Image Indexing By Sparse Spectral Hashing

With the rapid improvement of Internet, the time has been arrived that the amount of multimedia data on the internet is growing at an exponential rate. Therefore, the traditional index and searching method is no longer fit for the request of speed and preciseness from user as the data on the Internet has been getting more and more large and high-dimensional. As a result, new index algorithm which can deal with large scale of high dimensional data efficiently and accurately is not only very important in theoretical area, but also very urgent for modern Internet. This is why the research of index algorithm becomes hot spot issue recent years both here and abroad.Since some new problems come out for high-dimensional data, for example, the similarity between data is more difficult to measure, this paper starts with semantic hashing, proposes to introduce sparse principle component analysis into traditional spectral hashing and obtain the embedding subspace of high-dimensional features, which can remove the effect of redundant information caused by over-completeness of feature, as a result the index result will be more efficient and accurate. We call this Sparse Spectral Hashing (SSH).Meanwhile, in order to have SSH adapted to real data more properly, we introduce Boosting algorithm in machine learning to determine the threshold, which makes this algorithm more adaptable and widely available. Experimental result shows us that SSH can be efficiently used to not only image indexing, but also video and text indexing, while the result is almost always better than traditional index method like LSH, RBM, PSH and SH.