I am currently working on interpreting a trained CNN model used for image recognition. Specifically, I want to understand how color information is represented in the feature vector, particularly in the last layer that I use for tasks such as cosine similarity or matching with enrollments.
Consider the scenario where I have two images, and I want to identify the specific part of the feature vector that corresponds to color differences between these images.
I am seeking guidance on methods or techniques to extract information from the model that can help me distinguish color-related features within the feature vector.
What approaches or tools can be employed to achieve this? Any insights or suggestions would be greatly appreciated.
I found this answer but I am not sure about it.
Neural networks are all about taking raw input data (RGB values and pixel location) and learning useful features that are relevant to some downstream tasks. This process of aggregating raw inputs into higher-level features can start at the first layer past the inputs.
So yes, only the first layer of the network uses the actual raw color information from the image. Beyond that, the network has already started to put together nearby pixels and disparate color channels to find more complex patterns. Deeper layers in the neural network typically do further aggregation on features learned in earlier layers, rather than taking raw color information as input.
ref: https://ai.stackexchange.com/questions/5738/is-color-information-only-extracted-in-the-first-input-layer-of-a-convolutional