A data processing system is disclosed for selecting the correct form of a garbled input word misread by an optical character reader so as to change the number of characters in the word by character splitting or concatenation. Dictionary words are stored in the system, having characters which are flagged for segmentation or concatenation OCR misread propensity. The OCR word and a dictionary word are loaded into a pair of associated shift registers, aligning their letters on one end. The dictionary word characters are inspected for error propensity flags. When a splitting propensity, for example, is found for a character, special conductional probability values are accessed from a storage and a calculation is performed of the probability that the first character of the dictionary word was split by the OCR into the first and second characters of the OCR word. This regional context probability is compared with the probability of a simple substitution error for the characters. If the probability of segmentation is larger, the OCR characters in the first shift register are shifted one space with respect to the dictionary word characters in the second shift register so that subsequent character pairs to be compared are properly matched. The greater calculated probability is combined in a running product. The dictionary word with the largest running product is output by the system as the most likely correct form of the garbled OCR input word. In addition to optical character recognition, the system disclosed may be applied to correcting segmentation errors in phoneme-characters output from a speech analyzer. In addition to optical character recognition, the system disclosed may be applied to correcting character substitutions, transpositions, additions, and omissions inadvertently typed on a keyboard.
The invention provides a method of deriving generally improved statistical acoustic model of a first class of speech sounds, given a limited amount of sampling data from that first class. This is done by combining a first statistic calculated from samples of that class of speech sounds with a corresponding second statistic calculated from samples of a second, broader, class of speech sounds. Preferably the second statistic is calculated from many more samples than the first statistic, so it has less sampling error that the first statistic, and preferably the second class is a super-set of the first class, so that the second statistic will provide information about the first class. In one embodiment, the invention combines statistics from the models of a plurality of first classes of speech sounds to reduce the sampling error of such statistics and thus improve the accuracy with which such models can be divided into groups of similar models. The first and second statistics can be measurements of spread, of central tendency, or both. They also can relate to different types of parameters, including spectral parameters and parameters representing the duration of speech sounds.
Character string recognition and identification is accomplished with a combined, multi-phase top-down and bottom-up process. Characters in an applied signal are recognized with a process that employs a knowledge source which contains information both, about the basic elements in the signal and about strings of the basic elements in the signal. The knowledge source, which may be derived from a training corpus, includes word probabilities, word di-gram probabilities, statisitics that relate the likelihood of words with particular character prefixes, and rewrite suggestions and their costs. Higher level word n-grams, such as word tri-gram probabilities, can also be used. A mechanism is provided for accepting words that are not found in the knowledge base, as well as for rewrite suggestions that are not in the knowledge base.
There is disclosed an image editing method and apparatus wherein low resolution image data among hierarchically encoded image data are decoded, the decoded low resolution image data are subjected to editing, editing data representative of the editing are stored, the hierarchically encoded image data are decoded to obtain original image data, and the decoded original image data are subjected to the editing in accordance with stored editing data.
In a method for automatic character recognition, character strings marked by word start and word end are formed from the discrete characters calculated with the assistance of a character classifier. These character strings are checked with stored comparison strings of a context lexicon with respect to identity or similarity. The context lexicon is continuously updated by continuous read-in of strings containing no rejection characters, whereby the repeated read-in of identical strings is counted. Current strings are compared to the strings of the context lexicon and that string which is optimum with respect to similarity and frequency is selected for further evaluation. A correction provided with reference to the context comparison is only executed when the substitution transposition is probable based on the classifier characteristic for the characters under consideration.
This disclosure relates to a character recognition method and apparatus through which highly accurate character recognition is capable of being executed inexpensively and at high speed. Character recognition is raised in speed by executing segmentation of character images from an input image and character recognition from the segmented character images in parallel by separate processors without use being made of a special communication processor. After the character images have been segmented from the input image, the results of segmentation are evaluated and the character images are segmented further based upon the results of evaluation. Methods of evaluation include a method involving finding a standard character size from characters written in character images segmented initially, and adopting the difference between the standard size and the size of each segmented character image, and a method involving quantifying results of character recognition of character images segmented initially, and evaluating the results of quantification.
