list and clicking the Add button.

The order in which the operations are performed is always from the top to the bottom of the list.  Order is very important for the operations to work effectively.  For example, it is best to deskew an image before removing lines.  Line removal may not work as cleanly if the image is skewed.  Some operations also have image color depth requirements.  For example, if the image is a color image and the desired operation only works on black and white images, first perform the binarize operation to convert the image to black and then perform the desired operation.

The order can be adjusted by clicking on the up and down arrows next to the operation name. Using the arrows, the operations can be arranged into specific orders to achieve maximum effectiveness. 

It is allowable to add the same operation twice.  When the Add button is clicked, the operation is added, even if it already exists.  Using this method, the same operation can be performed twice.  While this may not be useful for all operations, there are times when it can help to run an operation, then run other operations, then again rerun the earlier operation and it can run with different detail configuration settings.  This makes it very flexible to arrange the image operations.

4.    Image enhancement features - Highlights
This section highlights a few of the available image enhancement operations.  The ruleset has more operations that are listed here.  It is also possible that more features can be added in future releases, so it is recommended to review the operations available in the drop-down list.

4.1.    Auto rotate
Auto rotate attempts to determine if the is in a 90, 180 or 270 degree rotation and adjust it to 0 degrees.  This is required for recognition and verification so text coordinates are correct and the text is in the correct order. The rotation operation doesn’t understand the language, it just understands the image’s geometry.  It may not always be able to determine the correct orientation, especially if there is a small amount of printing on the page.  It may also rotate some languages better than others.  For example, Hebrew text may be more frequently rotated to 180 degrees instead of 0 degrees.
The recognition actions also have the ability to rotate an image.  When performed by a recognition action, first configure the correct language for the document and then use the automatic rotation provided by the recognition actions. Typically, the recognition engine is more reliable at determining the 0 degree orientation as it understands the type of text that is expected on the page.  Accuracy can be reduced if there is a small amount of text on the page, or if there is text that exists in multiple directions on the page.  Nevertheless, the rotation feature provided by the engine is typically more reliable. This rotation can be performed before using the Image Enhancement ruleset or after, depending on the application’s needs.

4.2.    Close filter
The close filter performs dilate and erode in one step.  This combination provides a thickening or blackening of characters that can improve recognition.  This is often a good choice to try to improve recognition or if words are missed.  It may not work well on highly detailed characters such as Asian character sets.  Holes and broken lines in characters are a prone to recognition errors, this operation can fill in gaps that help reduce those errors.

4.3.    Color depth
Image Enhancement provides a range of image enhancements.  Some enhancements only work on specific color depths.  So if you scan an image at 1 bpp, you cannot perform enhancements such as Adjust brightness/contrast, Close Filter, Image Detergent, Smooth Background. There is a Binarize feature that converts an image down to 1 bpp, but earlier there was no method to increase the color depth. 

Color depth feature is added to increase/decrease the color depth which allows use of the full range of features regardless of the input image color depth. You can specify Bits per pixel, where valid values are 1,4,8 and 24. A 32 bit color image can be converted to a lower bit depth. A lower bit depth cannot be converted to a 32 bit color.

Note: In versions prior to Datacap 9.1.6, Datacap did not accept 32 bit image as input, but with this feature you can convert 32 bit image to lower bit and perform respective operation.

4.4.    Deskew
Deskewing is very important for good recognition and for text to be understood as being on one line.  It is necessary for good field alignment and for viewing and click-n-key in verify panels.  Like automatic rotation, the recognition engines can also deskew an image. Some engines allow selection of what marks on the page are used for the deskew process.  If the deskew in the image enhancement ruleset isn’t working well enough, the deskew provided by the recognition rotation actions can be used before or after image enhancement.  For example, it is possible to use a recognition action to rotate and deskew a page, then run image enhancements, then perform recognition.

4.5.    Dilate
Similar to the close filter, the dilate operation thickens lines and characters.  The close filter may be the best choice to try first to thicken or darken characters. Like the close filter, there are values that can be adjusted to control how aggressive the operation is on the image.

4.6.    Remove lines / Remove combs
Recognition is often improved if lines or comb boxes are removed first. This is true for full page and field-level recognition.  Typically, recognition is best if everything except the text itself is removed from the page. Line removal has adjustable values that determine the limits of the removal.  They indicate how long and thick the line needs to be to consider removal.  Values must be set to allow short lines are removed, but not so short that parts of characters are also removed.
The page should always be deskewed prior to line removal as straight lines are what are removed best.
If lines are thin, filled with holes, or have line breaks, the dilate and close filters can help to repair lines prior to removal so they are removed more fully.

4.7.    Rotate image
Rotate image is for when it is known that an image must always be rotated by a specific amount.  The recognition actions also have this ability, similar to the automatic rotation.  Either the rotation can be performed by image enhancement or the recognition actions, based on preference or need.

4.8.    Smooth objects
Clean sharp text recognizes the best.  Smooth objects can help fill in gaps within characters to create a clean solid character.

4.9.    Smooth background
Text without backgrounds recognizes best.  Light colored or gray backgrounds should not be dithered, leaving speckles of black dots that create a gray appearance.  Instead, backgrounds should be completely removed so they are white, leaving a clear background on which the text is printed.  This enhancement can help clean up backgrounds to improve text clarity.

4.10.    When an operation has no apparent change
When configuring an image operation, it is possible that it produces no visible change.  Here are some common reasons:

• The operation may require a specific color depth.  If there is an exclamation icon “!”, then hover the mouse over it to determine the issue.  The typical reason is that the operation requires a specific color depth and the current image is of a different color depth. For example, if the image is color and the operation requires a black and white image, then enable the Binarize operation before the operation that requires a 1-bit image.
• The configured settings of the operation may need to be adjusted.  Hovering the mouse over the selections displays information about the settings that guides use of the operation.

5.    Legacy image enhancement panel
In the Datacap Studio Zones tab, the “Image” panel allows loading of an image and has similar image operations.  This panel configures settings files for the ImageFix action library.  These actions are added to rules just like all other actions and can be configured to use the settings files created from the Image panel.

This legacy configuration panel has a few drawbacks compared to the Image Enhancement ruleset:

•    The operations take more steps to configure.
•    Operations cannot be ordered; they only run in a pre-defined order.
•    Operations cannot be run twice.

6.    Other image enhancement actions
The Image Enhancement ruleset provides a set of image enhancements that can be configured in a single step.  These are not the only image operations that are available within Datacap.  This section describes some of the other actions that perform image operations.  It highlights some of the available operations.  It is recommended that you review the actions within Datacap Studio to see the entire available list.

These image operations can be configured to run before or after using the Image Enhancement ruleset.

6.1.    ImageConvert action library
This library contains the largest additional set of image actions.  These are some of the available actions.

6.1.1.    Appending images
There are several actions to append images together to make one long image.  Be aware that if too many images are appended together, they can become too long for Datacap or other imaging applications to open and manipulate, so use this with care.

6.1.2.    ConvertToJPEG / ConvertToTIFF
These actions convert the current image to a JPEG or TIFF image.  These actions can also increase or decrease the image color depth at the same time.  If recognition is performed on an image, it is highly recommended to not use JPEG images as the lossy compression softens lines of characters, reducing recognition quality.  If you cannot prevent ingestion of JPEG images, they should be converted to a lossless format, such as TIFF with LZW or FAX compression, after virtual scan and before performing any additional manipulation to them to prevent further reduction of quality.

6.1.3.    RescaleImage / SetImageDPIByWidth
These actions allow images to be resized or to have the dots per inch, dpi, corrected.  Typically recognition requires a minimum resolution of 200 dpi.  Some actions have very specific image size requirements.  Images that have been taken with a camera, often have a large enough pixels for the x and y size of the image, but the dpi may be only 96 because a camera cannot determine the actual physical size of the page being photographed. Using these actions, the dpi and sizes of the images can be corrected or increased.

SetImageDPIByWidth is a good action to correct large enough images that have incorrect DPI settings. For example, if it is known that the page is expected to be 8.5” wide, then this action can be used to adjust the image to a proper size and DPI.  This works on photographs and on scanned pages that have incorrect DPI settings.

It is also possible to scale up a low resolution image to a higher resolution image.  While it will not sharpen an image, it is possible that this coupled with other image enhancement settings, recognition of low resolution images can be performed.

6.1.4.    SaveImageInformation
While this action does not change an image, it can be very useful to obtain information about an image and make decisions at runtime based on the image specifications.  The action saves information such as the image dpi, width, and height in DCO variables.  These variables can then be used as needed by the application.  
As an example, rules can call this action, then check the image size or the image DPI and then run corrective actions on the image that are outside of required specifications.  The Rulerunner actions have numerical comparison actions allowing logic such as “If dpi < 200 then run actions A, B & C”.

6.2.    Equalize action library
It is important for images to have an X and Y dpi that is identical, for example 300x300 dpi.  This is called an isotropic image.  Sometimes, ingested images may have non-matching dpi values, for example 100x200.  This kind of non-isotropic image is frequently seen with images created from a fax.  For fingerprinting, recognition, verification, and other Datacap operations to work correctly, the image must be converted so the x and y dpi are the same.  The action EqualizeUnbalancedImage adjusts an image so the dpi is the same.  For example, it takes a 100x200 image and converts it to a 200x200 image.

If it is possible that an application ingests non-isotropic images, then the first step should always be to call EqualizeUnbalancedImage.  This action should be called after VScan rules but before any other actions.  The action leaves an image unchanged, if the image is already balanced.

6.3.    ColorToBW action library
This action library converts a color image to black and white.  One feature is that this method allows control of the color dithering.  Typically, it is best to set dithering to “None”.  Dithering colors causes black speckles on the document which interferes with recognition.  When “None” is used, lighter colors in the image are simply converted to white, making the text clearer.

6.4.    Recognition actions
In addition to the previously discussed automatic rotation and deskew that is available with some of the recognition action libraries, some have additional image adjustments that can be made.  Reviewing the help for the recognition actions is recommended.  Keep in mind that if you are on an older version of Datacap, additional features may have been added.

7.    Final thoughts
The best recognition results are achieved when images have dark, sharp, solid text with high contrast on a clean background.  Pages should be straight, oriented correctly, preferably with lines removed.  If pages are not recognizing well as desired, image enhancement can improve recognition, even in cases where images already look acceptable.  For example, just running a close filter may help improve recognition by further darkening characters.

Use image enhancement settings that work best overall for the ingested images. If needed, depending on the ingested images, it may be possible to separate pages by different page types and apply different types of operations based on the type.  These types can be changed again after image enhancement, if necessary. The situations would be based on specific use cases and the expected quality of images.

Recognition is not 100% accurate and accuracy declines as image quality deteriorates.  There always are cases where there are recognition mistakes.  The approach is to get the optimal recognition for most pages then use corrective actions and verification to correct mistakes made by recognition engines.