Adversarial network for transforming handwritten text
Described herein are systems, methods, and other techniques for training a generative adversarial network (GAN) to perform an image-to-image transformation for recognizing text. A pair of training images are provided to the GAN. The pair of training images include a training image containing a set of characters in handwritten form and a reference training image containing the set of characters in machine-recognizable form. The GAN includes a generator and a discriminator. The generated image is generated using the generator based on the training image. Update data is generated using the discriminator based on the generated image and the reference training image. The GAN is trained by modifying one or both of the generator and the discriminator using the update data.
1. A method of training a generative adversarial network (GAN) to perform an image-to-image transformation for recognizing text, the method comprising:
providing a pair of training images to the GAN, the pair of training images including a training image containing a set of characters in handwritten form and a reference training image containing the set of characters in machine-recognizable form, wherein the GAN includes a generator, a word-level discriminator, and a character-level discriminator, wherein the word-level discriminator is separate from the character-level discriminator;
providing the training image to the generator;
generating, using the generator, a generated image based on the training image;
providing the generated image and the reference training image to the word-level discriminator and to the character-level discriminator;
generating, using the word-level discriminator and the character-level discriminator, update data based on the generated image and the reference training image, the update data comprising (i) word-level update data generated by providing the generated image and the reference training image to the word-level discriminator and (ii) character-level update data generated by providing the generated image and the reference training image to the character-level discriminator; and
training the GAN by modifying at least one of the generator, the word-level discriminator, or the character-level discriminator using the update data.
2. The method of claim 1 , further comprising:
separating the generated image for and based on each of the set of characters; and
separating the reference training image for each of the set of characters.
3. The method of claim 2 , further comprising:
providing the separated generated image and the separated reference training image to the character-level discriminator;
generating, using the character-level discriminator, character-level update data based on the separated generated image and the separated reference training image, wherein the GAN is trained using the character-level update data.
4. The method of claim 1 , wherein the generator is further provided with a random input, and wherein the generated image is generated further based on the random input.
5. The method of claim 1 , further comprising:
providing additional pairs of training images;
generating additional update data using the additional pairs of training images; and
training the GAN by modifying at least one of the generator, the word-level discriminator, or the character-level discriminator using the additional update data.
6. A non-transitory computer-readable medium comprising instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
providing a pair of training images to a generative adversarial network (GAN), the pair of training images including a training image containing a set of characters in handwritten form and a reference training image containing the set of characters in machine-recognizable form, wherein the GAN includes a generator, a word-level discriminator, and a character-level discriminator, wherein the word-level discriminator is separate from the character-level discriminator;
providing the training image to the generator;
generating, using the generator, a generated image based on the training image;
providing the generated image and the reference training image to the word-level discriminator and the character-level discriminator;
generating, using the word-level discriminator and the character-level discriminator, update data based on the generated image and the reference training image the update data comprising (i) word-level update data generated by providing the generated image and the reference training image to the word-level discriminator and (ii) character-level update data generated by providing the generated image and the reference training image to the character-level discriminator; and
training the GAN by modifying at least one of the generator, the word-level discriminator, or the character-level discriminator using the update data.
7. The non-transitory computer-readable medium of claim 6 , further comprising:
separating the generated image for and based on each of the set of characters; and
separating the reference training image for each of the set of characters.
8. The non-transitory computer-readable medium of claim 7 , further comprising:
providing the separated generated image and the separated reference training image to the character-level discriminator;
generating, using the character-level discriminator, character-level update data based on the separated generated image and the separated reference training image, wherein the GAN is trained using the character-level update data.
9. The non-transitory computer-readable medium of claim 6 , wherein the generator is further provided with a random input, and wherein the generated image is generated further based on the random input.
10. The non-transitory computer-readable medium of claim 6 , further comprising:
providing additional pairs of training images;
generating additional update data using the additional pairs of training images; and
training the GAN by modifying at least one of the generator, the word-level discriminator, or the character-level discriminator using the additional update data.
11. A system comprising:
one or more processors; and
a computer-readable medium comprising instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
providing a pair of training images to a generative adversarial network (GAN), the pair of training images including a training image containing a set of characters in handwritten form and a reference training image containing the set of characters in machine-recognizable form, wherein the GAN includes a generator, a word-level discriminator, and a character-level discriminator, wherein the word-level discriminator is separate from the character-level discriminator;
providing the training image to the generator;
generating, using the generator, a generated image based on the training image;
providing the generated image and the reference training image to the word-level discriminator and the character-level discriminator;
generating, using the word-level discriminator and the character-level discriminator, update data based on the generated image and the reference training image, the update data comprising (i) word-level update data generated by providing the generated image and the reference training image to the word-level discriminator and (ii) character-level update data generated by providing the generated image and the reference training image to the character-level discriminator; and
training the GAN by modifying at least one of the generator, the word-level discriminator, or the character-level discriminator using the update data.
12. The system of claim 11 , further comprising:
separating the generated image for and based on each of the set of characters; and
separating the reference training image for each of the set of characters.
13. The system of claim 12 , further comprising:
providing the separated generated image and the separated reference training image to the character-level discriminator;
generating, using the character-level discriminator, character-level update data based on the separated generated image and the separated reference training image, wherein the GAN is trained using the character-level update data.
14. The system of claim 11 , further comprising:
providing additional pairs of training images;
generating additional update data using the additional pairs of training images; and
training the GAN by modifying at least one of the generator, the word-level discriminator, or the character-level discriminator using the additional update data.