I really like your question. Because of your question I added Figure (1) and (2) in the article to make it clearer. For image classification, the output layer is a dense layer which outputs the labels. For image noise reduction, CNN is used in an autoencoder framework. I hope this helps. Thanks.