What is meant by "spatial encoding" in the context of convolutional neural networks?

Question

Consider the following excerpt from the abstract of the research paper titled Squeeze-and-Excitation networks by Jie Hu et al.

Convolutional neural networks are built upon the convolution operation, which extracts informative features by fusing spatial and channel-wise information together within local receptive fields. In order to boost the representational power of a network, several recent approaches have shown the benefit of enhancing spatial encoding.

The authors used the term "spatial encoding" and the excerpt implies that enhancing spatial encoding has the benefit of increasing the representational power of a convolutional neural network.

What is meant by the term "spatial encoding" in this context related to the convolutional neural networks?

Answer 1

The Convolution Layer processes a certain part of the picture tensors and compresses it to a lower dimension. The spatial encoding adds the information of where the pixels were located in the image. Loosely speaking: It tells you the pixels I just processed were in the top left corner of the image. That way, the classification in the fully-connected layer can additionaly use this information. For more information on CNNs click here.

Maybe it becomes clearer in text processing. Transformer models for NLP have so-called positional encoding which is the counterpart to spatial encoding in image processing. When the sentence is processed it, the positional encoding tells the index of the word in the whole sequence.