Nearly copied image detection method based on multi-target matching

Abstract

The invention discloses a nearly copied image detection method based on multi-target matching. The nearly copied image detection method comprises the following steps that step1, image preprocessing is performed, wherein local characteristic regions and characteristic vectors of all the characteristic regions are extracted from images selected from to-be-queried images and an image set respectively, and initial matching pair sets are screened according to the characteristic vectors of the characteristic regions; step2, characteristic geometric transformation space is established, similarity transformation matrixes corresponding to the initial matching pair sets are solved, wherein the characteristic geometric transformation space is obtained by just combining all the similarity transformation matrixes; step3, clustering is performed, wherein a nonlinear mean shift algorithm suitable for geometric space is used for gathering multiple independently distributed clusters in the characteristic geometric transformation space; step4, similarity magnitude is calculated, wherein the similarity magnitude of the images is defined according to the number of points in all the independently distributed clusters; step5, results are presented, the step1 to the step 4 are repeated, and after the steps are performed on the image data in the image set, result sets are detected according to similarity sequencing images.

Description

technical field [0001] The invention relates to a near-duplicate image detection method for multi-target matching, and belongs to the fields of computer vision, multimedia information technology, pattern recognition and the like. Background technique [0002] With the development of the Internet and multimedia technology, images have become an important content of multimedia information transmission on the Internet, and the corresponding research on images has also received extensive attention. Near-copy image detection, that is, to query images that are similar copies of the image to be checked from the image collection. This technology can be applied to news video retrieval, advertisement insertion detection, image forgery detection, image copyright protection, image retrieval deduplication, sub-image query, image spam filtering, etc. It is a new research direction in network image retrieval research. [0003] The research on near-duplicate image detection technology init...

AI Technical Summary

Problems solved by technology

The most traditional method in global image representation is to represent the image as a feature vector, query the approximate copy image through the distance between the feature vectors, ignore the object target in the image, and be robust to image translation, cropping and other transformation operations Poor; in the traditional method of approximate copy image detection based on local features, the local feature area of ​​the image is first detected and described, and the local area detection operators currently mainly used: MSER, Harris-Hessian Affine, DoG, etc. and feature description methods SIFT, and then realize the retrieval of a large number of local features, mainly using hash and quantization retrieval methods. Although the possible rotation, scaling, translation and other transformations of image objects are considered, the information and description of local descriptors are lost in quantization. The mutual positional relationship of visual words has polysemy, which reduces the accuracy rate

Images