This book "Deep Learning and Convolutional Neural Networks for Medical Image Computing" mentioned a term spatial feature encoding

On the other hand, CNN models have been proved to have much higher modeling capacity, compared to the previous image recognition mainstream pipelines, e.g., HAAR, SIFT, HOG image features followed by spatial feature encoding, then random forest or support vector classifiers. Given millions of parameters to fit during model training (much more than previous pipelines), CNN representation empowers and enables computerized image recognition models, with a good possibility to be able to handle more challenging imaging problems. The primary risk is overfitting since model capacity is generally high in deep learning but often very limited datasets are available (that are with good quality of labels to facilitate supervised training). The core topics of this book are represented by examples on how to address this task-critical overfitting issue with deep learning model selection, dataset resampling and balancing, and the proper quantitative evaluation protocols or setups.

enter image description here

I googled a lot and didn't find explanation or definition of "spatial feature encoding".

What is "spatial feature encoding"? Can anyone give a concrete example?


Chatfield et al. provides a good survey of feature encoding methods.

From the introduction to that paper (emphasis my own):

The typical object recognition pipeline is composed of the following three steps: (i) extraction of local image features (e.g., SIFT descriptors), (ii) encoding of the local features in an image descriptor (e.g., a histogram of the quantized local features), and (iii) classification of the image descriptor (e.g., by a support vector machine).

I believe that "spatial feature encoding" refers to step (ii) in the typical pipeline. In this respect, I believe that the simplest possible example (and a standard one) would be the histogram encoding. If "spatial feature encoding" is meant to imply the preservation of some geometric information, than spatial binning is another technique to look at.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.