US5315668A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Offline text recognition without intraword character segmentation based on two-dimensional low frequency discrete Fourier transforms
Assignee: United States Department of the Air Force
•
Filing: 1991-11-27
•
Granted: 1994-05-24
•
Expires: 2011-11-27
Abstract
Abstract Image analysis and recognition includes reading text, by digitally scanning a surface, locating the printed material in that digital image, and then recognizing words, phrases, or numbers based on their two dimensional, low frequency Fourier harmonics. One objective is to specifically apply this method of recognition to the postal industry, to include all shipping and labeling applications. Once the image of a word is digitized and isolated, a two-dimensional Fourier transform is computed of the digital image. The process is accomplished in the same manner regardless of the type of surface the printed text comes from, just as long as each word, phrase, or set of numbers to be recognized is isolated, stored in a digital form, and then Fourier Transformed. The sine and cosine coefficients from the Fourier Transform are then filtered to include only the low frequency, terms (i.e. DC term and first 5 harmonics in both vertical and horizontal axis). The sine and cosine terms (coefficients) then define 121 unique vectors which represent a 121 orthogonal vector space. The vector space is normalized to unity and each image of the word, phrase, or set of numbers defines a unique point along this 121 orthogonal vector hypersphere. A library of words, phrases, and/or numbers must be produced using many different font styles. The library when developed, will consist of sine and cosine coefficient values which represent each word, phrase, or number to be recognized. This library is uniquely fashioned by averaging the sine and cosine terms of similar font styles into what is called font groups.
Inventors
- 👤 Mark A. O'Hair
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-10-31 01:40
CPC Classifications
G; G06; G06V; G06V10/00; G06V10/70; G06V10/74; G06V10/75; G06F; G06F18/00; G06F18/20; G06F18/21
Claims Analysis
Total Claims: 7
Claim 1
Independent Claim
345 words
The method of recognizing text, with a system having representative digital images stored in memory, comprising the steps: (a) building a library of library images of known entities comprising words, phrases and numbers using low frequency filtered Fourier coefficients, using many different font styles, and assembling them into font groups; the font groups being created by averaging Fourier coefficients of a particular entity in different font styles, with first and second properties generated for each coefficient, the first property being a mean value which equates to a unique location in Fourier space, the second property being a standard deviation associated with the averaging of the coefficients, this standard deviation being considered as a radius of a chosen font group, a font group center being a mean value derived by averaging Fourier coefficients of a particular entity in the different font styles; (b) inputting an image representing text; (c) digitizing the image to form a digital picture thereof; (d) storing the image in memory; (e) breaking the image into blocks of data, wherein the blocks represent entire words, phrases, or numbers; (f) placing a window around the first individual block; (g) performing a two-dimensional discrete Fourier Transform (2DDFT) of the image within the window; (h) adjusting the punctuations; (i) filtering both the real (cosine) and imaginary (sine) coefficients to the first N harmonics in both the vertical and horizontal axis, these coefficients then make up a total of M unique vectors which defines a M orthogonal vector space, wherein N is at least five and M is (2N+1) 2 ; (j) energy normalizing to unity the M unique vectors, which will eliminate the effects due to image brightness; (k) searching the library of known coefficients and finding the closest match by computing the smallest error between library image and input image; (l) calculating a confidence value of the closest match based on the radius of the chosen font group and the distance between the font group center and the input image; (m) storing the match if confidence is high, or further investigating if confidence is low.
Claim 2
Independent Claim
345 words
The method of recognizing text, with a system having representative digital images stored in memory, comprising the steps: (a) building a library of library images of known entities comprising words, phrases and numbers using low frequency filtered Fourier coefficients, using many different font styles, and assembling them into font groups; the font groups being created by averaging Fourier coefficients of a particular entity in different font styles, with first and second properties generated for each coefficient, the first property being a mean value which equates to a unique location in Fourier space, the second property being a standard deviation associated with the averaging of the coefficients, this standard deviation being considered as a radius of a chosen font group, a font group center being a mean value derived by averaging Fourier coefficients of a particular entity in the different font styles; (b) inputting an image representing text; (c) digitizing the image to form a digital picture thereof; (d) storing the image in memory; (e) breaking the image into blocks of data, wherein the blocks represent entire words, phrases, or numbers; (f) placing a window around the first individual block; (g) performing a two-dimensional discrete Fourier Transform (2DDFT) of the image within the window; (h) adjusting the punctuations; (i) filtering both the real (cosine) and imaginary (sine) coefficients to the first N harmonics in both the vertical and horizontal axis, these coefficients then make up a total of M unique vectors which defines a M orthogonal vector space, wherein N is at least five and M is (2N+1) 2 ; (j) energy normalizing to unity the M unique vectors, which will eliminate the effects due to image brightness; (k) searching the library of known coefficients and finding the closest match by computing the smallest error between library image and input image; (l) calculating a confidence value of the closest match based on the radius of the chosen font group and the distance between the font group center and the input image; (m) storing the match if confidence is high, or further investigating if confidence is low.
Claim 3
Dependent Claim
254 words
The method according to claim 1, wherein the system has the capability to self-monitor its decisions and performance and make an appropriate change(s) to its algorithm to improve its level of recognition; wherein when confidence is low, the method of recognition enters into a self-monitoring mode; wherein this mode includes the following steps: (n) varying the total number of harmonics used and the specific harmonic field used for recognition; (o) monitoring the percentage of words which exceed a low confidence threshold and activating the self-retraining mode; (i) using a small sample of less than a given number of words; (ii) calculating the mean magnitude and direction of errors between the input words, phrases, or numbers and the library words or numbers; (iii) adding this mean magnitude and direction to the entire library font group(s); (iv) if the confidence values do not improve for the next set of said given number of input words, then returning the previous font group(s) to their values and repeating the first given number of inputs against the font styles in the library; using the most popular two or three font styles selected between the nearest neighbor matches, building a new font group using these two or three most popular font styles and the input font style; (v) if the confidence values do not improve for the third set of said given number of input words, then eliminating this new font group, using only the original font groups in the library, and flagging all words with a low confidence value.
Claim 4
Dependent Claim
79 words
The method according to claim 2, which further includes (p) activating solutions that include the steps: (i) breaking the input image into three parts; (ii) imaging the input into three distinct Fourier vector spaces and reaching a best solution based on their correlation; (iii) varying scanner resolution; (iv) searching strategic key subregions of the input image to eliminate confusion between words that differ only in details such as `came/come` or `after/often`; (v) post processing based on syntax and semantics.
Claim 5
Dependent Claim
44 words
The method according to claim 3, which further includes displaying, for a human operator, only those words, phrases, and numbers which exceed the low confidence threshold and can not be improved upon by steps (n), (o), or (p) or any combination of these steps.
Claim 6
Dependent Claim
69 words
The method according to claim 1, wherein the system has the capability to self-monitor its decisions and performance and make an appropriate change(s) to its algorithm to improve its level of recognition; wherein when confidence is low, the method of recognition enters into a self-monitoring mode; wherein this mode includes the following step: (n) varying the total number of harmonics used and the specific harmonic field used for recognition.
Claim 7
Dependent Claim
238 words
The method according to claim 1, wherein the system has the capability to self-monitor its decisions and performance and make an appropriate change(s) to its algorithm to improve its level of recognition; wherein when confidence is low, the method of recognition enters into a self-monitoring mode; wherein this mode includes the following steps: (n) monitoring the percentage of words which exceed a low confidence threshold and activating the self-retraining mode; (i) using a small sample of less than a given number of words; (ii) calculating the means magnitude and direction of errors between the input words, phrases, or numbers and the library words or numbers; (iii) adding this mean magnitude and direction to the entire library font group(s); (iv) if the confidence values do not improve for the next set of said given number of input words, then returning the previous font group(s) to their values and repeating the first given number of inputs against the font styles in the library; using the most popular two or three font styles selected between the nearest neighbor matches, building a new font group using these two or three most popular font styles and the input font style; (v) if the confidence values do not improve for the third set of said given number of input words, then eliminating this new font group, using only the original font groups in the library, and flagging all words with a low confidence value.
References Cited (Backward Citations)
US4327354A
Cited
US4254400A
Cited
US4513442A
Cited
US4651289A
Cited
US4490851A
Cited
US4513441A
Cited
US4817176A
Cited
US4764973A
Cited
US5191622A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 1997-05-23 | 1997-11-22 | 1998-05-24 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2001-05-23 | 2001-11-22 | 2002-05-24 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2005-05-23 | 2005-11-22 | 2006-05-24 | $7700.00 | Window Closed / Lapsed |