Synthetic handwriting word data generation using GAN

Question

I am trying to generate synthetic handwritten data. I came across GAN being used to generate a single character image.

Generator: upsamples a random tensor to an image.

Discriminator: uses real data to classify real/generated.

But we can't control which character output we need.

I have a handwritten dataset containing images of words (IAM).

If for word level do we need separate GAN's for each word?

I need to annotate the output image to a word(for word recognitions).

Is there a GAN network that output a synthetic handwritten images for words, not in the training dataset?

Answer 1

You are using a GAN network which has two parts :

• Generator : It forges a random noise into an image. It is trained using the loss produced by the discriminator.

• Discriminator : It distinguishes the real and predicted images. Accordingly, loss is calculated.

The discriminator matches the real and predicted images. If I show the discriminator an image of character "2" then it will compare the prediction with that image. Accordingly the loss will be calculated and the generator will learn from that loss.

Hence, if you show the image of character "3" then the generator will try to generate images of character "3" so that it can fool the discriminator.

You by giving different images to the generator, you can forge the random noise into any desired output.