CN106777090A - The medical science big data search method of the Skyline that view-based access control model vocabulary is matched with multiple features - Google Patents

Abstract

The medical science big data search method of the Skyline that view-based access control model vocabulary is matched with multiple features, belong to intelligent medical treatment with big number treatment crossing domain, be applied in the middle of the medical Image Retrieval Technology based on content for metric space Skyline inquiries by the system, and technical essential is：Extract the characteristics such as SIFT, Color of medical image, multiple low-level image features of image are merged using distributed Skyline operations, each characteristic similarity as Skyline evaluation objective, the result of return is all to compare the similar or extremely similar candidate image of certain one-dimensional characteristic on multidimensional characteristic with query image, Liu Shi treatment finally is carried out using the Spark systems of cloud computing, and is inquired about in real time or result.Effect is：The corresponding information for getting picture in user terminal is uploaded and is saved in cloud server, and then cloud server is processed, and is obtained optimal medical image clustering schemes and is fed back to user.

Description

Skyline medical big data retrieval based on visual vocabulary and multi-feature matching method

technical field

The patent of the present invention belongs to the intersection field of smart medical care and large number processing. It is a Skyline medical big data retrieval system based on visual vocabulary and multi-feature matching. The system applies metric space Skyline query to content-based medical image retrieval technology. , involving large-scale medical data analysis, massive data processing in the cloud computing environment, and intelligent data processing and application development.

Background technique

With the development of the Internet and the popularization of medical digital equipment, medical image data has grown exponentially, and the retrieval technology of related image data has attracted more and more attention. Massive data not only has the characteristics of large data volume, but also contains Huge commercial value. For example, analyzing the tumor growth of medical cancer users can guide doctors to recommend relevant personalized treatment plans; analyzing brain activity and heart rate records can provide hospital manufacturers and patients with diagnosis and treatment guidance or pre-disease warning for family monitoring. However, the explosive growth of massive medical image data has made the traditional stand-alone data analysis and processing technology less and less suitable for the current intensive data analysis and processing. In order to improve the efficiency of medical image retrieval under the premise of ensuring the accuracy of image retrieval, The metric space Skyline query (MetricSkylineQuery) algorithm has been well applied in the field of image processing. The algorithm can improve the efficiency of image retrieval by pruning the data in the metric space.

Most existing metric space Skyline algorithms for image data are based on general text semantics for metric space modeling. In the semantic image retrieval method with medical background, although the semantic information of the image is rich, there are also shortcomings such as complex semantic information, subjective semantic understanding, semantic extraction and expression difficulties, which affect the metric space modeling and medical image retrieval. In addition, due to the ambiguity of semantic information, most algorithms select multiple images to participate in the query according to semantic needs in order to improve the query accuracy, which greatly increases the amount of calculation in the query process. The large amount of calculation has become a major bottleneck of Skyline query in metric space, which is especially prominent in the processing of massive medical image data.

In recent years, content-based image retrieval technology has developed rapidly, and has gradually become the mainstream technology in the field of image retrieval. Aiming at the shortcomings of existing metric space algorithms for medical image data to select image semantic information for retrieval, starting from the content of medical images, the underlying features of images are selected as the research object in metric space. In order to improve the retrieval accuracy, in order to save computational overhead and speed up the calculation of similarity distance, the metric space Skyline algorithm was designed from the perspective of multi-feature fusion. Based on this, we designed and implemented this invention patent.

Contents of the invention

According to the defects and deficiencies in the above-mentioned background technology, the present invention applies the metric space Skyline query to the content-based large-scale medical image retrieval technology, and proposes a large-scale medical image based on the fusion of visual vocabulary and Skyline multi-features The retrieval method (BigFeatureFusionbySkyline, BSKFF) uses Skyline operation to fuse multiple features, and designs a new medical big data retrieval system based on visual vocabulary, which better solves the problem of medical large-scale image data retrieval.

In order to achieve the above object, the technical solution adopted in this patent is:

A kind of medical big data retrieval method based on the Skyline of visual vocabulary and multi-feature matching, it is characterized in that, comprises the steps:

S1. Extract the underlying features of medical images, cluster the underlying feature sets respectively, and construct a visual vocabulary, thereby quantizing the images in the image library into a vector of the frequency of occurrence of visual words to obtain the partition feature vector;

S2. Calculate the similarity distance between the query image and any image in the image library on each feature to construct image similarity vectors of different features;

S3. Invoking the Skyline-based multi-feature fusion method for distributed retrieval calculation decision-making.

Further, the step S1. extracting the feature data of the medical image, given a query image, extracting the underlying features of the image, includes the following steps:

S1.1. Extraction of Color features;

S1.2. Extraction of SIFT features;

S1.3. Build a visual vocabulary;

S1.4. Image quantization representation.

Further, the method for constructing image similarity vectors of different features in the step S2 is: an image library containing n pieces of medical images and the query image q, the medical image is expressed as a feature vector, the similarity distance between the query image q and any image o _i in the image library I on the t-th feature, which is expressed as the L1 distance _of the two vectors:

in Represents the t-th feature descriptor vector of image o _i , which is the k-dimensional vector of the t-th dimension bottom layer feature of image o _i ;

Based on the formula 1.3, the similarity distance between the query medical image q and any image o _i in the medical image library I on each feature is obtained, and the similarity vector of the image q and o _i is shown in Definition 1.2:

Definition 1.2: Let is an image library containing n images, q is a query image, and the similarity vector between query image q and any image o _i in image library I is expressed as an m-dimensional vector:

Vect _i (o _i ,q)＝＜dist(o _i .x ₁ ,qx ₁ ),dist(o _i .x ₂ ,qx ₂ ),...,dist(o _i .x _m ,qx _m )＞

Where i∈[1,n], m represents the number of underlying features, Vect _i (o _i ,q) represents the similarity vector between image q and image o _i , dist(o _i .x _k ,qx _k ) represents two images The similarity distance of the kth (k≤m) dimensional feature; all the images in the image library I and the query image q calculate the similarity distance on each dimensional feature, and construct and generate n similarity vectors.

Further, the specific method of the step S3:

Given a medical image library containing n images and a query image q, the set R is the query result of the multi-feature fusion method, for the m underlying feature vectors of each image

When an image o _i ∈ R, if and only if it satisfies the following conditions:

Then the R set contains the similarity vector Vect _i (o _i ,q)=<dist(o _i .x ₁ ,qx ₁ ),dist(o _i .x ₂ ,qx ₂ ) with the query image q in the X vector space ,...,dist(o _i .x _m ,qx _m )>The collection of all images not dominated by any other image similarity vectors on the medical image library I;

Furthermore, the result set of the multi-feature fusion method based on Skyline is a subset of the medical image library, and the image set is not dominated by any image in the image set in the multi-feature metric space, the query image q and any image o _i The SIFT and Color feature similarity distance values constitute a point, the abscissa of the point represents the similarity distance of the SIFT feature between the image o ₁ and the query image q, and the ordinate represents the similarity distance of the Color feature between the image o ₁ and the query image q , the similarity distance is calculated based on the bag-of-words model in the multi-feature metric space, and the smaller the similarity distance is, the more similar they are.

Further, use Spark for stream processing, decompose stream computing into a series of short batch jobs, and gradually integrate and recommend decision results.

Further, the extraction method of step S1.1.Color feature is as follows:

The Color feature is represented by the color attribute CN descriptor, which is composed of red, black, blue, green, brown, gray, pink, orange, white, purple, and yellow colors. The color attribute CN is defined as an 11-dimensional variable, which is an image All pixels in the image are assigned a color attribute label, which is a main factor of Skyline multi-factor analysis, and Spark is used for stream processing, and the results are gradually improved and output;

Further, the method of step S1.2.SIFT feature extraction is as follows:

It consists of two parts: detecting feature points and describing feature points. Scale conversion is performed on the original image to obtain the scale space representation sequence of the image, and then the image is processed to obtain the feature points. The 128-dimensional descriptor vector is used to represent the feature points, and the total 128-dimensional SIFT feature vector, using the feature points generated during the SIFT feature extraction process, using the feature point and its surrounding area as a local area, extracting the CN vector of each pixel in the local area, and obtaining SIFT and CN local feature vectors , this vector is used as a main factor of Skyline multi-factor analysis, using Spark for stream processing, and the results are gradually improved and output;

Further, step S1.3. The method for constructing a visual vocabulary is as follows:

Through the Spark-based multi-layer clustering algorithm k-means and its variants and oversampling correction, the Spark system is used to perform streaming training on the images in the image library, and gradually generate visual vocabulary for SIFT and Color feature vectors, and generate For the visual vocabulary, the data is first segmented, and the Spark system is used for distributed processing in a streaming manner, and the result set is incrementally exported;

Among them, the multi-layer k-means clustering algorithm is to find _k cluster centers C= _{{ c 1 ,} c ₂ , ...,c _k }, so that the sum of square errors from each feature point to the center of the cluster is the smallest; these cluster centers divide X into k disjoint clusters Y={Y ₁ ,Y ₂ ,..., Y _k }, such that for any 1≤i≠j≤k, For a cluster Y _i , its center point is:

Among them, the oversampling correction algorithm uses a SparkSpark job to select the center point and calculate the global error (different from the traditional MapReduce, we use Spark and use distributed cache for processing to speed up the iteration speed. The result is flow-increasing way), its objective function is:

The goal of the OnR clustering algorithm generated in each decomposition stage is to find an optimal partition C, so that the final global clustering error φ _X (C) of Spark is the smallest, where φ _X (C) uses the central point set C, for The global clustering error generated by the feature set X division, || || is the Euclidean distance. The SIFT and CN feature sets are clustered separately, and the k cluster centers obtained are their visual vocabulary.

Further, step S1.4. The method of image quantization representation is as follows:

Based on the visual vocabulary generated by the clustering algorithm, the SIFT descriptor of each image is quantized as a bag of words full of words. In the visual bag of words model, the visual vocabulary of a given feature Where j=1,...,m, k is the number of words in the visual vocabulary, in the image library, each image is quantified as a k-dimensional vector of the frequency of a visual word, and the Color feature is processed in the same way Quantization processing, and quantize each image to generate a corresponding feature vector, for the multi-feature quantization process, and so on, until all features are quantized, and the feature vector shown in Definition 1.1 is obtained;

Definition 1.1: In each data partition, find an image library containing n images Assume that each image o _i has a set of underlying features m is the number of underlying features, and the feature vector of each image o _i is expressed as <o _i .x ₁ ,o _i .x ₂ ,...,o _i .x _m >.

Beneficial effects: The medical big data retrieval system will obtain the corresponding information of the image on the client side through related technologies, upload it and save it to the cloud server, and then the cloud server will perform distributed processing to obtain the best medical image clustering scheme and gradually feed it back to the user.

Description of drawings

The system model of the feature fusion method of Fig. 1 of the present invention;

Fig. 2 the feature fusion process based on Skyline of the present invention;

Fig. 3 is the pseudo code of the SKFF algorithm of the present invention.

detailed description

Embodiment 1: With reference to Fig. 1, it is a kind of medical big data retrieval system based on visual vocabulary and multi-feature matching Skyline, said system is made up of a cloud center service system and a mobile phone intelligent mobile client software system. Among them, the cloud service system is responsible for the distributed and gradual extraction of feature data such as SIFT and Color of medical images, and uses the Skyline operation to fuse multiple underlying features of the image. The similarity of each feature is used as the evaluation target of Skyline. After calculation by Spark, The results are returned step by step, and the final returned results are candidate images that are similar to the query image in terms of multi-dimensional features or extremely similar in one-dimensional features; our mobile medical software will perform hierarchical clustering of medical large-scale images as needed Send medical images to the cloud center service system and receive cloud requests.

As an example, the execution flow of the Skyline medical big data retrieval system based on visual vocabulary and multi-feature matching is that when a mobile user collects and sends a request for relevant medical image retrieval through a medical image scanning instrument, the cloud system Extract feature data such as SIFT and Color of medical images, use Skyline operation to fuse multiple underlying features of the image, get the best clustering scheme and return it to the user step by step, if the time is long enough, the final result will be given to the user, In the middle, the step-by-step confirmation of the business and the confirmation of the final and complete results can be carried out through the mobile communication platform.

The processing steps of the SIFT and Color feature data algorithms are as follows: the Color feature is represented by the color attribute ColorNames (CN) descriptor, the color attribute CN is defined as an 11-dimensional variable, and a color attribute label is assigned to all pixels in the image. Tags were used as a principal factor in the Skyline multivariate analysis. SIFT feature extraction is to perform scale conversion on the original image to obtain the scale space representation sequence of the image, and then use the 128-dimensional descriptor vector to represent the feature points, and obtain a total of 128-dimensional SIFT feature vectors. Use the feature points generated during the SIFT feature extraction process, use the feature points and their surrounding areas as local areas, extract the CN vector of each pixel in the local area, and obtain SIFT and CN local feature vectors, which are used as Skyline multi-factor A major factor in the analysis. Then we will use Spark to stream process the collected CN tags and feature vectors, and the results will be gradually improved and output. Based on the extraction method of SIFT and CN feature vectors, through the Spark-based multi-layer clustering algorithm k-means and its variants and oversampling correction, using the Spark system, the images in the large-scale medical image database are streamed for training, and respectively To gradually generate a visual vocabulary for SIFT and Color feature vectors, we first segment the data, and use the Spark system to perform distributed processing in a streaming manner, and incrementally export the result set; among them, the multi-layer k-means clustering algorithm is in Find k cluster centers in the set of feature points of some dimensions (such as grid or higher-dimensional space), so that the sum of squared errors from each feature point to the center of the cluster (lesion area) where it is located is minimized. These clustering centers divide the set of feature points into k disjoint clusters (focus areas), so that for any , the focus points can be calculated for one cluster (focus area).

Based on the visual vocabulary generated by the clustering algorithm, the SIFT descriptor of each image is quantized as a bag-of-words filled with words. In the bag-of-visual-words model, given a feature’s visual vocabulary Where j=1,...,m, k is the number of words in the visual vocabulary (that is, the number of cluster centers). Therefore, in the medical image database, each medical image is quantified as a vector (k-dimensional vector) of the frequency of occurrence of a visual word. The Color feature is quantized in the same way, and each image is quantized to generate a corresponding feature vector. For the quantization process of multi-features (m≥2), and so on until all features are quantized.

As another embodiment, the oversampling correction algorithm is defined as: in each iteration, oversampling and refining (Oversampling and Refining, referred to as OnR) uses a SparkSpark job to carry out center point selection and global error calculation (different from traditional MapReduce The reason is that we use Spark and use distributed cache for processing to speed up the iteration speed, and the results are incrementally streamed), the OnR method is inspired by the scalablek-means++ method, in addition to the oversampling factor, it uses another The oversampling factor further increases the number of center points selected in the Map stage.

In each data partition, find an image library containing n medical images and the query medical image q, according to S1, the medical image is expressed as a feature vector. Therefore, the similarity distance between the query image q and any image o _i in the image library I on the t-th feature can be expressed as the L ₁ distance between two vectors. According to the formula, we get the query image q and any image o i in the image library I The similarity distance of image o _i on each feature, then the similarity vector of image q and o _i can be expressed as the similarity distance of the kth (k≤m) dimensional features of two images. Calculate the similarity distance between all the images in the image database I and the query image q on each dimension feature, and construct n similarity vectors.

Referring to Figure 3, calculate the similarity between each image in the image library and the query image on the feature SIFT and Color, and obtain a two-dimensional image similarity vector set; further, the SIFT and Color features of the query image q and any image o _i The similarity distance value constitutes a point, and the distributed computing decision is made through the multi-feature fusion method based on Skyline. The smaller the similarity distance, the more similar the two are. We use Spark for stream processing, and the results are gradually fused and recommended for decision results. Users The results obtained will be progressively more accurate over time.

Embodiment 2: A Skyline medical big data retrieval system based on visual vocabulary and multi-feature matching, which mainly extracts feature data such as SIFT and Color of medical images, and uses distributed Skyline operations to fuse multiple underlying features of images , each feature similarity is used as the evaluation target of Skyline, and the returned result is a candidate image that is relatively similar to the query image in terms of multi-dimensional features or a certain dimensional feature is extremely similar, and finally uses the cloud computing Spark system for stream processing. And get query or processing results in real time. It can be divided into the following three stages:

The first stage: extracting the features of the image. Given a query image, extract the underlying features of that image. Proceed as follows:

S1.Color feature extraction;

S2.SIFT feature extraction;

S3. Building a visual vocabulary;

S4. Image quantization representation.

Further, step S1.Color features are represented by the color attribute ColorNames (CN) descriptor, which consists of 11 basic colors, namely red, black, blue, green, brown, gray, pink, orange, white, purple and yellow, Therefore, the color attribute CN is defined as an 11-dimensional variable, and a color attribute label is assigned to all pixels in the image. This label is used as a main factor in Skyline multi-factor analysis. We use Spark for stream processing, and the results are gradually improved and output.

Further, the step S2.SIFT feature extraction process consists of two parts: detecting feature points and describing feature points. Scale conversion is performed on the original image to obtain the scale space representation sequence of the image, and then the image is correlated to obtain the feature points. A 128-dimensional descriptor vector is used to represent the feature points, and a total of 128-dimensional SIFT feature vectors are obtained. Use the feature points generated during the SIFT feature extraction process, use the feature points and their surrounding areas as local areas, extract the CN vector of each pixel in the local area, and obtain SIFT and CN local feature vectors, which are used as Skyline multi-factor As a main factor of the analysis, we use Spark for stream processing, and the results are gradually improved and output;

Further, step S3. Based on the extraction method of SIFT and CN feature vectors, through the Spark-based multi-layer clustering algorithm k-means and its variants and oversampling correction, using the Spark system to perform streaming training on the images in the image library , and gradually generate a visual vocabulary for SIFT and Color feature vectors respectively. The difference between us and the previous visual vocabulary is that we use the first data segmentation, and use the Spark system to perform distributed processing in a streaming manner, and incrementally export the result set ;

Among them, the multi-layer k-means clustering algorithm is to find k feature point sets X={x ₁ ,x ₂ ,...,x _n } in some dimensions (such as grid or higher-dimensional space) Clustering center C={c ₁ ,c ₂ ,...,c _k }, so that each feature point is located at the center of the cluster (in the tumor image, these cluster centers represent tumor lesions, or possible lesions) The minimum sum of squared errors (SumofsquaredError, SSE). These cluster centers divide X into k disjoint clusters Y={Y ₁ ,Y ₂ ,...,Y _k }, such that for any 1≤i≠j≤k, For a cluster Y _i , its center point (ie centroid) is:

Among them, the oversampling correction algorithm uses a SparkSpark job to select the center point and calculate the global error (different from the traditional MapReduce, we use Spark and use distributed cache for processing to speed up the iteration speed. The result is flow-increasing way), its objective function is:

The goal of the OnR clustering algorithm generated in each decomposition stage is to find an optimal partition C that minimizes the final global clustering error φ _X (C) of Spark. Among them, φ _X (C) is the global clustering error generated by dividing the feature set X by using the central point set C, and || || is the Euclidean distance. The SIFT and CN feature sets are clustered separately, and the k cluster centers obtained are their visual vocabulary.

Further, step S4. Based on the visual vocabulary generated by the clustering algorithm, the SIFT descriptor of each image is quantized into a bag of words filled with words. In the bag-of-visual-words model, given a feature’s visual vocabulary Where j=1,...,m, k is the number of words in the visual vocabulary (that is, the number of cluster centers). Therefore, in the image library, each image is quantized as a vector (k-dimensional vector) of the frequency of occurrence of a visual word. The Color feature is quantized in the same way, and each image is quantized to generate a corresponding feature vector. For the quantization process of multi-features (m≥2), and so on, until all the features are quantized, the feature vector shown in Definition 1.1 is obtained.

Definition 1.1 (partition feature vector): In each data partition, find an image library containing n images Assume that each image o _i has a set of underlying features m is the number of underlying features, and the feature vector of each image o _i is expressed as <o _i .x ₁ ,o _i .x ₂ ,...,o _i .x _m >.

The second stage is feature matching. Distributed computing queries the image and the similarity of SIFT and Color of each image in each data partition in the image database. Proceed as follows:

S1. Given a medical image, use Spark to gradually extract its SIFT features and Color features, and then quantify its feature descriptors into feature vectors according to the generated visual vocabulary. We use Spark for stream processing, and the results are gradually extracted and Quantify;

S2. Calculate the similarity of each feature between medical images;

Further, step S2. There is an existing image library containing n medical images and query image q, medical images are expressed as feature vectors according to S1. Therefore, the similarity distance between the query image q and any image o _i in the image library I on the t-th feature can be expressed as the L ₁ distance between two vectors:

in Represents the t-th feature descriptor vector of image o _i , that is, the k-dimensional vector representing the t-th dimension bottom layer feature of image o _i .

Based on Equation 1.3, we obtain the similarity distance between the query medical image q and any image o _i in the medical image library I on each feature. Then the similarity vectors of images q and o _i are shown in Definition 1.2:

Definition 1.2 (image similarity vector): Let is an image library containing n images, q is a query image, and the similarity vector between query image q and any image o _i in image library I can be expressed as an m-dimensional vector:

Vect _i (o _i ,q)＝＜dist(o _i .x ₁ ,qx ₁ ),dist(o _i .x ₂ ,qx ₂ ),...,dist(o _i .x _m ,qx _m )＞

Where i∈[1,n], m represents the number of underlying features, Vect _i (o _i ,q) represents the similarity vector between image q and image o _i , dist(o _i .x _k ,qx _k ) represents two images The similarity distance of the kth (k≤m) dimensional feature.

Calculate the similarity distance between all the images in the image database I and the query image q on each dimension feature, and construct n similarity vectors.

The third stage is feature fusion. The similarity vectors of different features are constructed into a new vector, and the Skyline-based multi-feature fusion method (SKFF) is called for distributed computing decision-making. Finally, we use Spark for stream processing, and the results are gradually integrated and recommended for decision-making results. The results obtained by users will be gradually accurate at any time.

S1. Distributed calculation of the similarity between each image in the image library and the query image on the feature SIFT and Color, to obtain a two-dimensional image similarity vector set;

S2. Use Skyline's multi-feature fusion to perform feature fusion, and the result of the previous feature matching can be used as the input of Skyline operation;

S3. Use the Spark system of cloud computing to perform streaming processing, and obtain query or processing results in real time.

Further, the definition (4.1) of the multi-feature fusion method based on Skyline is given.

Definition 1.4 (Skyline-based multi-feature fusion method): Given a medical image library containing n images and a query image q, the set R is the query result of the multi-feature fusion method. For the m underlying feature vectors of each image The R set contains the similarity vector Vect _i (o _i ,q)=<dist(o _i .x ₁ ,qx ₁ ),dist(o _i .x ₂ ,qx ₂ ), ...,dist(o _i .x _m ,qx _m )＞The set of all images not dominated by any other image similarity vector on the medical image database I, that is, when an image o _i ∈R, if and only When the following conditions are met:

Furthermore, the result set of the Skyline-based multi-feature fusion method (SKFF) is a subset of the medical image database, and an image collection that is not dominated by any image in the multi-feature metric space. The SIFT and Color feature similarity distance values between the query image q and any image o _i constitute points, as shown in Figure 2. For example, the abscissa of point p ₁ represents the similarity distance of SIFT features between image o ₁ and query image q, The ordinate indicates the similarity distance of the Color features between them, and these distances are calculated based on the bag-of-words model in the multi-feature metric space.

Further, the smaller the similarity distance, the more similar they are, so {p ₁ ,p ₂ ,p ₃ ,p ₄ } is the final Skyline result, indicating that there is no other better image than {o ₁ ,o ₂ ,o ₃ ,o ₄ } are more similar to the query image on both SIFT and Color features, that is, there are no images in the image library whose similarity vectors to the query image dominate them on SIFT and Color features.

S3.Spark performs stream processing, gradually integrates and recommends decision results.

Further, in step S2, the final Skyline result is {p ₁ , p ₂ , p ₃ , p ₄ }.

Further, use Spark for stream processing, and decompose stream computing into a series of short batch jobs. The entire streaming computing can superimpose the intermediate results according to the needs of the business, or store them in an external device, and gradually feed back the best medical clustering scheme to the user.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto, any person familiar with the technical field within the technical scope of the disclosure of the present invention, according to the present invention Any equivalent replacement or change of the created technical solution and its inventive concept shall be covered within the scope of protection of the present invention.

Claims (10)

1. a kind of medical science big data search method of the Skyline that view-based access control model vocabulary is matched with multiple features, it is characterised in that Comprise the following steps：

S1. the low-level image feature of medical image is extracted, low-level image feature set is clustered respectively, build visual vocabulary table, with this, It is a vector for the vision word frequency of occurrences by the image quantization in image library, obtains partition characteristics vector；

S2. similarity distance of the arbitrary image in query image and image library in each feature is calculated, to construct different spies The image similarity vector levied；

S3. call the multiple features fusion method based on Skyline to carry out distributed search and calculate decision-making.

2. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 1 is matched with multiple features Method, it is characterised in that the step S1. extracts the characteristic of medical image, gives a query image, extracts the image Low-level image feature, comprises the following steps：

The extraction of S1.1.Color features；

The extraction of S1.2.SIFT features；

S1.3. visual vocabulary table is built；

S1.4. image quantization is represented.

3. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 1 is matched with multiple features Method, it is characterised in that the method for the image similarity vector of different characteristic is constructed in the step S2 is：One is cured comprising n width Learn the image library of imageWith query image q, medical image is expressed as characteristic vector, query image q and image library I In arbitrary image o_iSimilarity distance in t-th feature, its L for being expressed as two vectors₁Distance：

WhereinRepresent image o_iT-th Feature Descriptor vector, be image o_iT tie up low-level image feature k dimensional vectors；

Based on formula 1.3, the arbitrary image o inquired about in medical image q and medical image storehouse I is obtained_iIt is similar in each feature Degree distance, image q and o_iSimilarity vector as define 1.2 shown in：

Define 1.2：IfIt is the image library comprising n width images, q is query image, is appointed in query image q and image library I It is intended to as o_iSimilarity vector be expressed as m dimensional vectors：

Vect_i(o_i, q)=＜ dist (o_i.x₁,q.x₁),dist(o_i.x₂,q.x₂),...,dist(o_i.x_m,q.x_m) ＞

Wherein i ∈ [1, n], m represent low-level image feature number, Vect_i(o_i, q) represent image q and image o_iSimilarity vector, dist(o_i.x_k,q.x_k) represent two images kth (k≤m) dimensional feature similarity distance；All images point in image library I Do not calculate similarity distance, construction n similarity vector of generation on each dimensional feature with query image q.

4. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 1 is matched with multiple features Method, it is characterised in that the specific method of the step S3：

A given medical image storehouse comprising n width imagesIt is multiple features fusion side with width a query image q, set R The Query Result of method, for the m low-level image feature vector of each image

As piece image o_i∈ R, and if only if meets following condition：

Then R set contains with query image q the similarity vector Vect in X vector spaces_i(o_i, q)=＜ dist (o_i.x₁, q.x₁),dist(o_i.x₂,q.x₂),...,dist(o_i.x_m,q.x_m) ＞ is by other any image phases on the I of medical image storehouse Like the set of all images for spending vector domination.

5. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 4 is matched with multiple features Method, it is characterised in that the result set of the multiple features fusion method based on Skyline is the subset in medical image storehouse, and in Duo Te Levy in metric space the image collection do not arranged by arbitrary image in image set, query image q and arbitrary image o_iSIFT Point is constituted with Color characteristic similarities distance value, the abscissa of point represents image o₁The phase of SIFT feature between query image q Like degree distance, ordinate represents image o₁The similarity distance of Color features between query image q, the similarity away from It is all based on what bag of words were calculated from multiple features metric space, similarity is more similar between the two apart from smaller.

6. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 5 is matched with multiple features Method, it is characterised in that carry out stream process using Spark, streaming is calculated and resolves into a series of short and small batch processing jobs, gradually Fusion is recommended with the result of decision.

7. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 2 is matched with multiple features Method, it is characterised in that the method for the extraction of step S1.1.Color features is as follows：

Color features describe son to represent with color attribute CN, by red, black, blue, green, brown, grey, powder, orange, white, purple, yellow face Colour cell into, color attribute CN is defined as one 11 variable of dimension, be that all pixels assign a color attribute label in image, This label carries out stream process as a main factor of Skyline multiplicities using Spark, as a result gradually improve with it is defeated Go out.

8. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 2 is matched with multiple features Method, it is characterised in that the method for the extraction of step S1.2.SIFT features is as follows：

It is made up of detection characteristic point and Expressive Features point two parts, spatial scaling is carried out to original image, obtains the yardstick of image Space representation sequence, then to image process obtaining characteristic point, and characteristic point is represented using the description subvector of 128 dimensions, The SIFT features vector of totally 128 dimensions is obtained, with the characteristic point generated in SIFT feature extraction process, by characteristic point and its place Peripheral region extracts the CN vectors of each pixel in regional area as regional area, obtain SIFT and CN local features to Amount, this vector carries out stream process as a main factor of Skyline multiplicities using Spark, as a result gradually improve with Output.

9. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 2 is matched with multiple features Method, it is characterised in that the method that step S1.3. builds visual vocabulary table is as follows：

It is right using Spark systems by the multi-level clustering algorithm k-means based on Spark and its mutation and over-sampling amendment Image in image library carries out streaming training, and respectively SIFT and Color characteristic vectors progressively generate visual vocabulary table, generation During visual vocabulary table, using first cutting data, and Spark systems are used, distributed treatment is carried out in a streaming manner, and be incremented by derivation Result set；

Wherein, multilayer k-means clustering algorithms are the set of characteristic points X={ x in some dimensions₁,x₂,...,x_nMiddle searching k Cluster centre C={ c₁,c₂,...,c_k, make each characteristic point to the square error and minimum at place cluster center；In these clusters X is divided into k disjoint cluster Y={ Y by the heart₁,Y₂,...,Y_kSo that for arbitrary 1≤i ≠ j≤k,For a cluster Y_i, its central point is：

Wherein, over-sampling correction algorithm is the meter that center point selection and global error are carried out using a SparkSpark operation Calculate and (be that we employ Spark with traditional MapReduce differences, processed using distributed caching, to accelerate to change The speed of band, is as a result carried out in the way of streaming is incremented by), its object function is：

The target of the OnR clustering algorithms that each catabolic phase is produced is to find an optimal division C so that Spark is most Whole global clustering error φ_X(C) it is minimum, wherein φ_X(C) it is, using center point set C, the overall situation for producing to be divided to characteristic set X and is gathered Class error, | | | | it is Euclidean distance.SIFT and CN characteristic sets are clustered respectively, the k cluster centre for obtaining is i.e. It is their visual vocabulary tables.

10. the medical science big data retrieval side of the Skyline that view-based access control model vocabulary as claimed in claim 2 is matched with multiple features Method, it is characterised in that the method that step S1.4. image quantizations are represented is as follows：

Based on the visual vocabulary table of clustering algorithm generation, SIFT description of each image are quantified as a word for filling word Bag, in vision bag of words, gives a visual vocabulary table for featureWherein j=1 ..., m, k are visions The number of word in vocabulary, in image library, each image is quantified as a k dimensional vector for the vision word frequency of occurrences, with Identical mode carries out quantification treatment to Color features, and each image is quantified into the corresponding characteristic vector of generation, for The quantizing process of multiple features, by that analogy, until all features are quantized, obtains as defined the characteristic vector shown in 1.1；

Define 1.1：In each data partition, an image library comprising n width images is searchedIt is assumed that every width figure As o_iThere is one group of low-level image featureM is the quantity of low-level image feature, each image o_iCharacteristic vector be expressed as ＜ o_i.x₁,o_i.x₂,...,o_i.x_m＞.