How does activations function work in CNN ?

Hello, I've trained a regression CNN for patches of size 16x16 and I want to process a Image of 1024x1024. I have tried the following:

1. Split the total image in patches of 16x16 then process each patch and joint all the patch outputs.
2. use the function activations with the total image.

The results in 2. was much better than 1., my question is, ¿How MATLAB process the total image in this case? The documentation of the function "activations" does not clarify this.

Answer 1

[Sai Bhargav Avula]

Hi,

Answering for the title, the activation function extracts the features of a trained network. For example, lets take pretrained VGG networks and your image dataset the activations function returns the features set from the images using the pretrained VGG. The features can be used for different utilities like transfer learning etc.,.

Coming to your question in the question section, it really depends on what you meant by the result from the total image is better ? activations function at the end returns the features extracted from the dataset that is passed to the network(pretrained) in single go.

For better understanding, the example in the link below explains in deeper

https://www.mathworks.com/help/deeplearning/ug/visualize-activations-of-a-convolutional-neural-network.html

Hope this helps!