Optical character recognition system with back propagation of an objective loss function
A document management system performs optical character recognition (OCR) on a high resolution version of an image of a document, obtaining a first set of text representative of the text of the document. The document management system applies a machine-learned model on a low-resolution version of the image of the document, producing a denoised image that is of a higher resolution than that input into the machine-learned model. The document management system performs OCR on the denoised image, obtaining a second set of text representative of the text of the document. The document management system compares the first and second sets of text in the form of an objective loss function. The document management system subsequently generates a feature vector from the comparison of the sets of text and retrains the machine-learned model with the generated feature vector.
1. A method comprising:
accessing an image of a document comprising text;
processing the accessed image of the document to produce a first image version at a first resolution and to produce a second image version at a second resolution lower than the first resolution;
performing optical character recognition on the first image version to obtain a first set of text representative of the text of the document;
applying a machine-learned model to the second image version, the machine-learned model configured to produce a denoised image of the document at a third resolution, the third resolution higher than the second resolution;
performing optical character recognition on the denoised image of the document to obtain a second set of text representative of the text of the document;
generating a feature vector comprising entries each representative of a performance of the machine-learned model, at least one feature vector entry based on a comparison of the first set of text and the second set of text; and
retraining the machine-learned model using the generated feature vector.
2. The method of claim 1 , wherein the comparison of the first set of text and the second set of text comprises a comparison of letters, words, or sentences in each of the sets of text.
3. The method of claim 1 , wherein the retrained machine-learned model is configured to identify text in the first set of text that is missing in the second set of text.
4. The method of claim 1 , wherein the generated feature vector comprises at least one feature vector entry based on a mean-squared error loss function, a divergence loss function, a cross-entropy loss function, or a VGG loss function.
5. The method of claim 4 , wherein the mean-squared error loss function is based on a comparison between pixels of the first image version of the document and the denoised image of the document.
6. The method of claim 1 , wherein the machine-learned model is a convolutional neural network.
7. The method of claim 1 , wherein the machine-learned model is periodically retrained using subsequently accessed images of documents.
8. Non-Transitory computer-readable storage media storing executable instructions that, when executed, cause one or more processors to:
access an image of a document comprising text;
process the accessed image of the document to produce a first image version at a first resolution and to produce a second image version at a second resolution lower than the first resolution;
perform optical character recognition on the first image version to obtain a first set of text representative of the text of the document;
apply a machine-learned model to the second image version, the machine-learned model configured to produce a denoised image of the document at a third resolution, the third resolution higher than the second resolution;
perform optical character recognition on the denoised image of the document to obtain a second set of text representative of the text of the document;
generate a feature vector comprising entries each representative of a performance of the machine-learned model, at least one feature vector entry based on a comparison of the first set of text and the second set of text; and
retrain the machine-learned model using the generated feature vector.
9. The non-transitory computer-readable storage media of claim 8 , wherein the comparison of the first set of text and the second set of text comprises a comparison of letters, words, or sentences in each of the sets of text.
10. The non-transitory computer-readable storage media of claim 8 , wherein the retrained machine-learned model is configured to identify text in the first set of text that is missing in the second set of text.
11. The non-transitory computer-readable storage media of claim 8 , wherein the generated feature vector comprises at least one feature vector entry based on a mean-squared error loss function, a divergence loss function, a cross-entropy loss function, or a VGG loss function.
12. The non-transitory computer-readable storage media of claim 11 , wherein the mean-squared error loss function is based on a comparison between pixels of the first image version of the document and the denoised image of the document.
13. The non-transitory computer-readable storage media of claim 8 , wherein the machine-learned model is a convolutional neural network.
14. The non-transitory computer-readable storage media of claim 8 , wherein the machine-learned model is periodically retrained using subsequently accessed images of documents.
15. A document management system comprising:
one or more processors; and
non-transitory computer-readable storage media storing executable instructions that, when executed, cause the one or more processors to:
access an image of a document comprising text;
process the accessed image of the document to produce a first image version at a first resolution and to produce a second image version at a second resolution lower than the first resolution;
perform optical character recognition on the first image version to obtain a first set of text representative of the text of the document;
apply a machine-learned model to the second image version, the machine-learned model configured to produce a denoised image of the document at a third resolution, the third resolution higher than the second resolution;
perform optical character recognition on the denoised image of the document to obtain a second set of text representative of the text of the document;
generate a feature vector comprising entries each representative of a performance of the machine-learned model, at least one feature vector entry based on a comparison of the first set of text and the second set of text; and
retrain the machine-learned model using the generated feature vector.
16. The document management system of claim 15 , wherein the comparison of the first set of text and the second set of text comprises a comparison of letters, words, or sentences in each of the sets of text.
17. The document management system of claim 15 , wherein the retrained machine-learned model is configured to identify text in the first set of text that is missing in the second set of text.
18. The document management system of claim 15 , wherein the generated feature vector comprises at least one feature vector entry based on a mean-squared error loss function, a divergence loss function, a cross-entropy loss function, or a VGG loss function.
19. The document management system of claim 18 , wherein the mean-squared error loss function is based on a comparison between pixels of the first image version of the document and the denoised image of the document.
20. The document management system of claim 15 , wherein the machine-learned model is a convolutional neural network.