Generally, CNNs are used to extract feature representations of an image. I'm right now dealing with the class of CNN that produces saliency maps, which are generally in the format of a mask. I'm trying to generate a feature representation of that specific Mask. What could be the best way to approach this problem?
I found a method to do it in the paper Cross-Modality Personalization for Retrieval (2020, accessed: 20-Feb-2020).
Representation. For images, we extract Inception-v4 CNN features . We then mask the image convolution feature with the BubbleView saliency map, by resizing the saliency map to the convolution feature size and multiplying them together. Finally, average pooling is performed to obtain a 1536-dimensional feature vector. We represent textual descriptions as their average 200-dimensional Glove embedding . For personality, we use a 10-dimensional feature vector containing the scores for the personality questions in . Below, we describe how we learn projections of these representations that place them in the same feature space.