Trellis based word decoder with reverse pass

Abstract

Systems, apparatuses, and methods to relate images of words to a list of words are provided. A trellis based word decoder analyses a set of OCR characters and probabilities using a forward pass across a forward trellis and a reverse pass across a reverse trellis. Multiple paths may result, however, the most likely path from the trellises has the highest probability with valid links. A valid link is determined from the trellis by some dictionary word traversing the link. The most likely path is compared with a list of words to find the word closest to the most.

Description

BACKGROUND

Field

This disclosure generally relates to techniques for recognizing text from captured images, and specifically for deciding on a word based on a trellis structure and possible characters decisions resulting from optical character recognition.

Background

Image and text recognition is providing important functionality to today's mobile devices. In particular, a user may travel to a country where the national language is unknown to the user, and then may translate signs, menus, or other text included in a camera image into the user's home language (i.e., from a first language to a second language). With some systems, a user may access additional information once the text in the image is recognized.

Some languages are especially difficult to perform word recognition based on the wide range of character combinations, a large dictionary and limited power. What is needed is a means for increasing efficiency of optical word recognition, more quickly determining a valid word, and reducing power consumption.

SUMMARY

Systems, apparatuses, and methods disclosed herein provide for efficient and accurate recognition of text in images.

According to some aspects, disclosed is a method to relate images of words to a list of words in an optical character recognition (OCR) system, the method comprising: receiving a plurality of OCR characters corresponding to an image of a word, wherein the plurality of OCR characters are from an OCR system; determining a most likely (ML) path based on the plurality of OCR characters applied to a loaded forward trellis and a loaded reverse trellis, thereby forming a decoded word; and displaying the decoded word.

According to some aspects, disclosed is a mobile device to relate images of words to a list of words in an optical character recognition (OCR) system, the mobile device comprising: a camera configured to capture an image of a word; a processor coupled to the camera, the processor comprising code to: receive a plurality of OCR characters corresponding to the image of the word; and determine a most likely (ML) path based on the plurality of OCR characters applied to a loaded forward trellis and a loaded reverse trellis, thereby forming a decoded word; and a display coupled to the processor and configured to display the decoded word.

According to some aspects, disclosed is a mobile device to relate images of words to a list of words in an optical character recognition (OCR) system, the mobile device comprising: means for receiving a plurality of OCR characters corresponding to an image of a word, wherein the plurality of OCR characters are from an OCR system; means for determining a most likely (ML) path based on the plurality of OCR characters applied to a loaded forward trellis and a loaded reverse trellis, thereby forming a decoded word; and means for displaying the decoded word.

According to some aspects, disclosed is a device to relate images of words to a list of words in an optical character recognition (OCR) system, the device comprising a processor and a memory wherein the memory includes software instructions to: receive a plurality of OCR characters corresponding to an image of a word, wherein the plurality of OCR characters are from an OCR system; determine a most likely (ML) path based on the plurality of OCR characters applied to a loaded forward trellis and a loaded reverse trellis, thereby forming a decoded word; and display the decoded word.

According to some aspects, disclosed is a non-transitory computer-readable storage medium including program code stored thereon, for a method to relate images of words to a list of words in an optical character recognition (OCR) system, comprising program code to: receive a plurality of OCR characters corresponding to an image of a word, wherein the plurality of OCR characters are from an OCR system; determine a most likely (ML) path based on the plurality of OCR characters applied to a loaded forward trellis and a loaded reverse trellis, thereby forming a decoded word; and display the decoded word.

According to some aspects, disclosed is a method to prepare a forward trellis and a reverse trellis in an optical character recognition (OCR) system, the method comprising: accessing a list of words; loading the forward trellis using the list of words to form a loaded forward trellis; and loading the reverse trellis using the list of words to form a loaded reverse trellis.

According to some aspects, disclosed is a server to prepare a forward trellis and a reverse trellis in an optical character recognition (OCR) system, the server comprising: a list of words; the forward trellis; the reverse trellis; a processor coupled to receive the list of words and coupled to load the forward trellis and the reverse trellis, wherein the processor comprises program code to: access the list of words; load the forward trellis using the list of words to form a loaded forward trellis; and load the reverse trellis using the list of words to form a loaded reverse trellis.

According to some aspects, disclosed is a mobile device to relate images of words to a list of words in an optical character recognition (OCR) system, the mobile device comprising: means for accessing the list of words; means for loading a forward trellis using the list of words to form a loaded forward trellis; and means for loading a reverse trellis using the list of words to form a loaded reverse trellis.

According to some aspects, disclosed is a server to prepare a forward trellis and a reverse trellis in an optical character recognition (OCR) system, the server comprising a processor and a memory, wherein the memory includes software instructions to: access a list of words; load the forward trellis using the list of words to form a loaded forward trellis; and load the reverse trellis using the list of words to form a loaded reverse trellis.

According to some aspects, disclosed is a non-transitory computer-readable storage medium including program code stored thereon for a server to prepare a forward trellis and a reverse trellis, the non-transitory computer-readable storage medium comprising program code to: access a list of words; load the forward trellis using the list of words to form a loaded forward trellis; and load the reverse trellis using the list of words to form a loaded reverse trellis.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows an optical character recognition (OCR) process, in accordance with embodiments of the invention.

FIGS. 2 and 3 show mapping Unicodes to character identifiers (char IDs), in accordance with embodiments of the invention.

FIG. 4 illustrates constructing a trellis structure with an example list of words, in accordance with embodiments of the invention.

FIGS. 5 and 6 illustrate word decoding based on OCR character probabilities, in accordance with embodiments of the invention.

FIG. 7 shows method to select a best link in a path between a previous character and a current character, in accordance with embodiments of the invention r.

FIG. 8 shows a method to determine the most likely path through a trellis, in accordance with embodiments of the invention.

FIGS. 9 and 10 illustrate a calculation of a minimum Hamming distance, in accordance with embodiments of the invention.

FIG. 11 shows a forward pass operating in cooperation with a reverse pass, in accordance with embodiments of the invention.

FIG. 12 shows an example of trellis construction, in accordance with embodiments of the invention.

FIG. 13 shows an example of determining a minimum hamming distance, in accordance with embodiments of the invention.

FIG. 14 shows an example of a reverse trellis, in accordance with embodiments of the invention.

FIGS. 15 and 16 show a mobile device and a server, in accordance with embodiments of the invention.

DETAILED DESCRIPTION

The features and advantages of the disclosed method and apparatus will become more apparent to those skilled in the art after considering the following detailed description in connection with the accompanying drawing.

After processing an image through text detection and Optical Character Recognition (OCR), a word decoder (that leverages a dictionary or list of words) is used to determine a most likely valid word. The importance of an efficient word decoder increases with the increase of dictionary size. Systems and techniques described herein provide a trellis based word decoder that efficiently searches through a list of words to find a valid word. The techniques may also include an additional reverse pass, where the list of words is processed in a reverse order to create the trellis. The forward pass and reverse pass results may then be combined to obtain the final output.

FIG. 1 shows an optical character recognition (OCR) process 100, in accordance with embodiments of the invention. At 102, a mobile device captures an image containing text to form a captured image 104. At 106, a processor in the mobile device extracts, from the captured image 104, an image of a word to form an extracted word 116. The processor may extract the word by de-skewing and forming a rectangular box encompassing pixels that are connected together. Pixels connected together form a connected component. At 118, the processor segments the extracted word 116 into a sequence of character images 120. At 122, the processor recognizes the character images 120 to form a set of characters 124. For example, an extracted word 116 may contain the image (), which the processor breaks down into the set of characters 124, such as {} or {}.

At 126, some embodiments begin with the processor determining a most likely path through a trellis based on the set of characters 124. The processor then forms a word decision 129. The set of characters 124 may be represented by a plurality of possible characters for each position with a corresponding probability (as shown in FIG. 5). Link possibilities 210 are provided as an input signal to the processor performing the trellis decoding. At 130, the word decisions 129 are compiled and displayed.

FIGS. 2 and 3 show mapping Unicodes to character identifiers (char IDs), in accordance with embodiments of the invention. In FIG. 2, a method 200 for offline building of a trellis structure is shown. A list of words 202 (e.g., from a dictionary or other list of words) is provided to 204. At 204, a processor optionally converts and maps Unicode to character identifiers (hereinafter, char IDs) to form a list of words 206. Some character sets are identified with ASCII or Unicode. There is great benefit in terms of size savings to map character sets that are identified with a multiple-byte Unicode (as opposed to character sets identified with ASCII or a single Unicode). More complicated character sets use one, two or three Unicodes in a string for various characters in the set. For example, some Unicodes (e.g., U+093E for ‘’, U+941 for ) are combined to place accent marks to one side of a base character to form a new character (e.g., U+0923 for and U+093E for are combined to form ). At 208, the processor constructs a trellis with the list of words 206 to form link possibilities 210. The list of words 202 (without mapping) and list of words 206 (after mapping) are used interchangeable below. The processor may be in the mobile device or alternatively in a remote server.

In FIG. 3, a table shows five Unicodes mapping Roman letter into char IDs. The table also shows six Unicodes mapping Devanagari characters (also referred to as Nāgarī characters, which are used in the Hindi language) into additional char IDs. The char IDs require less memory than the original Unicodes. Each set of Unicode-to-Char ID mapping identifies one character or letter in the example alphabet. In this case, the lists of words 202 and 206 contain a four word dictionary with words BAD, CAB, DAD and BED. The list of words 202 (using Unicodes U+0041 to U+0045) and list of words 206 (using char IDs 1 to 5) reference the five Roman letters ‘A’ through ‘E’. The list of words 202 also contains Unicodes U+0924, U+092C, U+092D, U+0930, U+093E and U+0941 and list of words 206 contain char IDs 6 to 11. Lists of words 202 and 206 may contain just Roman letter, just Devanagari character or both Roman letters and Devanagari characters. Example below use either Roman letters or Devanagari characters for convenience.

FIG. 4 illustrates constructing a trellis structure with an example list of words, in accordance with embodiments of the invention. During pre-processing, words are loaded into the trellis one at a time. A first stage contains a first letter (first column) of a loaded word, a second stage contains a second letter (second column), a third stage contains a third letter (third column), and so on. For example, the list of words consists of {BAD, CAB, DAD, BED}, which is shown as {B₁A₂D₃, C₁A₂B₃, D₁A₂D₃, B₁E₂D₃} when including column subscripts. A different trellis may be used for words of adifferent length.

The trellis loading process begins with a first word, for example, containing three letters. Either the letters may be individually loaded or links connecting two letters may be loaded. At a first stage, a first letter is read. For the first example word “BAD,” a link from letter B at a first stage (B₁) to letter A as a second stage (A₂) is shown. A second link from A₂ to D₃ completes the path for the first example word “BAD.” Therefore, the word “BAD” may be saved as two links: B₁→A₂ and A₂→D₃. The process continues with example words CAB, DAD and BED with corresponding links C₁ to A₂ and A₂ to B₃ for “CAB,” D₁ to A₂ and A₂ to D₃ for “DAB,” and B₁ to E₂ and E₂ to D₃ for “BED.” In sum, for each node after the first column of nodes (i.e., representing a letter after the first letter position), an input linking vector is shown that identifies valid paths to that node.

Pre-processing may also compute probabilities associated with each node for entering and/or exiting the node. For the example given, node B₁ has an equal 50% probability between two links: B₁→A₂ and B₁→E₂. Similarly, node A₂ has an even probability from three links to enter node A₂ (shown as 33%) and from two links to exit node A₂ (shown as 50%). Node D₃ has a ⅔ probability (67%) of entering from a previous ‘A’ and a ⅓ probability (33%) of entering from a previous ‘E.’

Words may be loaded into a trellis by creating input linking vectors as described above. During runtime, the input linking vectors are examined when checking for a valid link and possible valid paths. The input linking vectors may be created offline, for example, by a server during a pre-processing period. The input linking vector includes one bit for each character in the alphabet. For example for Roman letter, the length of the input linking vector may be 52 for 26 lower case letter and 26 upper case letters. The length may be increased to include 10 digits and various punctuation marks.

For Devanagari script, the length of the input linking vector may vary depending on the complexity of the system deployed. Devanagari vowels may be written independently or be used to modified a basic consonant with diacritical marks, which are written above, below, before or after the consonant they belong to. When a vowel modifies a basic consonant, the vowel is referred to as a “modifier.” The character formed by modifying a basic consonant with a vowel is referred to as a “conjunct.” When concatenating two or more consonant together, the new character is referred to as a “composite character” or a “compound character.” Compound characters may include both basic consonants, modified consonants and half characters. Of the 33 consonants, 24 of them have half forms. Characters in a word are joined by a horizontal bar referred to as a “Shiro-rekha” or a headline. The many forms of compound characters lead to a large number of resulting characters, which make OCR of Devanagari script very difficult.

Considering Devanagari script, the length of the input linking vector may be 44 for the 44 basic characters: 33 consonant and 11 vowels. Alternatively, the length of the input linking vector may be increase to include 10 modifiers. Additionally, the length of the input linking vector may be increase to include 24 half characters. Considering 44 basic characters (33 consonant referred to as “vyanjan” and 11 vowels referred to as “svar”), 10 modifiers and 24 half characters, the length of the input linking vector may be 78. Again, the length may be increased to include 10 digits and various punctuation. The input linking vector may be made shorter by just including common characters and excluding infrequently used characters.

In this example, the input linking vector is length five. A first node in the second column (denoted as A₂ representing a letter ‘A’ in the second position of a word) is shown having an input linking vector [01110]. The input linking vector [01110] may be interpreted from the individual binary values ‘0’ (which means a first link is not valid), ‘1’ ‘1’ ‘1’ (the next three links are valid) and ‘0’ (the last link is also not valid). A link may be considered valid if a word has a transition between a previous node and the current node. Nodes representing the letters B, C, and D have no valid input links so the vector is shown as [00000]. Only one valid input link is shown for E₂ (the final letter E in the second position in a word) from B₁; therefore, the node's vector is [01000], representing the only valid input link to E₂ is from B₁. For the final stage, representing a third letter in a word, input linking vectors are shown as [00000], [10000], [00000], [10001] and [00000], respectively for A₃, B₃, C₃, D₃ and E₃, representing transitions to that node in column 3 by the previous stage from column 2.

Also shown, a set of words includes a letter represented by each node. That is, a word-node set is computed while building the trellis to show what words pass through that particular node. For example, regarding the first stage (a first letter in a word) no word in the list of words 202 starts with A₁, so an empty set { } is shown. Two words start with B₁, which are shown as {BAD, BED}. One word starts with C₁, which is shown as {CAB}. Another word starts with D₁, which is shown as {DAD}. No word starts with E₁, so an empty set { } is shown. Regarding a second stage, three words contain A₂; therefore, the set is shown as {BAD, CAB, DAD}. No words contain B₂, C₂ or D₂ in the list of words 202; therefore, empty sets { } are shown. A single word includes E₂, so the set shows {BED}. For the final stage, letters A₃, B₃, C₃, D₃ and E₃ include sets { },{CAB}, { }, {BAD, DAD, BED] and { }, respectively.

The process of building the trellis by creating the input linking vectors and word-node sets may be performed during a pre-processing period by the mobile device or alternatively by a remote server and downloaded to the mobile device.

FIGS. 5 and 6 illustrate word decoding based on OCR character probabilities, in accordance with embodiments of the invention. Each letter position includes a primary letter and associated probability and a secondary letter and its associated probability. Even though only primary and secondary letters and probabilities are shown, OCR character probabilities may expand to tertiary and quaternary or more possibilities.

FIG. 5 shows an output from an OCR unit (e.g., performed at 122), where a letter and its probability of being that letter are provided for each letter position. In this example, a character image 120 in a first position has a probability of 60% or 0.6 of being a letter B (shown as B₁). The first character image 120 in the first position also has a probability of 40% or 0.4 of being a letter D (shown as D₁). That is, the primary letter is a B with probability 0.6 and a secondary letter is a D with probability of 0.4. For a second position, a second character image 120 has a probability of 40% or 0.4 of being a letter A₂ and a probability of 60% or 0.6 of being a letter D₂. A third letter is shown having a probability of 30% or 0.3 of being a letter B₃ and a probability of 70% or 0.7 of being a letter D₃. The process works for an OCR unit that provides two letter possibilities (as shown here) or three, four or more letter possibilities. Two possible letters are shown to exemplify the process described herein.

FIG. 6 shows resulting links based on the link possibilities 210 (for pre-processing computed offline) for the example OCR unit output (set of characters 124) shown in FIG. 5. Possible links for the set of characters 124 are shown with dotted lines. Links with ‘X’ through the dotted line are not in the trellis or are removed from the trellis during pre-processing to indicate no word exists with that link. The set of characters 124 with corresponding probabilities is used to traverse the trellis.

In the example shown, B₁ is assigned a value of 0.6 and D₁ is assigned a value of 0.4 to correspond to the table in FIG. 5. The table shows a first character to second character transition may be B₁→A₂, B₁→D₂, D₁→A₂, and D₁→D₂, however, links D₁→A₂, and D₁→D₂ do not exist in the trellis. In this case, D₂ is a possible OCD decode but is an impossible letter for the second position based on the trellis. Therefore, these two links are removed from consideration. Next, A₂ is assigned 0.4 and D₂ is assigned 0.6. The table shows a second character to third character transition may be A₂→B₃, A₂→D₃, D₂→B₃, and D₂→D₃, however, links D₂→B₃, and D₂→D₃ do not exist in the trellis. Therefore, these links may be eliminated. The only valid links are B₁→A₂, D₁→A₂, A₂→B₃, and A₂→D₃ resulting in four valid paths (B₁→A₂→B₃, B₁→A₂→D₃, D₁→A₂→B₃, D₁→A₂→D₃) with four respective cumulative probabilities (0.6*0.4*0.3=7.2%, 0.6*0.4*0.7=16.8%, 0.4*0.4*0.3=4.8% and 0.4*0.4*0.7=11.2%). The best path is shown connecting valid links having the highest probabilities from FIG. 5, namely, the second path B₁→A₂→D₃ with cumulative probability of 16.8%.

FIG. 7 shows method 300 to select a best link in a path between a previous character and a current character, in accordance with embodiments of the invention. In this forward pass, a first pair of previous and current characters is the first character paired with the second character. A second pair is the second character pair with the third character and so one. Method 300 selects a best path (shown as the solid line in FIG. 6). Method 300 may be performed on the mobile device, on a server or a combination of both.

At 302, a processor considers a first pair of characters. The first pair of characters contains a first character and a second character represented as a previous character and a current character, respectively. That is, the current character is initialized from the second character and the previous character is initialized from the first character.

At 304, the processor finds the highest probability for the current character and initializes an end of a possibly selected link. In the example above, the current character is be A₂ with probability 0.4 and D₂ with probability 0.6. The previous character is be B₁ with probability 0.6 and D₁ with probability 0.4. Thus, the highest probability for the current character is D₂ with probability 0.6.

At 306, a beginning of an outer loop starts. The processor selects the highest probability for the previous character. In the example above, the highest probability for the previous character is B₁ with probability 0.6.

At 308, a beginning of an inner loop starts. The processor makes a determination whether the selected current character and the selected previous character form a valid link in the trellis. For example, the processor examines the appropriate input linking vector. In the example case, the input linking vector is [01010], which indicates that B₁→A₂ and D₁→A₂ are valid links. If the current and previous characters form a valid link from the trellis, processing continues at 310. If not, processing continues at 312.

At 310, if the link exists in the trellis, a link between the current and previous characters is selected as the best link. The process continues at 302 with new current and previous characters to find the next best link between the next stages.

At 312, if no link exists in the trellis, the process advances to examine the next highest probability as the previous character at 314 and then returns to 308. Otherwise, at 318, a check is made as whether next character exists and may be set as the current character. If a next character exists, the next highest probability for the current character is selected at 316, then the process continues at 306. At 320, if no possible links exist, then the process selects the first link (having the highest probability) even though the link is not in the trellis.

FIG. 8 shows a method 400 to determine the most likely path through a trellis, in accordance with embodiments of the invention. At 402, a processor accepts a set of characters 124. The best set of links is selected for each pair of previous and current characters. The best set of links is set as the one or more links from a previous OCR recognized character to each current OCR recognized character. At 404, a check is made to see of the ML path 128 from the best links forms a valid word from the list of words 206. If so, at 406, that word is selected as the word decision 129. If not, at 408, the word from the list of words 206 with the minimum Hamming distance is selected as the word decision 129.

FIGS. 9 and 10 illustrate a calculation of a minimum Hamming distance, in accordance with embodiments of the invention. To calculate the minimum Hamming distance, the word-node sets may be used. A word-node set is a set of all words that goes through a particular node. For example, a list of words 202 contains four words in its dictionary {BAD, CAB, DAD, BED}. A word-node set for the first position being a letter ‘B’ (or B₁) is {BAD, BED}.

FIG. 9 shows a most likely (ML) path 128. A counter exists for each possible word along the best path. In the example shown, a ML path 128 includes nodes B₁→A₂→D₃. For a first stage, a processor increments all words containing B₁ in the list of words 202. The ML path 128 from B₁ has words {BAD, BED} in the word-node set. Therefore, counters for BAD and BED are each set to one as shown in FIG. 10. Counters for other words in the list of words 202 are left at zero. The count is now BAD=1, CAB=0, DAD=0 and BED=0.

For the second stage, the word-node set {BAD, CAB, DAD} is shown along the best path. The counters for these words are incremented. The count is now BAD=2, CAB=1, DAD=1 and BED=1. Finally, the last stage passes through D₃. The word-node set includes {BAD, DAD, BED} for this node. Counters for these words are similarly incremented. The count is now BAD=3, CAB=1, DAD=2 and BED=2. The word having the highest count (e.g., BAD=3) is the word having the minimum Hamming distance to the best path. Therefore, the processor selects the word having the highest count as the selected word for this forward pass.

FIG. 11 shows a forward pass operating 500 in cooperation with a reverse pass, in accordance with embodiments of the invention. At 502, a processor performs a forward pass operation, as described above. During a forward pass, pre-processing creates a standard dictionary or list of words sorted from first character position to last character position. During operations, a processor determines a most likely forward path through the trellis. The processor then determines an ML path 128-1 from the forward trellis. At 506, the processor computes a probability of the selected path.

Similar to a forward pass, pre-processing may also create a reverse-order dictionary where a list of words 202 is sorted from last character position to first character position. At 504, the processor performs a reverse pass with the reverse-ordered list of words 202 to form an ML path 128-2 from the reverse trellis. At 508, the processor computes a probability of the selected path. At 510, the processor compares the probabilities from 508 and 508 and then selects the greater probability of the two paths as the ML path 128. In some test results with noisy images, a 5% improvement was found by adding a reverse pass to a forward pass system.

In the examples above, a simple form of the English language using an alphabet of five characters was used. The method may be expanded to a 26-character alphabet or to an alphabet that includes both upper and lower case letters, numbers and punctuation. In the examples below, Devanagari is used to illustrate the method for more complicated characters sets. Most North Indic scripts (e.g., the Devanagari script, also called Nāgarī, which is used in India and Nepal among other countries) are written from left to right, do not have distinct character cases, and are recognizable by a horizontal bar or line that runs along the top of characters. Devanagari script is commonly used to write standard Hindi, Marathi, Nepali and Sanskrit. The Devanagari script may be used for many other languages as well, including Bhojpuri, Gujari, Pahari, Garhwali, Kumaoni, Konkani, Magahi, Maithili, Marwari, Bhili, Newari, Santhali, Tharu, Sindhi, Dogri and Sherpa.

In Devanagari in general, modifiers (e.g., upper and lower modifiers) add a great deal of complexity to the basic character set due to the large variety. In fact, over a 1000 character combinations and contractions are possible. Currently, OCR systems have difficulty identifying works with such a complex set of character variations. An OCR system may be simplified to 100 characters, which is most commonly used, in the Devanagari alphabet.

In some embodiments, each word is considered as a sequence of characters with a unique identifier. As an example (India) is represented as +++. A dictionary of words is a list of valid words, where each word is represented as a sequence of characters (as shown in the example above). An OCR unit outputs one, two, three, four or more possibilities for each OCR character with their corresponding likelihood. An OCR character is a character recognized by an OCR system. The table of FIG. 5 shows two possible OCR characters for each set of characters 124. For some embodiments, any remaining probability can be equally divided among all the other characters in the list of words. Using the current techniques, a processor finds the most likely (valid) word from the list of words, given outputs from character decoder. A trellis based algorithm is used to compute the most likely word.

In some implementations, forward and reverse trellises may be created offline (away from a mobile device) during a pre-processing period. A list of words may be provided. The list of words can be processed from left to right to generate a forward trellis, and processed from right to left to generate a reverse trellis. The information indicative of the forward and reverse trellises can then be stored as a trellis database for later access.

In order to identify text included in image data, a candidate sequence of decoded characters 124 is identified (e.g., at the mobile device or a network resource such as a server). The candidate sequence is processed using a trellis based decoder, which accesses the trellis database. Additionally, a reverse candidate sequence is generated that includes the decoded characters 124 in reverse order, which is also processed using the trellis based decoder. Based on processing the candidate sequence and the reverse candidate sequence, the candidate sequence may be matched to a dictionary word having the highest confidence or highest probability. The “highest confidence” may be indicated by a confidence score, which may be the fraction of match between the winning or ML path 128 and the decoded word. In some implementations, if a match cannot be determined with a minimum confidence, the algorithm may provide an indication that the search has failed or is not reliable.

FIG. 12 shows an example of trellis construction, in accordance with embodiments of the invention. In the following, N is the total number of character identifiers, L is the maximum number of characters in a word (for a given list of words), and W is the number of words in the list of words. In the example shown, N=7 (with characters ‘’), L=5 (length of second and last word), and W=3 (with words , and ).

For each node of the trellis, the following two quantities may be maintained for pre-processing. First, a (binary) characteristic vector or input linking vector in N bits long and represents the list of characters from previous level connected to the node. For the node highlighted with a dotted circle, the input linking vector is [0101000]. Second, a list of words that visit the node in the trellis may be maintained as a word-node set. The node highlighted node contains two words {, } passing through the node.

During runtime, the trellis structure may also contain a counter for each word at each character position. An OCR unit provides a set of decoded characters 124 and associated probabilities for matching within the trellis. The OCR unit may provide a single possible character for each position. Alternatively, the OCR unit provides multiple possible characters for each position along with an associated probability for each character. Passing a set of decoded characters 124 from the OCR unit through the trellis results in a best path. The word from the list of words that is closest to the ML path 128 is selected as the word decision.

Assume for this example an OCR unit provides three characters at each position. For the first character position, the OCR unit provides three possible characters and associated probabilities. For the second character position, the OCR unit provides another three possible characters with associated probabilities. Looking back from the second position to the first position, the link having the highest probability is saved for each of the possible second characters. That is, up to three links are saved. For each node in the second character position identified by the OCR unit, a link may be formed from each of the three possible characters from the first character position to the second character position. That is, the number of possibilities at each position is squared to determine how many possible links exist (e.g., 3²=9).

Instead of saving every link to between the positions, only the link with the highest probability is saved. For example, the first character position has possibilities of 0.7, 0.2 and 0.1 for the first, second and third positions of the first character positions, respectfully. The second character position has possibilities of 0.5, 0.3 and 0.2 for the first, second and third positions of the second character positions, respectfully. The different permutations from first to second character positions results in nine links: three links to the first node of the second position with link probabilities of (0.7*0.5=0.35, 0.2*0.5=0.10 and 0.1*0.5=0.05); three links to the second node of the second position with link probabilities of (0.7*0.3=0.21, 0.2*0.3=0.06 and 0.1*0.3=0.03); and three links to the third node of the second position with link probabilities of (0.7*0.2=0.14, 0.2*0.2=0.04 and 0.1*0.2=0.02).

Of the three links from nodes of the first position to the first node of the second position, only one link is saved. The saved link has the highest priority that is also a valid link in the trellis. If the highest priority link is an invalid link, it is discarded. Therefore, the link with the highest probability to the first node in the second position and found as a valid link from the input linking vector is saved. Similarly, the highest probability valid link to the second node in the second position determined from the input linking vector is saved. Finally, the link with the highest valid probability to the third node in the second position is saved.

If the highest probable link is determined to be an invalid link, the next highest probability is considered as long as it is a valid link. If also an invalid link, the next highest probability is considers and so on. If no valid links are found, the link with the highest probability is used.

As we progress from the first and second positions to the second and third positions, the same strategy of saving one link for each possible node (the highest valid link) is saved. We then progress to the next pair of positions and so on keeping track of the highest valid links and therefore, keeping track of a ML path 128 of entering into a node at the current level. At the final level, three paths are formed; one path into each OCR node. At this final stage, the node having the highest path likelihood is selected and referred to as the winning path, best path or most likely (ML) path. That is, the forward trellis process results in one winning path. Similarly, the reverse trellis also results in one winning path. The processor compares the likelihood of the forward path to the likelihood of the reverse path and selected the overall winning path as the ML path 128 having a higher likelihood between the two paths. Often the forward path and the reverse path are identical. In cases when they are different though, comparing the likelihoods and selecting the path with a greater likelihood as the ML path 128 improves performance over a system that only examines a single path.

There may be some issues with the ML path. For example, the final overall winning path may correspond to an invalid word (i.e., a word not in the list of words. As shown in FIG. 13, assume the case where the list of words contains three words: , and (or the characters <>, <>, and <>, respectively). The processor finds the ML path 128 corresponds to (or the characters <>). In order to mitigate this issue, the Hamming distance may be computed for the ML path 128 with words that touch this path in the list of words. In this example, the Hamming distance for the first dictionary word is 1 (1 character differs and 3 of 4 characters correspond between the first word in the dictionary and the word formed along the ML path). The Hamming distance for the second and third dictionary words are 4 and 2, respectively. Therefore, the word along the ML path 128 () is closest to the dictionary word with a minimum Hamming distance of 1. The word with the minimum Hamming distance may then be an output word.

Another issue is that the selected word could be longer than the input word. This happens when the initial segment of a longer word “matches” the path; e.g., if the input word is , the final word could be . This issue may be mitigated by imposing a length constraint on the output word.

As noted above, each word in the list of words is included in the reverse order, and another trellis is created (referred to as a reverse trellis). Upon receiving OCR output, we reverse the input string and search through the reverse trellis to obtain the most likely path. This step is referred to herein as the reverse pass.

If the reverse pass returns a word with higher confidence than the forward pass, then its output can be retained. If not, the output of the forward pass is retained. Note, in some embodiments the input string need not be completely reversed to create the reverse trellis. We can instead start somewhere in between, and traverse both forward and reverse to create the trellis.

The following process may be performed to incorporate a reverse pass in the word decoder. Each word in the list of words can be read in the reverse order and another trellis constructed. Upon receiving the OCR output, the input string can be reversed and the reverse trellis searched to obtain the ML path 128-2 (called “reverse pass”). As noted above, if the reverse pass returns a word with higher confidence than the forward pass, its output is retained. If not, the output of the forward pass can be retained.

As shown in FIG. 14, a reverse pass provides diversity (as the noise within character images may vary across the length), and hence may provide gain. The first word (mellon) is reversed in order and saved in the trellis as <>. Similarly, the second and third words are saved as <> and <>, respectively. The plurality of OCR characters are also reversed before the processor passes the characters through the reverse trellis.

FIGS. 15 and 16 show a mobile device 600 and a server 700, in accordance with embodiments of the invention. In FIG. 15, the mobile device 600 includes an optional wireless transceiver 610, such as a cellular receiver and/or a WiFi transceiver. The mobile device 600 may also include a wired transceiver. The mobile device 600 may also contain a GNSS receiver, such as a GPS receiver. The mobile device 600 also includes an optional sensor 620, such as an accelerometer, a gyroscope and/or a barometer. The mobile device 600 also includes a processor 630, memory 640, a camera 650 and a display 660. The processor 630 and memory 640 contain software instructions or code to perform the methods performed on a mobile device described herein. The camera 650 captures images for the processor 630 to perform OCR. The display 660 shows the results from the processor 630. A bus 670 couples these components together. Some components may connect directly together without using the bus 670.

In FIG. 16, the server 700 includes a transceiver 710, a processor 730 and memory 740 communicating over bus 770. The transceiver 710 communicates to the mobile device 600 via the Internet and/or with a wired or wireless transceiver. The processor 730 and memory 740 contain software instructions or code to perform steps and methods performed on a server describe herein. A bus 770 couples these components together. Components may connect directly together without using the bus 770.

Various wireless communication networks based on infrared, radio, and/or microwave technology can be used to implement described techniques. Such networks can include, for example, a wireless wide area network (WWAN), a wireless local area network (WLAN), a wireless personal area network (WPAN), and so on. A WWAN may be a Code Division Multiple Access (CDMA) network, a Time Division Multiple Access (TDMA) network, a Frequency Division Multiple Access (FDMA) network, an Orthogonal Frequency Division Multiple Access (OFDMA) network, a Single-Carrier Frequency Division Multiple Access (SC-FDMA) network, and so on. A CDMA network may implement one or more radio access technologies (RATs) such as cdma2000, Wideband-CDMA (W-CDMA), and so on. Cdma2000 includes IS-95, IS-2000, and IS-856 standards. A TDMA network may implement Global System for Mobile Communications (GSM), Digital Advanced Mobile Phone System (D-AMPS), or some other RAT. GSM and W-CDMA are described in documents from a consortium named “3rd Generation Partnership Project” (3GPP). Cdma2000 is described in documents from a consortium named “3rd Generation Partnership Project 2” (3GPP2). 3GPP and 3GPP2 documents are publicly available. A WLAN may be an IEEE 802.11x network, and a WPAN may be a Bluetooth network, an IEEE 802.15x, or some other type of network. The techniques may also be used for any combination of WWAN, WLAN and/or WPAN.

Those skilled in the art will understand that information and signals may be represented using any of a variety of different technologies and techniques. For example: data, information, signals, bits, symbols, chips, instructions, and commands may be referenced throughout the above description. These may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.

In one or more exemplary embodiments, the functions and processes described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Disk and disc, as used herein, includes compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and blu-ray disc where disks usually reproduce data magnetically, while discs reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media. The term “control logic” used herein applies to software (in which functionality is implemented by instructions stored on a machine-readable medium to be executed using a processor), hardware (in which functionality is implemented using circuitry (such as logic gates), where the circuitry is configured to provide particular output for particular input, and firmware (in which functionality is implemented using re-programmable circuitry), and also applies to combinations of one or more of software, hardware, and firmware.

For a firmware and/or software implementation, the methodologies may be implemented with modules (e.g., procedures, functions, and so on) that perform the functions described herein. Any machine readable medium tangibly embodying instructions may be used in implementing the methodologies described herein. For example, software codes may be stored in a memory, for example the memory of mobile station, and executed by a processor, for example the microprocessor of modem. Memory may be implemented within the processor or external to the processor. As used herein the term “memory” refers to any type of long term, short term, volatile, nonvolatile, or other memory and is not to be limited to any particular type of memory or number of memories, or type of media upon which memory is stored.

Moreover, the previous description of the disclosed implementations is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these implementations will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other implementations without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the features shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method to prepare and use a forward trellis and a reverse trellis in an optical character recognition (OCR) system, the method comprising: loading the forward trellis containing characters from a list of words to form a loaded forward trellis;loading the reverse trellis containing the characters from the list of words to form a loaded reverse trellis;computing a forward path probability of a forward path through the loaded forward trellis;determining a most-likely forward path based on the forward path probability;computing a reverse path probability of a reverse path through the loaded reverse trellis;determining a most-likely reverse path based on the reverse path probability; andselecting a path based on a greater of the forward path probability and the reverse path probability, wherein the path represents a set of selected characters from the list of words in the OCR system.

2. The method of claim 1, wherein the list of words comprise a Devanagari list of words.

3. The method of claim 1, further comprising forming a linking vector that indicates valid and invalid links into a node.

4. The method of claim 1, further comprising, for each node in the forward trellis and the reverse trellis having a previous character position entering the node, forming a linking vector that indicates valid and invalid links into a node.

5. The method of claim 1, further comprising, for each node in the forward trellis and the reverse trellis having a previous character position entering the node, forming a set of words indicating which words from the list of words have a path through the node.

6. The method of claim 1, wherein loading the reverse trellis comprises, for each word in the list of words: selecting a word from the list of words; andloading the reverse trellis at: a first character position with a final character;a second character position with a second to the final character; anda third character position with a third to the final character.

7. The method of claim 1, further comprising sending the loaded forward trellis and the loaded reverse trellis to a mobile device.

8. An apparatus to prepare and use a forward trellis and a reverse trellis in an optical character recognition (OCR) system, the apparatus comprising: the forward trellis;the reverse trellis; anda processor coupled to receive a list of words and coupled to load the forward trellis and the reverse trellis, wherein the processor comprises program code to: load the forward trellis containing characters from the list of words to form a loaded forward trellis;load the reverse trellis containing characters from the list of words to form a loaded reverse trellis;compute a forward path probability of a forward path through the loaded forward trellis;determine a most-likely forward path based on the forward path probability;compute a reverse path probability of a reverse path through the loaded reverse trellis;determine a most-likely reverse path based on the reverse path probability; andselect a path based on a greater of the forward path probability and the reverse path probability, wherein the path represents a set of selected characters from the list of words in the OCR system.

9. The apparatus of claim 8, wherein the list of words comprise a Devanagari list of words.

10. The apparatus of claim 8, further comprising a linking vector that indicates valid and invalid links into a node.

11. The apparatus of claim 8, further comprising, for each node in the forward trellis and the reverse trellis having a previous character position entering the node, a linking vector that indicates valid and invalid links into a node.

12. The apparatus of claim 8, further comprising, for each node in the forward trellis and the reverse trellis having a previous character position entering the node, a set of words indicating which words from the list of words have a path through the node.

13. An apparatus to prepare and use a forward trellis and a reverse trellis in an optical character recognition (OCR) system, the apparatus comprising a processor and a memory, wherein the memory includes software instructions to: load the forward trellis containing characters from a list of words to form a loaded forward trellis;load the reverse trellis containing the characters from the list of words to form a loaded reverse trellis;compute a forward path probability of a forward path through the loaded forward trellis;determine a most-likely forward path based on the forward path probability;compute a reverse path probability of a reverse path through the loaded reverse trellis;determine a most-likely reverse path based on the reverse path probability; andselect a path based on a greater of the forward path probability and the reverse path probability, wherein the path represents a set of selected characters from the list of words in the OCR system.

14. A non-transitory computer-readable storage medium including program code stored thereon for an apparatus to prepare and use a forward trellis and a reverse trellis, the non-transitory computer-readable storage medium comprising program code to: load the forward trellis containing characters from a list of words to form a loaded forward trellis;load the reverse trellis containing the characters from the list of words to form a loaded reverse trellis;compute a forward path probability of a forward path through the loaded forward trellis;determine a most-likely forward path based on the forward path probability;compute a reverse path probability of a reverse path through the loaded reverse trellis;determine a most-likely reverse path based on the reverse path probability; andselect a path based on a greater of the forward path probability and the reverse path probability, wherein the path represents a set of selected characters from the list of words in the OCR system.

15. The non-transitory computer-readable storage medium of claim 14, wherein the list of words comprise a Devanagari list of words.

16. The non-transitory computer-readable storage medium of claim 14, further comprising forming a linking vector that indicates valid and invalid links into a node.

17. The non-transitory computer-readable storage medium of claim 14, further comprising, for each node in the forward trellis and the reverse trellis having a previous character position entering the node, forming a linking vector that indicates valid and invalid links into a node.

18. The non-transitory computer-readable storage medium of claim 14, further comprising, for each node in the forward trellis and the reverse trellis having a previous character position entering the node, forming a set of words indicating which words from the list of words have a path through the node.

19. The non-transitory computer-readable storage medium of claim 14, wherein loading the reverse trellis comprises, for each word in the list of words: selecting a word from the list of words; andloading the reverse trellis at: a first character position with a final character;a second character position with a second to the final character; anda third character position with a third to the final character.

20. The non-transitory computer-readable storage medium of claim 14, further comprising sending the loaded forward trellis and the loaded reverse trellis to a mobile device.