Method and system for extracting image fingerprints based on representative local mode

Abstract

The invention discloses a method and system for extracting image fingerprints based on a representative local mode, and relates to the field of image processing. The method comprises: performing image attack analog processing of database images to generate new database images; extracting key points of the database images and the new database images; obtaining local blocks according to the key points; generating local modes and establishing a local mode database according to the local blocks; obtaining the representative local mode from the local mode database; establishing image fingerprints of the database images and the new database images according to the representative local mode; storing the image fingerprints into an image fingerprint database; obtaining a new image and extracting fingerprints of the new image; comparing the fingerprints of the new image with the image fingerprints in the image fingerprint database; and searching an image corresponding to the new image in the database images. According to the invention, the required memory is small; accelerated matching is performed by using an optimized machine instruction; the method and system are suitable for massive image copying and detection.

Description

technical field [0001] The invention relates to the field of image processing, in particular to an image fingerprint extraction method and system based on representative local patterns. Background technique [0002] Image copy detection has always been one of the hot research areas of computer vision. Its general strategy is to extract a few features from the image and integrate them into a feature vector that can reflect the content of the image. From the perspective of the composition of image features, it is generally divided into global features and local features. The following is a brief discussion of the development status of these two features. As early as 1999, Naphade was the first method to extract the histogram of the image in the LUV color space as an image signature, and then Mohan used the block method to extract the average gray value of each block, according to the gray value of these blocks Size sorting construction features; a relatively influential image...

AI Technical Summary

Problems solved by technology

[0003] The application of existing global features and local features to large-scale image copy detection applications mainly has the following shortcomings and deficiencies: Poor robustness: the existing global features are constructed based on the overall information of the image such as the color histogram, so when the image occurs In the case of partial occlusion and cropping, it will fail; feature extraction speed is slow: for large-scale image retrieval, the extraction speed of image features is very high, especially for emerging mobile phone platforms, whose hardware performance is limited, so feature extraction is required The complexity should not be too high

At present, the mainstream global descriptor GIST needs to calculate five kinds of feature descriptions, and the local feature SIFT needs to calculate time-consuming DOG operators. These algorithms cannot be applied to large-scale image copy detection; memory consumption is large: feature matching complexity is high, In order to ensure its robustness and differentiation, the existing feature descriptions often have a high dimensionality. Taking the SIFT descriptor as an example, the algorithm divides the blocks around the feature points into 16 regions, and calculates 8 Gradient direction histogram in each direction, a total of 128-dimensional floating-point features, if each floating-point type occupies 4 bytes, a SIFT feature consumes 500 bytes of memory, and an image will have hundreds or thousands SIFT features, such high-dimensional features bring enormous pressure to storage and matching

Images