SVD Based Robust Image Content Retrieval

Abstract

In some image content retrieval applications, the query maybe some image processed, geometrically transformed, or noise contaminated versions of their original ones, and therefore requires robustness against these unmalicious modifications, while possessing good discrimination ability for different image contents. We investigate a robust way for image content retrieval based on singular value decomposition to improve the performance of content discrimination. The singular value is used to gain robustness against geometrical variance. To gain higher robustness against other signal processing modifications, an adaptive image thresholding method is used as a preprocessing to SVD. Our experiments are based on 1,000 original images (with different contents) and their 23,000 modified versions, showing improved results of discrimination ability for image contents.