CN110489585B

CN110489585B - Distributed image searching method based on supervised learning

Info

Publication number: CN110489585B
Application number: CN201910609588.3A
Authority: CN
Inventors: 胡海峰; 熊键
Original assignee: Nanjing University of Posts and Telecommunications
Current assignee: Nanjing University of Posts and Telecommunications
Priority date: 2019-07-08
Filing date: 2019-07-08
Publication date: 2022-12-02
Anticipated expiration: 2039-07-08
Also published as: CN110489585A

Abstract

The invention discloses a distributed image searching method based on supervised learning, which comprises the steps of firstly carrying out classification marking on images, videos and files in a database of each node, initializing a classification matrix, an encoding matrix, a Hash code matrix and corresponding Lagrangian multipliers, then introducing a minimized classification error and a reconstructed error to construct an objective function, solving the objective function and updating a parameter matrix; the data nodes are communicated with the central node, whether the conversion matrixes of the nodes tend to be consistent or not is judged, the Lagrangian multiplier is updated, and finally the similarity searching process is carried out; the invention solves the problems that large-scale data is stored and the scale required during calculation is too large, and the algorithm model is not suitable for centralized training; and the data node and the central node do not exchange original information in communication, so that the problem of overlarge transmission communication can be effectively solved, and the data on the node keeps independence.

Description

Distributed image searching method based on supervised learning

Technical Field

The invention relates to an image searching method, in particular to a distributed image searching method, and belongs to the field of machine learning.

Background

With the continuous development of social networks, electronic commerce, mobile internet and the like, the scale of data required to be stored and processed becomes larger and larger, and the stand-alone system cannot meet the increasing requirements. Internet companies such as Google and Alibaba successfully bring forward two popular fields of cloud computing and big data, and the cloud computing and the big data are both applications constructed on distributed storage. The core of cloud storage is a large-scale distributed storage system at the back end, large data not only needs to store massive data, but also needs to be analyzed through a proper framework and tools to obtain a useful part, and if distributed storage is not available, analysis on the large data is not referred to. Although research on distributed systems has been conducted for many years, until recently the rise of internet big data has made distributed systems to be applied in engineering practice on a large scale. The distributed system utilizes a plurality of computers to cooperatively solve the problems of calculation and storage which cannot be solved by a single computer, and the biggest difference between the distributed system and a single computer lies in the scale of the problems. It is a system composed of a plurality of nodes, often called a node, a server or a process on a server, which are not isolated but communicate with each other via a network to transfer information. In addition, due to the rapid development of mobile terminals such as smart phones and the like, the smart phones store a large amount of information such as pictures, texts and videos, the smart phones can also be regarded as an independent node, and the data processing capacity is improved through the base station or through distributed cooperation among the smart phones.

Supervised learning (Supervised learning), a type of algorithm in machine learning, can learn or create a pattern from training data, and infer a new instance based on the pattern. The training data consists of input objects (usually vectors) and expected outputs. The output of the function may be a continuous value (called regression analysis), or a predictive classification label. There is also a large class of algorithms in machine learning, called Unsupervised learning (Unsupervised learning), which is to model and learn training data without labels directly, and it is noted that the data here is unmarked data, and the most basic difference from supervised learning is that the modeled data is labeled one and unlabeled one. Compared with unsupervised learning, the supervised learning has the advantages that the known marking information can be fully utilized, more information is fused into the constructed model, and the reliability of the model is effectively improved.

In addition, with the wide spread of the internet and the development of multimedia technology, data of various industries is rapidly increased, and modern information technology infrastructure has to deal with huge databases. In fact, retrieving relevant content in large-scale databases is a more challenging task compared to storage costs, especially in searching multimedia data, such as audio, image and video content. When the traditional nearest neighbor algorithm is used for processing a large-scale image retrieval problem, the characteristic dimension of sample data can reach thousands of dimensions, and the problem of large storage space consumption and low retrieval speed can be caused by dimension disaster. In recent years, the hash algorithm, as a representative nearest neighbor search technology, can meet special requirements for storage space and search time in large-scale search. The purpose of the hashing algorithm is to represent the image as a set of fixed-length binary codes, i.e., hash codes, in which bits are typically represented using-1/1 or 0/1. The Hash algorithm solves the unreasonable requirements of the traditional retrieval problem on large-scale data storage space and retrieval time, greatly reduces the requirements on the storage space and the retrieval time, and can obtain good retrieval effect, so that the Hash algorithm becomes a sharp for processing the large-data problem and is widely concerned in the field of computer vision. However, most of the hash algorithms are centralized at present, and a plurality of problems such as large calculation amount of a single node exist, so that how to apply the hash algorithm in a distributed scene is an interesting problem.

In summary, there is no disclosure in the prior art of how to implement distributed image search using a supervised hash algorithm.

Disclosure of Invention

The invention aims to provide a distributed image searching method based on supervised learning, which is mainly used for solving the problems that the number of samples such as images, videos, texts and the like is large, semantic neighbors cannot be accurately found, and if the samples are trained together, the transmission quantity and the calculated quantity are overlarge.

The invention provides a distributed image searching method based on supervised learning, which comprises the following steps:

step 1: classifying and marking images, videos and files in a database of each node;

step 2: initializing a classification matrix, an encoding matrix, a Hash code matrix and a corresponding Lagrange multiplier;

and step 3: introducing a minimum classification error and a reconstruction error to construct an objective function;

and 4, step 4: solving the objective function, updating a classification matrix, an encoding matrix and a Hash code matrix;

and 5: the data nodes are communicated with the central node, whether the conversion matrixes of the nodes tend to be consistent or not is judged, and the Lagrangian multiplier is updated;

and 6: a proximity search process is performed.

As a further limitation of the present invention, in step 1, it is assumed that there are N nodes, and each node corresponds to a database X _i ，X _i The databases of the ith node are independent from one another, and information is not expected to be shared among different nodes, each database has c types of category labels, and different samples are marked with different labels.

As a further limitation of the present invention, in step 2, initialization setting is performed on the classification matrix, the coding matrix, the hash code matrix and the corresponding lagrangian multiplier at each node, the initialization classification matrix is set to be a d × r dimensional unit matrix, the corresponding initialization lagrangian multiplier is a d × r dimensional all-0 matrix, the initialization classification matrix is an r × c dimensional unit matrix, the corresponding initialization lagrangian multiplier is an r × c dimensional all-0 matrix, the initialization hash code matrix is an r × n dimensional matrix with an absolute value of each element of 1, d represents a dimension of the original feature space of the sample, r represents the number of coding bits, c represents the number of classes, and n represents the number of samples.

As a further limitation of the present invention, in step 3, a minimum classification error and a minimum reconstruction error are introduced into the objective function, and the original feature space is mapped to the hash codes through the coding matrix, so that the classification accuracy based on the hash codes is as high as possible, to ensure the validity of the hash codes, and meanwhile, to reduce the correlation between the hash codes, an orthogonal constraint is added, and to reduce the quantization error, a discrete constraint is added, that is, the hash codes are forced to be equal to 1 or-1.

The constructed objective function is as follows in sequence:

in the above formula X _i Samples representing the ith node, i.e. database X _i ,C _i 、B _i Respectively representing the coding matrix and the Hash code matrix of the ith node, pi _i Representing dual variables, wherein rho is a Lagrange multiplier, Z is a global parameter introduced by consistency, the constraint in the formula is composed of two parts, the first part is the global consistency constraint of an Alternating Direction Multiplier Method (ADMM), and the second constraint is the mutually independent constraint of hash codes:

in the above formula Y _i Sample flag, W, representing the ith node _i 、B _i The classification matrix and the hash code matrix of the ith node are respectively represented, lambda is a Lagrange multiplier, U is a global parameter introduced by the Alternating Direction Multiplier Method (ADMM) consistency, and the constraint is global consistency constraint.

In the above formula Y _i Sample flag, W, representing the ith node _i 、B _i 、C _i And v is a balance parameter, and the added constraint is to ensure that the middle value of each bit of the hash code is discrete.

As a further limitation of the present invention, in step 4, when solving the coding matrix C, it is necessary to solve by using singular value decomposition since it involves solving a problem of minimizing the matrix trace under the condition of orthogonal constraint.

As a further limitation of the present invention, in step 5, when the coding matrix C and the classification matrix W are optimized in a distributed manner, in addition to the N data nodes, there is a central node for performing global update on W and C, and parameter information is transmitted between the central node and the data nodes, so as to ensure parameter consistency.

As a further limitation of the present invention, in step 6, for a new query sample, broadcast it to all nodes, after mapping the coding matrix, calculate hamming distances between the new sample and the node samples, and take the sample corresponding to the top k minimum distances as the result of the proximity search.

Compared with the prior art, the invention adopting the technical scheme has the following technical effects:

1. the method solves the problems that many traditional methods are too unilateral when used for neighbor searching, marking information is not considered when neighbor searching is carried out, discretization processing is not carried out in the intermediate process, and the performance of the algorithms in the practical application of approximate searching is poor;

2. the problem that the large-scale data is stored and the required scale is too large during calculation, the calculation capacity of a single calculation node is exceeded, and an algorithm model is not suitable any more in a centralized training mode is solved;

3. the parameter matrix is used for carrying out communication among the nodes, original information is not exchanged in the communication among the nodes, the problem of overlarge transmission communication can be effectively solved, and meanwhile, good performance is achieved.

Drawings

FIG. 1 is a system block diagram of the present method.

Fig. 2 is a flow chart of the distributed training of the present method.

Fig. 3 is a flowchart of neighbor search in the present method.

Detailed Description

The technical scheme of the invention is further explained in detail by combining the drawings as follows:

a system framework diagram of the method is shown in fig. 1, the whole method process can be divided into a distributed training process and an approximation search process, and specific flows are respectively shown in fig. 2 and fig. 3, wherein the first step to the fifth step adopt a mode shown in fig. 2, and the sixth step adopts a mode shown in fig. 3.

In the first step, images, videos, files, etc. are labeled in a database of each node by classification.

Assuming that there are N nodes, each node corresponds to a database X _i ，X _i Databases representing the ith node, the databases in different nodes being independent of each other and notThe same nodes do not want to share information, each database has n samples, each database has c kinds of category labels, and different labels are marked on different samples.

And secondly, initializing a classification matrix, an encoding matrix and a Lagrange multiplier matrix, and initializing a Hash code matrix.

Initializing a classification matrix, a coding matrix and a corresponding Lagrange multiplier in each node, initializing a Hash code matrix, and setting an initialization classification matrix C of the ith node _i Is an identity matrix with dimension d multiplied by r, the corresponding initialized Lagrange multiplier is a full 0 matrix with dimension d multiplied by r when the ith node optimizes C, and a classification matrix W is initialized _i The matrix is an r × c-dimensional unit matrix, the initialized Lagrange multiplier corresponding to the optimized W is an r × c-dimensional all-0 matrix, and the initialized Hash code matrix B is an r × n-dimensional matrix with the absolute value of each element being 1. d represents the dimension of the original feature space of the sample, r represents the number of encoding bits, c represents the number of classes, and n represents the number of samples. The conversion matrix and the initialization value of the lagrangian multiplier are the same for all nodes.

And thirdly, introducing a minimized classification error and a reconstruction error, and constructing an objective function by discretization and orthogonal constraint.

It should be noted that the key point of the present invention is to obtain the coding matrix and to utilize the coding matrix, so that the original objective function and the specific optimization process of the objective function are not listed in the key point, and only the result after the objective function is optimized is given; optimization C constructed by ith node _i The latter objective function is as follows:

in the above formula X _i Sample representing the ith node, C _i 、B _i Respectively representing the coding matrix and the Hash code matrix of the ith node, wherein rho is a Lagrange multiplier, pi _i Representing dual variables, Z is a global parameter introduced by the Alternating Direction Multiplier Method (ADMM) consistency, and the constraint in the formula is composed of two parts, wherein the first part is an alternating directionThe second constraint is the constraint that the hashes are independent of each other, which is the global consistency constraint of the multiplier method (ADMM).

Optimization W constructed by ith node _i The latter objective function is as follows:

Optimization B constructed by ith node _i The latter objective function is as follows:

in the above formula Y _i Sample flag, W, representing the ith node _i 、B _i 、C _i The classification matrix, the hash code matrix and the coding matrix of the ith node are respectively represented, v is a balance parameter, and the added constraint is to ensure that the intermediate value of each bit of the hash code is discrete.

And fourthly, solving the objective function and updating the parameter matrix.

Respectively solving the three objective functions to optimize W _i ,C _i All adopt an Alternative Direction Multiplier Method (ADMM) to solve and optimize B _i The data information of each node is directly utilized to solve, and the nodes are completely distributed during optimization.

And fifthly, the data nodes are communicated with the central node, whether the conversion matrixes of the nodes tend to be consistent or not is judged, and the Lagrangian multiplier is updated.

The data nodes and the global nodes are communicated, that is, each node transmits the parameter matrix calculated by the node to the central node, the central node performs global optimization, and then transmits the parameter matrix of the global optimization to each data node for next iteration updating, and the parameter transmission can ensure that the training process meets the idea of consistency.

And if the conversion matrixes of all the nodes do not tend to be consistent, iteratively updating the Lagrange multiplier, and repeating the third step.

And sixthly, performing an approximation search process.

For a new query sample x _c Inputting the code matrix into all distributed nodes, and assuming that the code matrix enters the ith node, obtaining the code matrix C by using a distributed training process _i For query sample x _c And data node samples are encoded, then x is calculated _c The hamming distance from them (i.e., the number of bits corresponding to the hash code); obtaining a query sample x _c And after the distances between the nodes and other samples in each node, sequencing the distances to obtain the first k minimum distances, wherein the samples corresponding to the first k minimum distances are the results obtained by the approximation search.

The above description is only an embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can understand that the modifications or substitutions within the technical scope of the present invention are included in the scope of the present invention, and therefore, the scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A distributed image searching method based on supervised learning is characterized by comprising the following steps:

and step 3: introducing a minimum classification error and a reconstruction error to construct a target function, introducing the minimum classification error and the reconstruction error in the target function, and mapping an original characteristic space to a hash code through a coding matrix, so that the classification accuracy based on the hash code is as high as possible to ensure the effectiveness of the hash code, wherein the constructed target function is as follows in sequence:

in the above formula X _i Samples representing the ith node, i.e. database X mentioned earlier _i ，C _i 、B _i Respectively representing the coding matrix and the Hash code matrix of the ith node, pi _i Representing dual variables, wherein rho is a Lagrange multiplier, Z is a global parameter introduced by consistency, the constraint in the formula is composed of two parts, the first part is global consistency constraint of an Alternating Direction Multiplier Method (ADMM), and the second constraint is constraint that hash codes are mutually independent:

in the above formula Y _i Sample flag, W, representing the ith node _i 、B _i Respectively representing a classification matrix and a Hash code matrix of the ith node, wherein lambda is a Lagrange multiplier, U is a global parameter introduced by an Alternative Direction Multiplier Method (ADMM) consistency, and the constraint is global consistency constraint;

in the above formula Y _i Sample flag, W, representing the ith node _i 、B _i 、C _i The classification matrix, the Hash code matrix and the coding matrix of the ith node are respectively represented, v is a balance parameter, and the added constraint is to ensure that the middle value of each bit of the Hash code is discrete;

and 4, step 4: solving the target function, updating a classification matrix, a coding matrix and a Hash code matrix, wherein when the coding matrix C is solved, singular value decomposition is needed for solving the problem of minimizing matrix traces under the condition of orthogonal constraint;

and 5: the data nodes are communicated with the central node, whether the conversion matrixes of the nodes tend to be consistent or not is judged, and when the Lagrange multiplier is updated and the distributed optimization coding matrix C and the classification matrix W are updated, besides the N data nodes, the central node is used for globally updating the W and the C and transmitting parameter information between the central node and the data nodes, so that the consistency of parameters is ensured;

step 6: and performing an approximation search process, broadcasting a new query sample to all nodes, mapping the new query sample by using a coding matrix, calculating Hamming distances between the new sample and the node samples, and taking the samples corresponding to the first k minimum distances as the result of the approximation search.

2. The supervised learning-based distributed image searching method of claim 1, wherein in step 1, assuming that there are N nodes, each node corresponds to one database X _i ，X _i The databases of the ith node are independent from one another, and information is not expected to be shared among different nodes, each database has c types of category labels, and different samples are marked with different labels.

3. The supervised learning-based distributed image searching method of claim 2, wherein in step 2, the classification matrix, the coding matrix, the hash code matrix and the corresponding lagrangian multiplier are initialized, the initialized classification matrix is a unit matrix of d × r dimension, the corresponding initialized lagrangian multiplier is a full 0 matrix of d × r dimension, the initialized classification matrix is a unit matrix of r × c dimension, the corresponding initialized lagrangian multiplier is a full 0 matrix of r × c dimension, the initialized hash code matrix is a matrix of r × n dimension with each element having an absolute value of 1, d represents a dimension of a sample original feature space, r represents a number of coding bits, c represents a number of categories, and n represents a number of samples.