Efficient Extraction of Object Level Signature for Large-Scale Image Copy Detection

The problems of conventional ‘visual invariant’ detection methods were high computational complexity, indiscriminate processing in foreground and background regions, and generating redundant and low distinctive features. An object level image signature called BLIP is proposed to deal with these problems. At first, we develop efficient affine-invariant local detector and descriptor, speeding up the extraction from second magnitude to millisecond. Next, improved BING is used to detect foreground object regions, and the context of BLIP is encoded into binary signatures. Experiments demonstrate BLIP is robust to common copy transformations in Internet. The signature can be extracted in 8.9 ms and only consumes 0.17%, 2% memory cost of SIFT and BRISK separately.