A Universal Visual Dictionary Learned from Natural Scenes for Recognition

摘要

Inspired by the efficient coding hypothesis and simple-to-complex cell hierarchy of the visual system, we study a universal visual dictionary learned from natural scenes using sparse coding for recognition. The vocabularies are similar to V1 simple cells receptive fields. Max pooling is done in a local region ("block") so that the features are translation invariant, which is the function of complex cells. Macro-features of a grid of overlapping spatial blocks are built and fed to a linear SVM classifier for recognition. We have tested the learned universal visual dictionary on different recognition tasks and demonstrated the effectiveness of the model.