Hi,
If I input an image of a dog and add noise and then the neural network predicts the noise and then it gets subtracted, then it will generate the image of a dog, but how it will generate other images from it, like a cat? Which video explains it and any link to a related research paper?
regards
Faiz
Hi @faizulhaque it is a good point.
To generate different types of image, such as a cat, you would need a separate diffusion model that was trained on a diverse dataset containing images of both dogs and cats. Each model is specialized in generating images from its specific dataset.
About the paper, I recommend:
“Denoising Diffusion Probabilistic Models” source: https://arxiv.org/pdf/2006.11239.pdf
1 Like
@nydia @faizulhaque
Thanks for the question and thanks for the refernce @Nydia. One question: Lets say I have a sensor with a particular type of noise. If I have a lot lot of images from this sensor:
a) can it get trained to know what the noise statistics are and automatically remove them? or would it need also the grown truth images from a different sensor?
b) what if there is no noise, but there is a camera distortion? Have people used NNs to take multiple images of camera (eg from a checker board, or even better a moving video of a stationary scene) to calibrate or auto-calibrate a camera?
Thanks in advance, Esfandiar
Hi, I think what you are looking is simpler than diffusion models. You can read more about time series and forecasting.
a) can it get trained to know what the noise statistics are and automatically remove them? or would it need also the grown truth images from a different sensor?
You can use directly the noise (the Fourier space signal) and apply do some cleaning on the data.
b) what if there is no noise, but there is a camera distortion? Have people used NNs to take multiple images of camera (eg from a checker board, or even better a moving video of a stationary scene) to calibrate or auto-calibrate a camera?
I think is is something that is widely used in the signal processing field.