Analysis of activation maps through global pooling measurements for texture classification

We analyzed the effects of global pooling measurements on extracting relevant texture information from a given set of activation maps. Initially, using a layer-by-layer approach (GP-CNN), we experimentally demonstrated that layers at various depth levels could provide high-quality texture information. Based on this finding, we developed RankGP-CNN, a method that performs multi-layer feature extraction. More specifically, RankGP-CNN treats every CNN model as a vast collection of deep composite functions, where each function computes a 2D activation map for every input image. A feature ranking approach then assigns a score to each deep composite function by processing the activation maps generated for a particular dataset bank. Eventually, RankGP-CNN uses the top-ranked deep composite functions to compute feature vectors for different texture datasets. Experiments on a dedicated classifier showed that RankGP-CNN achieves good results and can adapt to different texture problems. Finally, we present RankGP-3M-CNN as the version of RankGP-CNN that considers multiple CNN models. Overall, RankGP-3M-CNN achieves promising results with the advantage of only using the default scale of the input images.