# What is a feature vector and how do they differ from feature maps

I'm reading a paper which reviews image upscaling techniques. A model named SRCNN is described as:

• First layer: creates feature maps from low-resolution input images
• Second layer: converts these feature maps into high dimensional feature vectors

source snippet: https://prnt.sc/vy2sah

From what I understand a feature map (in this context) is the output of a filter that has been applied to an image.

What is a "high dimensional feature vector" and how do they differ from feature maps?

Sometimes, it's better to go to the original paper, from which I borrowed the below figure.

Each of the patches in the original image is represented with a single value in the second stage (after convolution). There are $$n_1$$ feature maps so each patch is represented with $$n_1$$ values, where $$n_1=64$$. They call this 64-length representation as a high-dimensional vector, which I believe is not that high (but that's subjective).

Moreover, I don't think the description in the survey is accurate, because the authors of the original paper talks about these high dimensional vectors in the first layer, while the explanation in the survey says the patches are converted into high-dimensional vectors in the second (nonlinear transform) layer:

Patch extraction and representation: this operation extracts (overlapping) patches from the low resolution image Y and represents each patch as a high-dimensional vector

Non-linear mapping: this operation nonlinearly maps each high-dimensional vector onto another high-dimensional vector.

What I'd normally expect from "converting patches to high dimensional vectors" is flattening, but it's not the case in this paper. It follows three convolutional layers, but due to the convolution sizes, the operations performed has other interpretations.

• Thanks for all your effort. Your answer really helped me :) Dec 8, 2020 at 13:37