That's pretty much how AI image compression works - you're describing standard autoencoder (encoder/decoder) architecture with some adaptiveness.