What are Denoising autoencoders? They are very similar to the basic model we saw in previous sections, the difference is that, the input is corrupted before being passed to the network. By matching the original version (not the corrupted one) with the reconstruction at training time, this autoencoder gets trained to reconstruct the original input image from the corrupted image. The ability to reconstruct original image from corrupted image makes autoencoder very smart.
An additive noise autoencoder uses the following equation to add corruption to incoming data:
xcorr = x + scale*random_normal(n)
The following is the detail describe about the preceding equation:
- x is the original image
- scale is the multiplier for a random normal number generated from n
- n is the number of training samples
- xcorr is the corrupted output