Multi-characteristic Signature-based Image Retrieval Technology

The traditional float vector based image feature is the base of content based image retrieval techniques. Float vector based image feature has high dimension and usually makes use of Euclidean distance as its similarity definition. Although float vector based image feature can describe an image with precise, but there are also some drawbacks inside. For example, it costs more for storage and the distance computation is quite complex. With the number of images grows, it will be unsuitable for vector based image feature to stay in memory. It will also suffer from the time cost of distance computation.Aiming at the issues above, a signature based image feature has been put forward. By using Principal Component Analysis(PCA) and Vector Quantization(VQ), this method reduces the dimension of float feature vectors and then projects them onto feature signatures. Hamming Distance is introduced to represent the distance between feature signatures. Experiments demonstrate that, compared to float vector based image features, this method can solve the storage and matching issues of image features.By using signature based image feature, we stepped forward and proposed a multi-signature based image retrieval system and a multi-signature based duplicate image detecting method.After extracting multi-features from images, multi-signature based image retrieval system reduces the dimension of multiple float feature vectors and then projects them onto multi-signatures. Hamming distance is defined as the distance between images. Experiments show that the system can solve the storage and matching issues of feature in image retrieval system and it will return more query results. Moreover, the system has excellent feature extensibility.Duplicate image detection is a new issue oriented from the explosive growth of images on internet, whose task is to detect duplicate images in a search result page and replace all of them with one image, so as to improve user experience. As another application of feature signature, multi-signature based duplicate image detecting method improved the mapping and encoding during vector quantization. It is proved by experiments that the new mapping method has higher recall rate. In addition, compared to single signature, multi-signature can improve the recall rate further in duplicate image detection.