Hi
I have a doubt concerning the number of filters in a specific layer when related to color images. In the input mage, we consider Nc to be the number of channels while from the first layer on we consider it to be the number of filters. If Nc[1] = 6, does that mean we have 6 filters each doing a task like vertical edge detection or does it mean it is a set of 2 filters with 3 channels one for each color? To put it simply do colors not matter beyond the input image? Thank you.
A kernel is a 2D matrix (K, K) that is part of a 3D feature detector. This feature detector is called a filter and it is basically a stack of 2D kernels. Each kernel is convolved with a 2D input channel (i.e. feature-map) so if there are Cin channels in the input, then there are Cin kernels in a filter (C == Cin)
Each filter is convolved with the entire input to create a single output channel (i.e. feature-map). If there are Cout output channels, then there are Cout filters (F == Cout).
But in the link provided by you, if you notice the channels keeps increasing as the image depth increasing, explaining image color more on grayscale basis.
But in case one wants to also reduce channel with dimensions, then you can reduce the number channels too and resultant output will have reduced dimension height, width and channel based on requirement.
The RGB channel are grayscale representations of each color channel, where white represents 100% color, and black represents no color.
So if your question is more from a color significance point of view, then it depends on the model being created like if someone is looking for multi-channel image detections or in case of medical imaging it is usually a black-white image detection, where a darker hue of white can be considered obstruction, as well as darker hue of black can be considered as soft tissue lesion.
Thank you for answering my question but Iām still a bit confused.
In the lectures, Andrew mentions that kernels and filters are synonymous with each other and he prefers to use the term filters. If my question is from a color significance point of view, then if the number of filters is f = 6 does that mean that each of those filters representing a feature detector contains 3 kernels each representing the RGB colours?
