CN114996493A - Electric power scene image data screening method based on data elimination and redundancy elimination - Google Patents

Abstract

The invention discloses a power scene image data screening method based on data elimination and redundancy elimination, belongs to the technical field of data screening, and aims to provide a power scene image data screening method based on data elimination and redundancy elimination, which solves the problem of low image data screening efficiency caused by data redundancy of the existing power scene image data. Aiming at data characteristics under different service scenes, a Hash-based graph retrieval method and a local sensitive Hash approximate nearest neighbor statistical coding method are fused, and a retrieval result is obtained by combining specific local graph characteristics obtained by utilizing a deep neural network convolution layer, so that the optimization of a characteristic extraction and nearest neighbor search technology related to image data is realized, a characteristic vector capable of efficiently representing an image is extracted, rapid visual content search is carried out, and efficient screening of the image data is achieved. The method is suitable for the electric power scene image data screening method based on data elimination and redundancy elimination.

Description

Electric power scene image data screening method based on data elimination and redundancy elimination

Technical Field

The invention belongs to the technical field of data screening, and particularly relates to a power scene image data screening method based on data elimination and redundancy elimination.

Background

In an electric power scene, because electric power is large, data types are various, data quantity is large, and screening efficiency is low due to data redundancy, how to eliminate redundancy among data is one of key links for data screening.

With the popularization and application of multimedia and the development of information and network technologies, a large amount of various visual contents are generated, such as remote sensing images of power routing inspection and the like. Eliminating redundancy of image data requires comparison between all images, direct comparison of images is certainly impossible, and a comparison method based on image visual content retrieval is an effective method for solving the problems.

Disclosure of Invention

The invention aims to: the method for screening the image data of the power scene based on the data elimination and redundancy elimination solves the problem that the screening efficiency of the image data is low due to data redundancy of the image data of the existing power scene.

The technical scheme adopted by the invention is as follows:

a power scene image data screening method based on data elimination and redundancy elimination comprises the following steps:

(1) the image retrieval method based on the Hash is characterized in that high-dimensional and detail invariance characteristics of images in an image database are extracted, a characteristic space vector is constructed, and the characteristics are subjected to Hash coding;

(2) inputting a query image, extracting high-dimensional and detail invariance characteristics of the query image based on a Hash image retrieval method, constructing a characteristic space vector, and carrying out Hash coding on the characteristics;

(3) carrying out similarity sequencing and rearrangement based on Hamming distance between Hash codes to obtain a target retrieval result; and finishing image data screening.

Further, in the step (2), the image retrieval method based on the hash is coded into a binary hash through feature learning, and XOR operation is utilized, so that the calculation response time is shortened while the memory consumption is reduced;

and expanding the data into a high-dimensional space, dividing the space into a plurality of sub-regions by adopting a random hyperplane to construct a hash function, performing barrel sorting storage in an index stage of hash coding by combining with the characteristic data, and obtaining final neighbor by mapping and comparing to form a local sensitive hash which does not utilize data in the stage of constructing the hash function and has no training relaxation constraint in the stage of constructing the hash function.

Further, a graph retrieval method based on Hash and an approximate nearest neighbor statistical coding method based on locality sensitive Hash are fused, a retrieval result is obtained by combining specific local graph characteristics obtained by utilizing a deep neural network convolution layer, and the method specifically comprises the following steps:

s1, selecting the number of preset cluster types, generating a quantizer through clustering, quantizing the original features of the image, filtering the noise of the features, enabling similar features to be matched, and establishing an index;

s2, calculating the distance between the candidate list feature vector obtained by establishing the index and the feature vector of the query image, Reranking the result, and finally returning the nearest neighbor result;

by utilizing a Hamming distance threshold value, the length of reordering candidates is reduced and quantization noise and retrieval complexity are balanced when the vector is embedded into a binary Hash space.

Further, the mapping matrix learning algorithm for vector embedding into the binary hash space is as follows:

inputting: learning set

L _N In order to learn the number of vectors contained in the set,

let d be the original characteristic latitude, d _b The dimension of the binary characteristic after the embed is obtained;

step a: randomly generating a Gaussian matrix, performing QR orthogonal decomposition, and extracting front d of orthogonal matrix Q _b The row vector construction dimension is d _b x _d An orthogonal projection matrix P;

step b: for each vector x in the learning set omega _i Projecting by using matrix P to obtain vector set

Step c: z obtained by step b _ih Obtaining a median projection vector Tl, Tlh ═ mean { Z ] of the first cluster _ih },h∈[1,d _b ]；

Step d: quantizing x with a coarse quantizer q, and projecting x to z using P;

step e: mapping x into a binary vector using a fine quantizer;

and (3) outputting: projection matrix P, binary vector. .

In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:

1. according to the invention, aiming at data characteristics under different service scenes, a Hash-based graph retrieval method and a local sensitive Hash approximate nearest neighbor statistical coding method are fused, and a retrieval result is obtained by combining specific local graph characteristics obtained by utilizing a deep neural network convolution layer, so that the optimization of the characteristic extraction and nearest neighbor searching technology related to image data is realized, a characteristic vector capable of efficiently representing an image is extracted, rapid visual content searching is carried out, and the efficient screening of the image data is realized.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are required to be used in the embodiments will be briefly described below, it should be understood that the following drawings only illustrate some embodiments of the present invention and therefore should not be considered as limiting the scope, and that for those skilled in the art, other relevant drawings can be obtained according to the drawings without inventive efforts, wherein:

FIG. 1 is a schematic flow chart of a hash-based image retrieval method according to the present invention;

FIG. 2 is a schematic diagram of the approximate nearest neighbor search of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all embodiments of the present invention.

Thus, the following detailed description of the embodiments of the present invention, presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

A power scene image data screening method based on data elimination and redundancy elimination comprises the following steps:

(1) the image retrieval method based on the Hash is characterized in that high-dimensional and detail invariance characteristics of images in an image database are extracted, a characteristic space vector is constructed, and the characteristics are subjected to Hash coding;

(2) inputting a query image, extracting high-dimensional and detail invariance characteristics of the query image based on a Hash image retrieval method, constructing a characteristic space vector, and carrying out Hash coding on the characteristics;

(3) carrying out similarity sequencing and rearrangement based on Hamming distance between Hash codes to obtain a target retrieval result; and finishing image data screening.

Further, in the step (2), the image retrieval method based on the hash is coded into a binary hash through feature learning, and XOR operation is utilized, so that the calculation response time is shortened while the memory consumption is reduced;

and expanding the data into a high-dimensional space, dividing the space into a plurality of sub-regions by adopting a random hyperplane to construct a hash function, performing barrel sorting storage in an index stage of hash coding by combining with the characteristic data, and obtaining final neighbor by mapping and comparing to form a local sensitive hash which does not utilize data in the stage of constructing the hash function and has no training relaxation constraint in the stage of constructing the hash function.

Further, a graph retrieval method based on Hash and an approximate nearest neighbor statistical coding method based on locality sensitive Hash are fused, and a retrieval result is obtained by combining specific local graph characteristics obtained by utilizing a deep neural network convolution layer, wherein the method specifically comprises the following steps of:

s1, selecting the number of preset cluster types, generating a quantizer through clustering, quantizing the original features of the image, filtering the noise of the features, enabling similar features to be matched, and establishing an index;

s2, calculating the distance between the candidate list feature vector obtained by establishing the index and the feature vector of the query image, Reranking the result, and finally returning the nearest neighbor result;

by utilizing a Hamming distance threshold value, the length of reordering candidates is reduced and quantization noise and retrieval complexity are balanced when the vector is embedded into a binary hash space.

Further, the mapping matrix learning algorithm for vector embedding into the binary hash space is as follows:

inputting: learning set

L _N In order to learn the number of vectors contained in the set,

let d be the original characteristic latitude, d _b The dimension of the binary characteristic after the embed is obtained;

step a: randomly generating a Gaussian matrixPerforming line QR orthogonal decomposition to extract the front d of an orthogonal matrix Q _b The row vector construction dimension is d _b x _d An orthogonal projection matrix P;

step b: for each vector x in the learning set omega _i Projecting by using matrix P to obtain vector set

Step c: z obtained by step b _ih The median projection vector Tl, Tlh ═ mean { Z } is obtained for the first cluster of calculations _ih },h∈[1,d _b ]；

Step d: quantizing x with a coarse quantizer q and projecting x to z using P;

step e: mapping x into a binary vector using a fine quantizer;

and (3) outputting: projection matrix P, binary vector.

In the implementation process of the invention: a method for eliminating data redundancy by data visual content retrieval comparison is provided. In order to retrieve new training samples from samples obtained from iteratively learned training data distributions, the synthetic data is treated as instances of content data information variables derived from the content retrieval learned distributions of a given training set. In the aspect of image retrieval, as the retrieval result of an object or a target is influenced by factors such as environment, equipment and the like, the global expression of the image is constructed by different coding modes on the basis of selecting invariance local features with better anti-interference performance in a search mode, and the overall higher retrieval precision is realized.

1. The image retrieval method based on Hash comprises the following steps:

the image retrieval method based on the Hash realizes the extraction of high-dimensional and detail invariance features of an image to construct a feature space vector, so that the learning of a Hash function is carried out, the Hash coding is carried out on the features, and the similarity sequencing and the rearrangement are carried out by combining the Hamming distance between the codes to obtain a target retrieval result, as shown in FIG. 1.

2. Approximate nearest neighbor search

The image retrieval method based on the Hash is coded into binary Hash through feature learning, and the XOR exclusive OR operation in the computer is utilized, so that the memory consumption is reduced, and the calculation response time is shortened. Extending to a high-dimensional space, as shown in (a) of fig. 2, adopting a random hyperplane to divide the space into a plurality of sub-regions to construct a hash function, performing bucket ordering storage in an indexing stage of hash coding by combining feature data, and obtaining final neighbors through mapping comparison to form a local sensitive hash which does not utilize data in the hash function constructing stage and has no trained relaxation constraint in the hash function constructing stage, as shown in (b) of fig. 2.

3. Deep neural network retrieval algorithm based on SPoC characteristics

A graph retrieval method based on Hash and an approximate nearest neighbor statistical coding method of locality sensitive Hash are fused, and a retrieval result is obtained by combining specific local graph characteristics obtained by utilizing a deep neural network convolution layer. The aggregation characteristic of the image is obtained by adopting the aggregation characteristic SPoC generated by accumulation summation, and specific characteristic vectors of the high-efficiency representation image are obtained by continuously adjusting parameters and learning training data points for a network according to the data characteristics under different service scenes and optimizing the related characteristic extraction and neighbor search technology, so that the rapid visual content search is carried out.

The retrieval performance of large-scale high-dimensional vectors is generally divided into search optimization and vector optimization, and the search optimization of data is considered, namely, the performance optimization is carried out by optimizing a retrieval structure, the vectors are not changed, a matching target with approximate nearest distance is found by using an approximate nearest neighbor ANN algorithm, the search space is reduced, and the search efficiency can be improved. The local sensitive hash algorithm is characterized in that two adjacent data points in an original data space are still adjacent with a high probability after the same mapping or projection transformation is carried out, so that all data in an original data set can be subjected to hash mapping, the data points are subjected to hash function mapping transformation and are regularly dispersed into buckets of hash tables, the original data set is divided into subsets with a plurality of adjacent buckets and a small number of data, the problem is converted into the problem of searching adjacent elements in a small space domain, and the calculated amount is reduced. In addition, the hash function needs to satisfy the following two conditions:

1) if d (x, y) < ═ d1, then the probability of h (x) ═ h (y) is at least p 1;

2) if d (x, y) > < d2, the probability of h (x) > < h (y) is at most p 2;

where d (x, y) represents the distance between x and y, d1< d2, h (x) and h (y) represent the hash of x and y, respectively.

Considering that visual content information under an electric power scene is clear, a typical approximate nearest neighbor algorithm of inverted file index is adopted, and the whole optimization retrieval structure process is divided into two steps:

s1, selecting a proper cluster number, generating a quantizer through clustering, quantizing the original features of the image, wherein the quantization operation can filter the noise of the features, so that similar features can be matched, and further, an index is established. But noise is introduced, the cluster number K is small, the complexity is low, the reverse candidate elements are more, the rearrangement complexity is high, and the noise is also high.

And S2, calculating the distance between the candidate list feature vector obtained by establishing the index and the feature vector of the query image, Reranking the result, and finally returning the nearest neighbor result. By utilizing a Hamming distance threshold value, the length of reordering candidates is reduced and quantization noise and retrieval complexity are balanced when the vector is embedded into a binary hash space. The mapping matrix learning algorithm of the vector embedding process is as follows:

inputting: learning set

L _N In order to learn the number of vectors contained in the set,

let d be the original characteristic latitude, d _b The dimension of the binary characteristic after the embed is obtained;

step a: randomly generating a Gaussian matrix, performing QR orthogonal decomposition, and extracting front d of orthogonal matrix Q _b The row vector construction dimension is d _b x _d An orthogonal projection matrix P;

step b: for each vector x in the learning set omega _i Projecting by using matrix P to obtain vector set

Step c: z obtained by step b _ih Obtaining a median projection vector Tl, Tlh ═ mean { Z ] of the first cluster _ih },h∈[1,d _b ]；

Step d: quantizing x with a coarse quantizer q and projecting x to z using P;

step e: mapping x into a binary vector using a fine quantizer;

and (3) outputting: projection matrix P, binary vector.

Example 1

A power scene image data screening method based on data elimination and redundancy elimination comprises the following steps:

(1) the image retrieval method based on the Hash is characterized in that high-dimensional and detail invariance characteristics of images in an image database are extracted, a characteristic space vector is constructed, and the characteristics are subjected to Hash coding;

(2) inputting a query image, extracting high-dimensional and detail invariance characteristics of the query image based on a Hash image retrieval method, constructing a characteristic space vector, and carrying out Hash coding on the characteristics;

(3) carrying out similarity sequencing and rearrangement based on Hamming distance between Hash codes to obtain a target retrieval result; and finishing image data screening.

Example 2

On the basis of the embodiment 1, in the step (2), the image retrieval method based on the hash is coded into the binary hash through feature learning, and the XOR operation is utilized, so that the memory consumption is reduced, and the calculation response time is shortened;

and expanding the data into a high-dimensional space, dividing the space into a plurality of sub-regions by adopting a random hyperplane to construct a hash function, performing barrel sorting storage in an index stage of hash coding by combining with the characteristic data, and obtaining final neighbor by mapping and comparing to form a local sensitive hash which does not utilize data in the stage of constructing the hash function and has no training relaxation constraint in the stage of constructing the hash function.

Example 3

On the basis of the embodiment, a graph retrieval method based on Hash and an approximate nearest neighbor statistical coding method based on locality sensitive Hash are fused, and a retrieval result is obtained by combining specific local graph characteristics obtained by utilizing a deep neural network convolution layer, and the method specifically comprises the following steps:

s1, selecting the number of preset cluster types, generating a quantizer through clustering, quantizing the original features of the image, filtering the noise of the features, enabling similar features to be matched, and establishing an index;

s2, calculating the distance between the candidate list feature vector obtained by establishing the index and the feature vector of the query image, and Reranking the result, and finally returning the nearest neighbor result;

by utilizing a Hamming distance threshold value, the length of reordering candidates is reduced and quantization noise and retrieval complexity are balanced when the vector is embedded into a binary Hash space.

Example 4

On the basis of the above embodiment, the mapping matrix learning algorithm for vector embedding into the binary hash space is as follows:

inputting: learning set

L _N In order to learn the number of vectors contained in the set,

let d be the original characteristic latitude, d _b The dimension of the binary characteristic after the embed is obtained;

a, step a: randomly generating a Gaussian matrix, performing QR orthogonal decomposition, and extracting front d of orthogonal matrix Q _b The row vector construction dimension is d _b x _d An orthogonal projection matrix P;

step b: for each vector x in the learning set omega _i Projecting by using the matrix P to obtain a vector set

Step c: z obtained by step b _ih Obtaining a median projection vector Tl, Tlh ═ mean { Z ] of the first cluster _ih },h∈[1,d _b ]；

Step d: quantizing x with a coarse quantizer q and projecting x to z using P;

step e: mapping x into a binary vector using a fine quantizer;

and (3) outputting: projection matrix P, binary vector.

The above description is an embodiment of the present invention. The foregoing is a preferred embodiment of the present invention, and the preferred embodiments in the preferred embodiments can be combined and used in any combination if not obviously contradictory or prerequisite to a certain preferred embodiment, and the specific parameters in the embodiments and examples are only for the purpose of clearly illustrating the verification process of the invention and are not intended to limit the patent protection scope of the present invention, which is subject to the claims and all equivalent changes made by using the contents of the description and the drawings of the present invention should be included in the protection scope of the present invention.

Claims (4)

1. A method for screening electric power scene image data based on data elimination and redundancy elimination is characterized by comprising the following steps:

(1) the image retrieval method based on the Hash is characterized in that high-dimensional and detail invariance characteristics of images in an image database are extracted, a characteristic space vector is constructed, and the characteristics are subjected to Hash coding;

(2) inputting a query image, extracting high-dimensional and detail invariance characteristics of the query image based on a Hash image retrieval method, constructing a characteristic space vector, and carrying out Hash coding on the characteristics;

(3) carrying out similarity sequencing and rearrangement based on Hamming distance between Hash codes to obtain a target retrieval result; and finishing image data screening.

2. The method for screening power scene image data based on data elimination and redundancy elimination according to claim 1, wherein in the step (2), the image retrieval method based on hash is encoded into binary hash through feature learning, and XOR exclusive OR operation is utilized to reduce memory consumption and shorten calculation response time;

and expanding the data into a high-dimensional space, dividing the space into a plurality of sub-regions by adopting a random hyperplane to construct a hash function, performing barrel sorting storage in an index stage of hash coding by combining with the characteristic data, and obtaining final neighbor by mapping and comparing to form a local sensitive hash which does not utilize data in the stage of constructing the hash function and has no training relaxation constraint in the stage of constructing the hash function.

3. The method for screening image data of power scene based on data elimination and redundancy elimination according to claim 2, characterized by fusing a graph retrieval method based on hash and a approximate nearest neighbor statistical coding method based on locality sensitive hash, and obtaining a retrieval result by combining specific local graph features obtained by utilizing a deep neural network convolutional layer, comprising the following specific steps:

s1, selecting a preset cluster number, generating a quantizer through clustering, quantizing the original features of the image, filtering the noise of the features, enabling similar features to be matched, and establishing an index;

s2, calculating the distance between the candidate list feature vector obtained by establishing the index and the feature vector of the query image, and Reranking the result, and finally returning the nearest neighbor result;

by utilizing a Hamming distance threshold value, the length of reordering candidates is reduced and quantization noise and retrieval complexity are balanced when the vector is embedded into a binary hash space.

4. The method for screening power scene image data based on data elimination and redundancy elimination according to claim 3, wherein the mapping matrix learning algorithm of vector embedding into the binary hash space is as follows:

inputting: learning set

L _N In order to learn the number of vectors contained in the set,

let d be the original characteristic latitude, d _b The dimension of the binary characteristic after the embedded is obtained;

step a: randomGenerating a Gaussian matrix, performing QR orthogonal decomposition, and extracting the front d of the orthogonal matrix Q _b The row vector construction dimension is d _b x _d An orthogonal projection matrix P;

step b: for each vector x in the learning set omega _i Projecting by using the matrix P to obtain a vector set

Step c: z obtained by step b _ih The median projection vector Tl, Tlh ═ mean { Z } is obtained for the first cluster of calculations _ih },h∈[1,d _b ]；

Step d: quantizing x with a coarse quantizer q, and projecting x to z using P;

step e: mapping x into a binary vector using a fine quantizer;

and (3) outputting: projection matrix P, binary vector.

Images