Whats the difference between a convolutional autoencoder (CAE) and a convolutional neural network (CNN)

I'm working on a bachelor's project that involves using a convolutional autoencoder [1]. I used the code from this blog. Now the goal was to make a model that could take as input a pixelated image with text and as output, predict the image with depixelated text. The only change I made from the "convolutional autoencoder" code in the reference is that I also gave labels to my training process. After training several models, I concluded that it is pretty easy to reconstruct pixelated text.

Now while I'm writing a paper about the project, I'm really struggling to understand what exactly is a convolutional autoencoder and what makes it a convolutional autoencoder.

I'm completely new to any type of ML. When I did research on autoencoders in general, I found that autoencoders are neural networks that aim to minimize the difference between the output and input. And what makes it a "convolutional" autoencoder is the fact that it uses convolutions in its encoder to detect edges etc.

-But now in my case I depixelated pixelated text in images, so the output is not meant to be as close as possible to the input image but rather be the depixelated version.

-Another thing is that everywhere I look for answers, it is stated that autoencoders are primarily used for unsupervised learning. While in my case, I use supervised learning since I pass the labels to the training process.

-Lastly, since I don't try to minimize the difference between the output and input image, AND I use supervised learning, then what exactly is the difference between a convolutional autoencoder and a convolutional neural network?

[1] https://blog.keras.io/building-autoencoders-in-keras.html

An auto-encoder (AE) is a self-supervised neural network model, that aims to learn a representation of its inputs, e.g. images. I'll explain myself:

• The AE is made of two NNs: the encoder f(x) -> z, and the decoder g(z) -> x; a third important piece is the z, called the latent space or bottleneck.

• AEs aim at learning z, where its size is smaller than x, by minimizing a reconstruction objective. Basically the latent space is a compact representation of x.
• There are many variations of AEs, one is called Denoising AE (DAE) and is the one your are using, because DAE aims at learning a "clear" version of x that is corrupted by some noise.
• A CAE is an AE in which the encoder and decoder are CNNs. Actually, there are more strict definitions which require z to be a 3D tensor (4D if you consider the batch size too) called the convolutional bottleneck (because it has the three dimensions as an image).
• The AE is usually self-supervised because you its own inputs are supervision: the images.

So a CAE is an AE with convolutional layers, and a convolutional bottleneck. Instead, you can consider a CNN as a building block for other kinds of models too: classifiers, regressors, AEs itself, and others.