Region-based high-level semantics extraction with CEDD

Abstract

Given an image, our proposed model can extract its dominant high-level semantics information through low-level feature extraction and image classification. It contains 3 main parts: image segmentation, feature extraction and classification. To our knowledge, this is the first model that applies Color and Edge Directivity Descriptor (CEDD), a multiple feature extraction algorithm, into the high-level semantics extraction field. Further, we also introduce a new padding strategy for region representation, which is especially suitable for widely-used non-arbitrary over-segmentation. Finally, our experiment shows that CEDD performs equally or better than traditional texture-based Gabor method. Meanwhile, new padding strategy outperforms other relevant methods.