Image Indexing By Structured Sparse Spectral Hashing

Due to high dimensional features of images, it's difficult to construct an efficient index for large scale of images. Semantic hashing is a good idea to solve this problem. However, the traditional semantic hashing based image indexing methods neglect the structural information embedded in images, which means the relationship of features and part information of images. NMF used in face recognition illustrates that it can improve the efficiency of image similarity analysis by utilizing the structural information.Motivated by the advantages of structure property of features and the significant part information of images, this paper introduces structured sparse principal component analysis into traditional spectral hashing to boost image retrieval, which is called structured sparse spectral hashing (S3H). S3H is not only similarity preserving but also structure preserving.Furthermore, in consideration of the different data distribution on different datasets, this paper further introduces Boosting SSC algorithm into SjH to determine the specific threshold for each dataset, so that S3H can adapt to different datasets more properly.In experiments section, this paper compares S3H with other indexing algorithms on feature-based datasets and pixel-based datasets. Experimental results demonstrate that SJH generally outperforms other algorithms like LSH, RBM, SH and SSH.