Generalized pooling pyramid with hierarchical dictionary sparse coding for event and object recognition

Feature coding and vector pooling are essential for image recognition in bag-of-visual-words (BoW) method. Encoding the low-level feature to rich one and pooling it without any information loss are very challenging works. In this paper, generalized pooling pyramid with hierarchical dictionary sparse coding is introduced to get rich sparse codes and alleviate the information loss in the phase of pooling. It includes two modules: First, with the low-level feature, hierarchical dictionary is learned for sparse coding to generate the hierarchical sparse representation. Second, in the phase of vector pooling, we present generalized pooling pyramid by utilizing the probabilistic function to model the statistical distribution of sparse codes. In the generalized pooling pyramid, the Fisher vectors which are computed with Gaussian Mixture (GMM) in different levels, are fused to represent the images. The performance of our method outperforms state-of-the-art performance in a large number of image categorization experiments on the event dataset (UIUC-Sport dataset) and the object recognition dataset (Caltech101 dataset).