Learning primitive and scene semantics of images for classification and retrieval
Cheong Yiu Fung, Kai Fock Loe · 1999
We present a learning-based semantics approach for classifying and retrieving images.Our approach defines semantics at two levels: (1) primitive semantics at the patch level, extracted automatically from pixel characteristics of patches with supervised learning; and (2) scene semantics at the image level, recognized from the association of primitive semantics of the patches in a self-organizing manner.Images are classified and retrieved according to the similarity in scene semantics.Our experiments so far have yielded highly accurate scene classification results and very promising retrieval performance on a set of diverse natural scene images.