Hierarchical mix-pooling and its applications to biomedical image classification

This paper introduces Hierarchical Mix-pooling (HMP), a translation-invariant image representation improving the discriminative power of pooling representations by capturing intermediate-size structure information in images. HMP consists of two levels, one traditional pooling (e.g., sum pooling) applied to intermediate-size regions to collect the statistics of local features, and one different pooling (e.g., max pooling) collecting statistics of the previously region-based pooled results. Classification experiments show that HMP considerably improves accuracies with much smaller sizes of dictionaries compared to traditional pooling. The superior performance of HMP is confirmed by experiments with different local features and classifiers on two public biomedical datasets (ICPR HEp-2 cells and IRMA radiology).