System and method of character recognition using fully convolutional neural networks
Embodiments of the present disclosure include a method for extracting symbols from a digitized object. The method includes processing the word block against a dictionary. The method includes comparing the word block against a word in the dictionary, the comparison providing a confidence factor. The method includes outputting a prediction equal to the word when the confidence factor is greater than a predetermined threshold. The method includes evaluating properties of the word block when the confidence factor is less than the predetermined threshold. The method includes predicting a value of the word block based on the properties of the word block. The method further includes determining an error rate for the predicted value of the word block. The method includes outputting a value for the word block, the output equal to a calculated value corresponding to a value of the word block having the lowest error rate.
1. A computer-implemented method, comprising:
obtaining a word block input, the word block input including at least one symbol;
processing the word block against a first dictionary, the first dictionary containing a collection of most common words;
comparing the word block against a word in the first dictionary, the comparison providing a confidence factor that the word block corresponds to the word;
outputting a prediction equal to the word when the confidence factor is greater than a predetermined threshold;
evaluating properties of the word block when the confidence factor is less than the predetermined threshold, the properties being at least one of the length of the word block and the number of symbols in the word block;
resizing the word block based upon the determined length of the word block or the number of symbols in the word block, the resizing corresponding to at least one of a height or a width of the word block, the resizing enabling the word block to be evaluated by a fully convolutional neural network without adjusting parameters of the fully convolutional neural network;
predicting a value of the resized word block based on the properties of the resized word block, wherein predicting a value of the resized word block is performed using a neural network, the neural network including a hierarchy of receptive fields to make a prediction associated with values of the symbols in the resized word block, wherein the neural network is a fully convolutional neural network and the hierarchy of receptive fields include a plurality of portions of the symbols before or after each symbol;
determining an error rate for the predicted value of the resized word block, the error rate corresponding to a calculated difference between the predicated word and one or more words from the first dictionary; and
outputting a value for the resized word block, the output equal to a calculated value corresponding to a value of the resized word block having the lowest error rate when the confidence factor is less than the predetermined threshold.
2. The computer-implemented method of claim 1 , further comprising:
receiving a digital image, the digital image including the at least one symbol;
extracting areas of text from the digital image;
splitting the areas of text into lines of text;
splitting the lines of text into the one or more word blocks, the one or more word blocks including the at least one symbol.
3. The computer-implemented method of claim 1 , wherein resizing the word block comprises resizing the word block to canonical width using at least one of blank prediction or length prediction to thereby resize a width of each symbol forming the word block to be equal.
4. The computer-implemented method of claim 1 , wherein determining the error rate comprises calculating a probabilistic character error rate comprising:
symbol probabilities associated with the symbols forming the resized word block;
word probabilities from the first dictionary; and
contextual probabilities associated with one or more different word blocks associated with the resized word block.
5. The computer-implemented method of claim 1 , wherein determining the error rate comprises calculating a probabilistic character error rate comprising:
assigning word probabilities to the resized word block, a value of the resized word block determined by evaluating the at least one symbol forming the resized word block with a fully convolutional neural network; and
matching the resized word block against a word from the first dictionary, the matching including inserting, deleting, or substituting symbols in the word block to correspond to the word from the first dictionary.
6. The computer-implemented method of claim 5 , wherein the probabilistic error rate penalizes symbol insertion and deletion more than symbol replacement.
7. The computer-implemented method of claim 5 , further comprising classifying the resized word block prior to predicting a value of the resized word block, the classification associated with a property of the resized word block.
8. The computer-implemented method of claim 7 , wherein the property of the resized word block is a class of characterization, a font, a language, a script, or a combination thereof.
9. The computer-implemented method of claim 1 , further comprising evaluating a second dictionary data comprising a list of second dictionary words, the evaluation comparing the prediction equal to the word against a second dictionary word and outputting the second dictionary word when the prediction equal to the word corresponds to the second dictionary word.