Fast multiresolution image querying

摘要: We present a method for searching in an image database using query that is similar to the intended target. The may be hand-drawn sketch or (potentially low-quality) scan of retrieved. Our algorithm makes use multiresolution wavelet decompositions and images. coefficients these are distilled into small “signatures” each image. introduce “image querying metric” operates on signatures. This metric essentially compares how many significant has common with potential targets. includes parameters can tuned, statistical analysis, accommodate kinds distortions found different types queries. resulting simple, requires very little storage overhead signatures, fast enough performed 20,000 images at interactive rates (on standard desktop machines) as sketched. experiments hundreds queries databases 1000 show dramatic improvement, both speed success rate, over conventional L1, L2, color histogram norm. CR