I have read through several articles that said stacking of conv layers consume lots of computer resources(I guess only run time not memory right?) https://www.jeremyjordan.me/semantic-segmentation/. But in regards to accuracy, why encoder/decoder is better? I think same level of deep conv layers(increase the channels) can extract same level of deeper information as encoder/decoder.

  • $\begingroup$ You need to give us some context. What exactly have you read? What are the sources for those statements? "Encoders" and "decoders" for images usually use convolutional layers, so it is unclear what you mean. $\endgroup$ – Tim Nov 4 '19 at 7:46