What are 'Convolutional Autoencoders', and how do they apply to image denoising?

Example answer

A convolutional autoencoder uses convolutional operations to encode an image and decode an output under a reconstruction-related objective. A bottleneck is common, but every autoencoder is not required to have a lower-dimensional latent representation; architecture and regularization determine what information is retained.

For denoising, one approach trains on paired noisy inputs and cleaner reference targets. Another synthetically corrupts an available image and uses the uncorrupted version as a self-supervised target. The latter learns the supplied corruption model and does not establish that the original target was noise-free. Methods that learn from noisy-only observations need additional assumptions; the autoencoder architecture alone does not remove the target-data problem.

The network estimates an output from evidence and learned patterns. It does not necessarily separate a truly noise-free latent state or recover details that were destroyed. Smoothing can erase a small real object, and a plausible reconstruction can invent structure.

Evaluate on independent data representative of the actual noise and signal. Use suitable reference measures, visual inspection and downstream errors, including detail preservation and unexpected artifacts. Compare with a simpler denoising baseline and measure runtime on the target hardware before claiming real-time suitability. For consequential uses, image clarity alone is insufficient validation.

Source: denoising autoencoder objectives.