Figure 7. Illustration of self-supervised features’ learning using image denoising.

(left): A noisy image is created by injecting noise to the original image. (middle) an auto-encoder model learns representations by compressing the noisy image into a latent space (Z) via the encoder network, while the decoder tries to reconstruct the compressed image from the latent space. (right): A denoised image close to the original image. Image credit: Celine.