Exploiting local dependencies with spatial-scale space (S-Cube) for near-duplicate retrieval

Image Near-Duplicate (IND) plays an important part in many real-world multimedia applications. At the same time, both the accuracy and speed are key problems in INDs. This paper presents an efficient and effective solution for retrieving Image Near-Duplicate. Different from previous methods, we analyze the local dependencies among the descriptors in the spatial-scale space (S-Cube). Such local dependencies in spatial-scale space (S-Cube) encodes not only visual appearance but also the spatial and scale co-occurrence of them. The local dependencies are exploited over the cube-space of neighboring spatial locations and multiple adjacent scales to form the new image representation, which is invariant to spatial transformation and scale change. ed up the retrieval process, the SuperNodes are built to incorporate the neighbor information. We evaluate our proposed spatial-scale (S-Cube) method for IND retrieval using two existing benchmarks as well as a new dataset extracted from the keyframes of TRECVID corpus. Compared to the state-of-the-art results, our proposed local dependencies in S-Cube plus SuperNodes approach has shown a high accuracy for IND retrieval, as well as a significant time reduction.