Real-Time Image Semantic Retrieval Based on VQ

Image semantic retrieval usually involves two steps, namely image annotation and annotation-based retrieval. Although there is a rich literature on image auto-annotation, little focuses on bridging the gap between annotation and retrieval. In fact, the efficiency of an image semantic retrieval system depends not only on the precision of annotation but also the way to use the annotation result in the retrieval step. This paper proposes a novel scheme for image semantic retrieval based on Vector Quantization (VQ). The annotation and retrieval steps, mapped as the encoding and decoding processes of VQ respectively, are closely linked by the VQ codebook. Experiments on the general image database show that the retrieval efficiency has been improved dramatically to the real-time level.