Attribute graph based image representation method

Abstract

The invention provides an attribute graph based image representation method. The image representation method is characterized in that an indirect complete graph (V, E), which called as an attribute graph, is adopted to represent an image; V={V1, Vg} is vertexes of the graph; V1 represents local nodes and describes all objects contained in the image; Vg represents global nodes and describes background information or the whole scene information of the image; E represents edges of the graph, wherein the edges are divided into local edges and global edges; and the edges in the attribute graph represent a spatial relationship between objects in the image and spatial positions of the objects in the whole image. According to the method provided by the invention, the attribute graph is adopted to represent an image, and images which contain the same objects but represent different image semantics can be accurately classified. The attribute graph not only considers local features of the image, but also considers global features of the image. Meanwhile, information of the edges is used to describe the spatial relationship between objects and the spatial positions of the objects in the whole image, thereby having positive significance for image classification and recognition.

Description

technical field [0001] The invention relates to an image representation method, in particular to an attribute graph-based image representation method, which belongs to the technical field of image recognition. Background technique [0002] Learning-based image recognition methods have made great progress in the past few years. For specific object classes, especially human faces and cars, there are relatively reliable and efficient recognition based on underlying features (such as SIFT features or HOG features). But the underlying features of these images cannot reflect the image category information well. [0003] Recently, new research material proposes methods for classification using intrinsic properties of images. Attributes refer to the characteristics that can be named by people and can be observed in the image. It can indicate whether the object in the image exists, can describe the color, shape, material, component, category and function of the object in the image,...

AI Technical Summary

Problems solved by technology

Fahadi et al (A.Farhadi, I.Endres, D.Hoiem, and D.Forsyth.Describing objects by their attributes.In CVPR, 2009.2, 5, 6) directly represent images with attributes and apply it to object classification; Patterson etc. (G.Patterson andJ.Hays.Sun attribute database: Discovering, annotating, and recognizing scene attributes.In CVPR, 2012.2) use attributes to describe scene images and apply it to the classification of scene images; While Lan et al. (T.Lan , W.Yang, Y.W.0003, and G.Mori.Image retrieval with structured object queries using latent ranking svm.In ECCV, 2012, 2) It is believed that the spatial relationship of all objects in the image can be established and used to represent the image, but it is not considered The geometric distribution of all objects throughout the image, and the characteristics of individual objects; Kulkarni (G. Kulkarni, V. Premraj, V. Ordonez, S. Dhar, S. Li, Y. Choi, A. C. Berg, and T. L. Berg. Babytalk : Understanding and generating simple image descriptions) established a model of the relationship between objects and objects, but did not take into account the overall background information of the image; Zheng et al. (L.Zheng, S.Wang, Z.Liu, and Q.Tian.Packing and padding: Coupledmulti-index for accurate image retrieval.In CVPR, 2014.2) Add the underlying feature SIFT and color information of the image to the multidimensional index, and use this method to represent the image. Douze et al. (M.Douze, A.Ramisa , and C.Schmid.Combining attributes and fisher vectors for efficient image retrieval.In CVPR, 2011.2, 6, 7) adding attributes to Fisher vectors to represent images, these techniques only consider the overall information of the image, but do not pay attention to the image objects and the local features of these objects; Cao et al. (X.C ao, X.wei, X.Guo, Y.Han, and J.Tang. Augmented image retrieval using multi-orderobject layout with attributes.In ACM MM, 2014.2, 6, 7) constructed an attribute-based triangular object structure to Represents an image, but does not consider the global scene information of the image;

[0004] To sum up, some existing technologies describe images directly with attributes, but ignore the spatial information of objects in the image; some use the relationship between objects to represent images, but do not take into account the overall background of the image information; some consider the overall characteristics of the image, but ignore the objects in the image and the local characteristics of these objects

Images