Per-pixel classification using deep learning models

Question

I want to train a model to classify image pixels in which neighbouring pixels are not considered, only channels (bands) for each pixel. I'm thinking about defining a CNN model which stacks several layers of 1x1 convolutions preserving the size of the image up to the output (only the number of channels varies).

I'd like to receive some feedback on whereas this is an overkill and inefficient use of CNNs and on alternative deep learning approaches that can lead to more optimal models.

Answer 1

This would be equivalent to running the same fully connected network for each pixel.

As for "efficiency", the different memory access patterns of CNNs and fully connected networks might mean that the fully connected implementation could be faster than a 1x1 conv implementation. I don't know if this is a significant difference though.