CN1710557A - Impliet semanteme picture searching method based on non-negative array decomposition - Google Patents

Abstract

Using prototype image library, the method constructs matrix of feature-image. Using training algorithm of non-negative matrix decomposition decomposes the matrix. A low dimensional semantic space is constructed by using decomposed basis matrixes. Each column of vector in coefficient matrix is semantic feature relevant to each image in prototype image library. Then, using test algorithm of non-negative matrix decomposition projects queried image to the semantic space to obtain semantic feature. Finally, in the low dimensional semantic space, similarity measurement is carried out between semantic features of prototype images and semantic features of queried image, and result of searching images is returned to user according to magnitude of similarity. The method possesses higher searching accuracy rate.

Description

Implicit semantic image search method based on the nonnegative matrix decomposition

Technical field

The present invention relates to a kind of implicit semantic image search method that decomposes based on nonnegative matrix, relate to fields such as pattern-recognition, matrix analysis and image retrieval, can directly apply to CBIR.

Background technology

The eighties of last century later stage along with continuing to bring out of extensive image data base, causes people's attention gradually for big data quantity management of information like this and effective application, and CBIR just becomes a hot research field.Traditional image indexing system is when being described picture material, directly from view data, analyze mostly and extract bottom visual signature (for example color of image, shape, texture, spatial relationship etc.), and the people can't directly obtain from view data the understanding of picture material, judge according to people's knowledge.This process combines a large amount of experiences that accumulate in the daily life.(Liu Zhongwei such as Liu Zhongwei, Zhang Yujin, based on the image inquiry and the searching system of feature, application foundation and engineering science journal 2000.8 (1): 69-77) inquiry and the search method of having inquired into single features such as utilizing color, texture, shape and having fully utilized different characteristic.But because the feature of using all is the visual signature of bottom, they can't react these experimental knowledgees, and therefore the image retrieval on traditional bottom visual signature basis can not be obtained good effect.Realize that like this mapping between bottom visual signature and the high-level semantic just becomes the key that addresses this problem, but in fact the visual signature of Yu Yi notion and bottom is not to concern one to one, that is to say that similar low-level image feature may represent different semantic concept, and different low-level image features may corresponding similar semantic content, the polysemy and the synonymy problem that exist in Here it is the CBIR.

Exist similar problem in text retrieval, for example similar word has the different meanings, expresses and similar semanteme has different words.(LatentSemantic Indexing, LSI) technology solves this problem (Zhou Wen, Gong Liming, Jiang Lan, implicit semantic retrieval and Chinese sample analysis example, computer utility, 2004.24 (6): 273-276) to the implicit semantic indexing of utilization in the text retrieval.It by to word frequency-document matrix utilization svd (Singular Value Decomposition, SVD) the contact model between Mi Jianli word document, thereby find semantic relation implicit between word and the document, improved the retrieval performance of system.

When this technology is applied in the image retrieval, feature-image array can be decomposed the linear space that obtains a low-dimensional by svd, set up the contact between feature and image, thereby find semantic relation implicit between low-level image feature and the picture material.But promptly include on the occasion of component in the matrix that obtains with svd the negative value component is arranged also, what it was represented but is the frequency (as the word frequency in the text retrieval) that feature occurs, and the negative value component just can't carry out proper explanations like this.Under this external given dimension, it is the sample distribution Normal Distribution that svd can keep the precondition of good approximation, but this hypothesis usually can't satisfy for the frequency that feature in the image retrieval occurs, and the distribution of Poisson or other nonnegativities is more suitable.

Summary of the invention

The objective of the invention is to deficiency at above-mentioned implicit semantic indexing method based on svd, propose a kind of implicit semantic image search method, improve the precision of image retrieval based on nonnegative matrix decomposition (Non-negative Matrix Factorization).

For realizing this purpose, the present invention utilizes the prototype image library to come construction feature-image array, at first use nonnegative matrix decomposition-training algorithm this matrix is carried out matrix decomposition, and construct the semantic space of a low-dimensional with the basis matrix that obtains after decomposing, and each column vector of matrix of coefficients is exactly the semantic feature of each image correspondence in the prototype image library.With nonnegative matrix decomposition testing algorithm query image is projected to this semantic space then and obtain semantic feature.At last, in this low-dimensional semantic space, carry out the similarity measurement of the semantic feature and the query image semantic feature of all prototype figure pictures, and return to the user according to big young pathbreaker's retrieving images result of similarity.

The implicit semantic image search method that decomposes based on nonnegative matrix of the present invention carries out as follows:

1. structural attitude-image array: at first each image in the prototype image library is all extracted colourity-saturation degree and mix histogram feature, structural attitude-image array then, each row of this matrix are corresponding to an image, and each row mixes the one-component of histogram feature corresponding to colourity-saturation degree.

2. constructing semantic space and generate the semantic feature of prototype figure picture: use nonnegative matrix decomposition-training algorithm and feature-image array is decomposed obtain basis matrix and matrix of coefficients, construct the semantic space of a low-dimensional with the base vector of this basis matrix.The value of the dimension r of semantic space will satisfy (n+m) r＜nm, and on behalf of colourity-saturation degree, n mix the dimension of histogram feature herein, and m represents the number of prototype figure picture.Decompose the matrix of coefficients that obtains this moment is the projection of prototype figure picture at this semantic space, so can be regarded as the semantic feature of prototype figure picture.

3. the generation of query image semantic feature: for query image, at first extract colourity-saturation degree and mix histogram feature color histogram feature, use nonnegative matrix to decompose testing algorithm then the low-level image feature of query image is projected in the semantic space that second step obtained, thereby obtain the semantic feature of this query image.

4. similarity measurement and result return: by the step of front, obtained the semantic feature of each prototype figure picture and the semantic feature of query image.Calculate the similarity of the semantic feature of each prototype figure picture of query image and database at last, sort, the some width of cloth images the most similar to retrieving images are returned according to the size of similarity.

Method of the present invention can obtain the higher search accuracy rate.Because the implicit semantic indexing method that decomposes based on nonnegative matrix can decompose the semantic space that obtains a low-dimensional with feature-image array, has set up the contact between feature and image, can find semantic relation implicit between low-level image feature and the picture material.Directly use in the bad application of the direct retrieval effectiveness of bottom visual signature at some, method of the present invention has more use value.

The implicit semantic image indexing system based on the nonnegative matrix decomposition that the present invention sets up can be used for retrieving the image that the user needs more accurately based on picture material and semantic retrieval.

Description of drawings

Fig. 1 obtains basis matrix and matrix of coefficients synoptic diagram among the present invention with nonnegative matrix decomposition-training algorithm characteristics of decomposition-image array.

Fig. 2 is the query image and the result for retrieval synoptic diagram of the embodiment of the invention.

Embodiment

Below in conjunction with specific embodiment technical scheme of the present invention is described in further detail.

The image data base that the embodiment of the invention adopts has 500 samples, stores from the image of the various semantic classess of network collection, comprising: outdoor landscape, plant, automobile, animal, indoor landscape, artificial building etc.The bottom visual signature adopts colourity-saturation degree to mix histogram feature.The feature vector representation, $T=\left\{{\stackrel{\&RightArrow;}{x}}_l\right\}.$ (l＝1，2，...，500)， ${\stackrel{\&RightArrow;}{x}}_l=\{x_{\mathrm{<figure-callout\; id="1"\; label="l"\; filenames="A20051002682400081.png"\; state="\{\{state\}\}">l</figure-callout>}1},x_{\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="A20051002682400081.png"\; state="\{\{state\}\}">l</figure-callout>}2},\&CenterDot;\&CenterDot;\&CenterDot;,x_{\mathrm{lp}},\&CenterDot;\&CenterDot;\&CenterDot;,x_{l100}\}$ Contain 100 features.

Return 12 images the most similar with query image at every turn.

The total system implementation procedure is as follows:

1. structural attitude-image array:

At first each image in the prototype image library is all extracted colourity-saturation degree and mix histogram feature, structural attitude-image array V then ^P, its size is 100 * 500, wherein 100 is dimensions that colourity-saturation degree is mixed histogram feature, and the 500th, the number of prototype figure picture.

2. constructing semantic space and generate the semantic feature of prototype figure picture:

With nonnegative matrix decomposition-training algorithm with feature-image array V ^PBe decomposed into basis matrix W ^PWith matrix of coefficients H ^PHere the dimension of lower dimensional space is made as r, and then the size of basis matrix and matrix of coefficients is respectively 100 * r and r * 500.W wherein ^PEach row comprise a base vector, according to the character of nonnegative matrix decomposition algorithm, W ^PIn all base vectors just constructed a semantic space, and H ^PEach row can regard V as ^PIn each image in the feature of this semantic space.If the n of j image dimension color histogram is characterized as $V_j^P={\{V_{j1}^P,V_{j2}^P,\&CenterDot;\&CenterDot;\&CenterDot;V_{\mathrm{jn}}^P\}}^T,$ So corresponding coefficient vector is $H_j^P={\{H_{j1}^P,P_{j2}^P,\&CenterDot;\&CenterDot;\&CenterDot;,H_{\mathrm{jr}}^P\}}^T,$ It is exactly the semantic feature of j prototype figure picture.Provided the synoptic diagram of this step among Fig. 1.In Fig. 1, can see that feature-image array V is broken down into basis matrix W and matrix of coefficients H.Each sample all is expressed as the low-level image feature vector of n dimension in feature-image array, shows j image pattern as grey color part.And each sample all is expressed as the semantic feature of r dimension in matrix of coefficients, grey color part wherein just j image pattern at the corresponding coefficient of semantic space.

3. the generation of query image semantic feature:

For query image q, at first extract colourity-saturation degree and mix histogram feature $V^q={\{V_1^q,V_2^q,\&CenterDot;\&CenterDot;\&CenterDot;V_n^q\}}^T,$ Size is 100 * 1, uses nonnegative matrix to decompose testing algorithm then the low-level image feature of query image projected in the semantic space that second step obtained, thus the semantic feature of obtaining

$H^q={\{H_1^q,H_2^q,\&CenterDot;\&CenterDot;\&CenterDot;H_r^q\}}^T.$

4. similarity measurement and result return:

Obtained the semantic feature of each prototype figure picture in the image library respectively (by H by nonnegative matrix decomposition-training and testing algorithm ^PEach tabulation show) and the semantic feature H of query image ^q, calculate at last each prototype figure picture of query image and database apart from d (H ^q, H _j ^P), j=1 ... 500, and some images nearest apart from query image are returned to the user as result for retrieval.Fig. 2 is the query image and the result for retrieval synoptic diagram of the embodiment of the invention.In Fig. 2, the image on the left side is a query image, the snow-capped snow mountain of the content of image, the image on the right is the result who retrieves out with method of the present invention, in 12 images that return, in semantically relevant (including snow mountain) 10 width of cloth images are arranged with query image as can be seen, incoherent have only 2 width of cloth.

Claims (1)

1, a kind of implicit semantic image search method that decomposes based on nonnegative matrix is characterized in that comprising as follows

Concrete steps:

1) structural attitude-image array: at first each image in the prototype image library is all extracted colourity-saturation degree and mix histogram feature, structural attitude-image array then, each row of this matrix are corresponding to an image, and each row mixes the one-component of histogram feature corresponding to colourity-saturation degree;

2) constructing semantic space and generate the semantic feature of prototype figure picture: use nonnegative matrix decomposition-training algorithm and feature-image array is decomposed obtain basis matrix and matrix of coefficients, then construct the semantic space of a low-dimensional with this basis matrix, the value of the dimension r of semantic space will satisfy (n+m) r＜nm, wherein on behalf of colourity-saturation degree, n mix the dimension of histogram feature, m represents the number of prototype figure picture, decompose the matrix of coefficients that obtains this moment is the projection of prototype figure picture at this semantic space, can be regarded as the semantic feature of prototype figure picture;

3) generation of query image semantic feature: for query image, at first extract colourity-saturation degree and mix histogram feature color histogram feature, use nonnegative matrix decomposition testing algorithm that the low-level image feature of query image is projected in the semantic space that obtains previously then, thereby obtain the semantic feature of this query image;

4) similarity measurement and result return: according to the semantic feature of each the prototype figure picture that obtains and the semantic feature of query image, calculate the similarity of the semantic feature of query image and each prototype figure picture of database, and sort according to the size of similarity, the some width of cloth images the most similar to retrieving images are returned.

Images