Image analysis – Image sensing – Optical
Reexamination Certificate
1998-12-11
2002-02-26
Couso, Jose L. (Department: 2724)
Image analysis
Image sensing
Optical
C382S311000, C382S312000
Reexamination Certificate
active
06351574
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to apparatus and methods for facilitating OCR verification processes.
BACKGROUND OF THE INVENTION
Many key-in applications are automated by scanning the hard-copy pages and using Optical Character Recognition (OCR) techniques to recognize the text written in various fields on the pages. No OCR technique is immune to errors, and hence, the automatic OCR phase is typically followed by a verification phase in which the OCRed characters are verified either automatically and/or manually. For most of these applications the scanned pages are forms in which characters may be interrelated within the field or across fields by some logical relationship such as arithmetic, dictionarial and/or logical relationship. This interrelation between characters can be utilized to automatically verify and/or increase the confidence level of part of the characters involved in the relationship. However, basing the verification on solely automatic methods is seldom sufficient. For the vast majority of applications, and in order to achieve a high accuracy level of final recognition, the intervention of a human operator to manually verify characters is normally needed. A useful techniques to implement this manual key-in verification is to display characters to the operator on a computer screen and have some navigation application for the operator to mark and/or correct the erroneous characters.
In many of the key-in verification techniques, fields are extracted from the scanned images and displayed to the operator along with the OCR results. The operator uses a keyboard or mouse to point at the erroneous characters and mark and/or correct them. This technique is called a video coding technique. U.S. Pat. No. 5,455,875 to Chevion et al describes a method, called SmartKey, for organizing the data on the screen and for utilization of a mouse in such a way that yields an improvement factor of 3-6 in productivity relative to ordinary video coding schemes. That is, each operator can key-in 3-6 times more characters than in conventional video coding techniques, or, equivalently, a lesser number of human operators is required to key-in the data.
The disclosures of all publications mentioned in the specification and of the publications cited therein are hereby incorporated by reference.
SUMMARY OF THE INVENTION
The present invention seeks to provide improved apparatus and methods for facilitating OCR verification processes.
To date, two complementary methods are used in order to compose an application: a manual phase, in which new information is obtained from a human operator, generally via a keyboard, and an automatic phase in which new information may be concluded from the logical interrelationships between characters. Both methods are stochastic in nature since they both contain errors. Also, there is no guarantee on the final error rate in the final data. Moreover, these methods cannot be dynamically tuned to approach a desired error rate.
The present invention seeks to provide a new key-in method that interactively combines manual key-in with automatic logical phases such that the manual and automatic phases are data driven. Multiple phases are typically interleaved in such a way that the data acquired in each phase is fed into the next phase typically with the aim of minimizing the number of manual key-strokes. Preferably, a goal of the present system is to optimally use the logical interrelationships between the characters in order to achieve the lowest possible number of key strokes by the human operators, and thus to further increase the overall productivity, measured by number of characters verified per man-hour. The stochastic characteristics of both the automatic logical phase and the manual human operator performance is typically measured, on-line, and the application can be tuned to approach any desired accuracy level, for any type of data, at minimal manual labor time.
The method of the present invention typically achieves an additional productivity factor of 3-10 relative to the aforementioned prior art SmartKey technique. The improvement factor with respect to the prior art SmartKey method depends on the quantity of logical relationships present in the application forms.
The method of the present invention is also termed herein the KIM (Key-In-Management) method. It typically manages OCRed data to/from various key-in stations, where each key-in station (KIS) is a physical or virtual station in which new information is added to the characters sent to these stations, generally via a human operator who supplies feedback regarding data displayed on a computer screen.
A preferred embodiment of the present invention seeks to obtain final OCRed data that best complies with the various logical rules of the form with minimum human labor cost (or maximum efficiency). This goal is typically accomplished by using a set of key-in stations that may supply information on characters at certain accuracy and certain cost. Examples of such key-in stations are human operators keying-in data by video coding techniques using Smartkey carpets, triplets, and/or fields, where Smartkey is a video coding technique described in U.S. Pat. No. 5,455,875 to Chevion. The method of the present invention uses a dispatcher that sends characters to the key-in stations e.g. in order to optimize some predefined cost (of system effectiveness), a logic module that supplies services regarding the inter and intra fields logic rules, and an OCR engine that supplies the probabilities of the OCRed characters. The dispatcher obtains services from the key-in stations and reports to the manager level of the system of the present invention.
Many document processing applications involve extracting ASCII information from images of scanned papers. The basic tool to supply this information is an OCR (Optical Character Recognition) engine that is capable of identifying characters and/or other symbols and marks in a given image. Some OCR engines supply only the most probable guess for each of the characters out of some alphabet of characters applicable for the field containing the character. This type of OCR output is called a hard decision output in which the OCR supplies the final decision for that character. Generally, a confidence level attribute is supplied with the output guess, which is, in most cases, normalized to the range [0,1]. Other OCR engines may supply more information, such as the probability vector of the character image to be one of each of the alphabet characters. OCR engines of this kind are said to supply a soft decision where the calling application may use these probabilities to generate a more global decision for the character images.
The OCR used by the present invention may even include a combination of several types of OCR classifiers via some voting scheme such that the OCR output may be more complex and/or data dependent.
No matter how the OCR of the application is implemented internally it will make errors. In most cases, the error level of the OCR is above the desired error rate required by the application. In order to supply the ASCII data at the error level requested by the application, some of the errors must be corrected. The application may thus use the OCR engine as a tool within a more complex system architecture that eventually supplies the ASCII data at the desired error rate.
In order to decrease the error rate beyond that supplied by the OCR engine, other information regarding the characters is incorporated. A major source of information in many document processing applications is the logical relationships between the application characters. There are many possible types of logical relationships such as: arithmetic (e.g., summation, multiplication, equality, inequality, etc.), syntax (e.g., dates), mathematical formulas (e.g., check-sum digits), and dictionaries (in which words or even phrases are taken from an a priori known dictionary). These logic rules are known a priori, and are defined specifically for each application. In most cases, the
Shevach Ronen
Yair Eyal
Couso Jose L.
Darby & Darby
Do Anh Hong
LandOfFree
Interactive verification of OCRed characters does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Interactive verification of OCRed characters, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Interactive verification of OCRed characters will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2977371