Recognition Systems Reduce Cost

Many businesses are actively implementing Optical Character Recognition (OCR) or Intelligent Character Recognition (ICR) or Optical Mark Recognition (OMR) applications. These "recognition systems" convert images to alpha-numeric data, as if the data had been entered directly from a keyboard. They do this with software programs, operating by template matching, features extraction, neural networking or a combination of these approaches. Utilizing the computer to replace clerical staff in capturing information from forms can be a significant cost savings for the right application.
What many who work with these applications don't realize is that choosing the right application to use these computer tools and implementing them correctly can make all the difference in the world. Secondly, everyone should recognize that use of these technologies is NOT about choosing between using the computer and using clerical labor to capture data. The decision is about how to achieve proper blending in the use of both the computer and clerical staff to achieve the desired quality result at the lowest cost.
Establishing Accuracy Requirements
We all know that the cost of repairing errors after the fact can be much more than building the accuracy into the capture process. You should start your implementation of this technology with a clear understanding of the accuracy you require in the resultant data. The higher the accuracy requirement the more you should use secondary sources and clerical staff doing sight reviews to correct errors.
Get Control of Forms Design
Choosing the right application for implementation of recognition systems usually starts with having forms design control. This includes the use of drop out ink, the use of boxes on the form in order to facilitate the proper spacing for words and numbers and, most importantly, the creation of registration marks. Registration usually consists of four "marks" at the four corners of the form. Lack of good registration marks on a form will not only prevent the recognition system from finding the proper place on the form for doing its recognition, but it will also cause a misalignment of the form for image keying. Control of your forms also includes such things as attempting to control the location of folds in the forms prior to processing. Folding can place a line in the image that will affect the resultant accuracy of your recognition system.
If you don't have control, you may not have the right application to use recognition systems. You can still attempt to use it but the accuracy will be reduced depending on the missing controls. A general rule is that the highest accuracy occurs with OMR implementation and the lowest relative accuracy occurs in ICR applications.
Implementing the application correctly starts with your scanning setup. Scanning should be at 200 dpi or better for proper resolution of the image. Resolutions below this will not provide enough clarity for recognition software to do its job properly. Resolutions above this will start to slow your scanning and recognition processing and add to storage and image transmission costs but can improve quality. It is important for scanning software or post-scanning software to do image clean-up before trying to use the recognition systems. This includes image de-speckle. This step ensures that the software trying to recognize the information on the form has the clearest possible image.
Decipher Different Technologies
There are a multitude of vendors selling recognition systems. Choose one that fits your needs. OCR is font recognition and more straightforward than ICR, which is attempting to capture handwritten data. Most software packages also support OMR, which is used for identifying if a box is checked. Make sure your software has an automated way to identify the box itself where forms design is not able to incorporate drop-out ink and can recognize the check in the box, while adjusting for the removal of the box itself. Make sure that your software solution can also perform "percent filled." This will allow you to detect when something is in a field that is not a box. It is particularly useful if you want to know if someone wrote anything in a certain section of the form. Finally, when looking for OCR/ICR software, look for a solution that has multiple engines and does internal comparisons for enhancing the accuracy of the result delivered. You will also want a "confidence" level returned with the result. So, if the software says it saw a "7" but only with a 60% confidence you can determine whether you want to rely on that result. Proper implementation of this kind of software is critically dependent on the effective use of this confidence level in order to prevent inaccuracies with word recognition and importantly, to prevent what is known as "substitution errors" with numbers. This is the software telling you it is a "7" when in fact it is a "1." When your application can not tolerate this kind of error you must implement controls to insure you achieve the accuracy you require.
The best example of a low tolerance for substitution error is the remittance world where the system is attempting to recognize the amount written on a check for payment. This example is the best way to talk about the next point, proper blending of technology and clerical resources to achieve the desired accuracy. In the remittance example you would want to have secondary sources to compare the result from your ICR process. For example, the account would have a Minimum Payment Amount and a Full Payment Amount. Statistically, you would know what percent of the time your customers pay one amount or the other. You can use this amount to compare to the ICR. If it matches, even with a lower confidence level, you are comfortable that you have captured the right number. If it doesn't match either of those two numbers and if the confidence level of the capture is not very high, you should use clerical labor to perform a sight review. While the remittance example is a clear situation where clerical staff is dictated, that mix is generally recommended for more situations than not.
Quality Assurance Checkpoints
Sight review processing can be very automated and can leverage the registration and other features that help the OCR/ICR/OMR process. Since the software knows your form, you can use zooms to facilitate the operator seeing only the field that you want sight verified without having to look around the form. There are several different ways to implement a sight review step. You might want to set your recognition system confidence level high and then look at all results that don't meet the high confidence level.
For OMR applications you can establish the recognition for high on "false positive" and then sight review all selected boxes. For fields that don't have external sources for validating the OCR/ICR/OMR accuracy (like ZIP Code files, account name and address files), you should look at clerical sight review in the same way you would look at a second step if your data capture were totally manual. If you had no automated way to capture information from the form you would probably use a key entry followed by either a second independent keying step if accuracy requirements were extremely high or a sight verify to validate the accuracy of what was keyed. You should look at implementation of recognition systems as a replacement for the first key. Depending on the forms control and secondary data sources for comparison, your OCR/ICR/OMR is your first key and the amount of secondary clerical review is driven by your accuracy need.
Build-In Flexibility
A final note. An important key to any system is on-going feedback and adjustment. Since your system should and will have operators doing sight review of what was captured by the recognition system, you know how often the software is wrong and have the ability to tweak it. The best example here is having a percent filled box for OMR recognition. The sight verify will tell you whether the recognition system is saying the box is checked and it is not (software threshold is set too low and it is picking up noise) or the system is saying the box is not checked and it is (software threshold is set too high). You use this information to adjust the software settings, often at a field level.
While proper guidance for implementation of these kinds of systems should consume volumes, the overview issues remain the same. Your goal is achieving an accuracy level while reducing the cost of capture. Both are achievable if you leverage these new software tools, but only if you never lose sight of your accuracy goal and properly blend the technology implementation with use of clerical staff to make sure that you achieve the accuracy goal. Remember that recognition systems can be profitable, but only if you implement them skillfully.