Method and system for removing noise in documents for image processing
A method and system are provided for removing noise from document images using a neural network-based machine learning model. A dataset of original document images is used as an input source of images. Random noise is added to the original document images to generate noisy images, which are provided to a neural network-based denoising system that generates denoised images. Denoised images and original document images are evaluated by a neural network-based discriminator system, which generates a predictive output relating to authenticity of evaluated denoised images. Feedback is provided backpropagation updates to train both the denoising and discriminator systems. Training sequences are iteratively performed to provide the backpropagation updates, such that the denoising system is trained to generate denoised images that can pass as original document images while the discriminator system is trained to improve the accuracy in predicting the authenticity of the images presented.
1. A method for removing noise from document images, the method comprising:
receiving one or more scanned paper document images comprising text to be identified using optical character recognition;
adding random noise to the text of the one or more document images to generate one or more noisy images;
removing noise from the one or more noisy images via a first neural network-based denoising system to generate one or more denoised images;
evaluating the one or more denoised images via a second neural network-based discriminator system to generate a predictive output relating to authenticity of the one or more denoised images; and
using feedback from a first backpropagation update to train the first neural network-based denoising system to denoise document images comprising text to be identified using optical character recognition, wherein the feedback is a propagation of errors based on the gradient of an error function of the discriminator system.
2. The method according to claim 1 , further comprising:
evaluating the one or more document images via the second neural network-based discriminator system to generate a predictive output relating to authenticity of the one or more document images; and
using feedback from a second backpropagation update to train the second neural network-based discriminator system.
3. The method according to claim 2 , further comprising:
evaluating the one or more denoised images via the second neural network-based discriminator system to generate a predictive output relating to authenticity of the one or more denoised images; and
using feedback from a third backpropagation update to train the second neural network-based discriminator system.
4. The method according to claim 3 , wherein the predictive output relating to authenticity is an indicator of whether the evaluated one or more denoised images is determined as being a synthetically-generated document image corresponding to the one or more noisy images or an original document image corresponding to the one or more document images.
5. The method according to claim 3 , wherein the first, second and third backpropagation updates correspond to respective computations of errors by the second neural network-based discriminator system.
6. The method according to claim 5 , wherein a plurality of training sequences are iteratively performed to provide the respective first, second and third backpropagation updates.
7. The method according to claim 6 , wherein the plurality of training sequences are iteratively performed to facilitate learning, by the second neural network-based discriminator system, of underlying patterns of noise in the one or more noisy images.
8. The method according to claim 3 , wherein the second neural network-based discriminator system receives an equal number of the one or more denoised images and the one or more document images, and wherein the one or more denoised images and the one or more document images are provided to the second neural network-based discriminator system on an unpaired basis at different times.
9. The method according to claim 3 , wherein the one or more documents images constitute a synthetic dataset comprising original document images.
10. The method according to claim 9 , wherein the first neural network-based denoising system and the second neural network-based discriminator system are trained to reach an equilibrium state, whereby the first neural network-based denoising system generates denoised output images such that the predictive output is indicative that the denoised output images are not distinguishable from the original document images.
11. A system for removing noise from document images, the system comprising a processor, for executing computer program instructions stored in a memory, which when executed by the processor, cause the processor to perform operations comprising:
receiving one or more scanned paper document images comprising text to be identified using optical character recognition;
adding random noise to the text of the one or more document images to generate one or more noisy images;
removing noise from the one or more noisy images via a first neural network-based denoising system to generate one or more denoised images;
evaluating the one or more denoised images via a second neural network-based discriminator system to generate a predictive output relating to authenticity of the one or more denoised images; and
using feedback from a first backpropagation update to train the first neural network-based denoising system to denoise document images comprising text to be identified using optical character recognition, wherein the feedback is a propagation of errors based on the gradient of an error function of the discriminator system.
12. The system according to claim 11 , the operations further comprising:
evaluating the one or more document images via the second neural network-based discriminator system to generate a predictive output relating to authenticity of the one or more document images; and
using feedback from a second backpropagation update to train the second neural network-based discriminator system.
13. The system according to claim 12 , the operations further comprising:
evaluating the one or more denoised images via the second neural network-based discriminator system to generate a predictive output relating to authenticity of the one or more denoised images; and
using feedback from a third backpropagation update to train the second neural network-based discriminator system.
14. The system according to claim 11 , wherein the predictive output relating to authenticity is an indicator of whether the evaluated one or more denoised images is determined as being a synthetically-generated document image corresponding to the one or more noisy images or an original document image corresponding to the one or more document images.
15. The system according to claim 13 , wherein the first, second and third backpropagation updates correspond to respective computations of errors by the second neural network-based discriminator system.
16. The system according to claim 15 , wherein a plurality of training sequences are iteratively performed to provide the respective first, second and third backpropagation updates.
17. The system according to claim 16 , wherein the plurality of training sequences are iteratively performed to facilitate learning, by the second neural network-based discriminator system, of underlying patterns of noise in the one or more noisy images.
18. The system according to claim 13 , wherein the second neural network-based discriminator system receives an equal number of the one or more denoised images and the one or more document images, and wherein the one or more denoised images and the one or more document images are provided to the second neural network-based discriminator system on an unpaired basis at different times.
19. The system according to claim 13 , wherein the one or more documents images constitute a synthetic dataset comprising original document images.
20. The system according to claim 19 , wherein the first neural network-based denoising system and the second neural network-based discriminator system are trained to reach an equilibrium state, whereby the first neural network-based denoising system generates denoised output images such that the predictive output is indicative that the denoised output images are not distinguishable from the original document images.