CN112464005B - Depth-enhanced image clustering method - Google Patents

Depth-enhanced image clustering method Download PDF

Info

Publication number
CN112464005B
CN112464005B CN202011343296.9A CN202011343296A CN112464005B CN 112464005 B CN112464005 B CN 112464005B CN 202011343296 A CN202011343296 A CN 202011343296A CN 112464005 B CN112464005 B CN 112464005B
Authority
CN
China
Prior art keywords
clustering
prototype
input
bernoulli
logistic
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202011343296.9A
Other languages
Chinese (zh)
Other versions
CN112464005A (en
Inventor
陈志奎
金珊
高静
李朋
张佳宁
宋鑫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Dalian University of Technology
Original Assignee
Dalian University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Dalian University of Technology filed Critical Dalian University of Technology
Priority to CN202011343296.9A priority Critical patent/CN112464005B/en
Publication of CN112464005A publication Critical patent/CN112464005A/en
Application granted granted Critical
Publication of CN112464005B publication Critical patent/CN112464005B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/55—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a depth-enhanced image clustering method, which belongs to the technical field of image clustering and data mining, and comprises the following steps of 1) pre-training an encoding and decoding network and initializing a potential feature space; 2) initializing clustering centroids in a potential feature space by adopting a traditional K-means method, and distributing Bernoulli-logistic units to the centroids; 3) calculating a logistic regression parameter and Bernoulli distribution between the point and the unit; 4) a reward regression strategy is used for dynamically distributing temporary rewards, and the movement locus of each mass center is calculated by combining with the distribution of auxiliary targets; 5) and calculating weight, and iteratively optimizing the clustering unit until a convergence condition is met, thereby completing the depth enhanced image clustering process. The method is based on the reinforcement learning idea, utilizes the potential characteristics to express and adjust the clustering center of mass in a combined manner by using the reward regression strategy, fully applies all clustering information, particularly clustering information of adjacent areas to the clustering analysis process, effectively improves the clustering fuzzy problem in the interaction of environment and behavior, and effectively improves the clustering performance.

Description

Depth-enhanced image clustering method
Technical Field
The invention belongs to the technical field of image clustering and reinforcement learning, and relates to a depth image clustering method based on reinforcement learning.
Background
With the rapid development of the internet of things technology and the network information technology, the popularization range of electronic products such as smart phones and tablet computers is wider, more and more data can be collected, the data structure is more and more complex, and especially the amount of unstructured image data is increased explosively. The image data contains rich semantic information for research and use in various fields, but the rich semantic information in the data is difficult to accurately acquire due to the influence of complex data structure and high dimensionality. Therefore, a new method is urgently needed to be researched to deeply mine abundant information in massive image data.
The clustering usually performs data analysis in an unsupervised learning or self-supervised learning mode, is one of important research contents in the fields of data mining, image processing and the like, and effectively solves a plurality of problems of data mining by the idea of 'same type and different types'. Conventional clustering algorithms typically take a given data feature representation as input and then cluster the feature representation using different models, but for some high-dimensional, complex image data, it may be difficult to find its intrinsic pattern structure, i.e., "dimensional disaster," due to a lack of a sample similarity measure that is effective in a high-dimensional space. For high-dimensional data clustering, the conventional corresponding method comprises subspace clustering, feature dimension reduction, feature extraction and the like. In recent years, deep learning brings a new solution for high-dimensional and complex data clustering due to the unique advantages of the deep learning on feature representation, and a plurality of effective deep clustering algorithms are derived. The more representative deep clustering method, namely a Deep Embedded Clustering (DEC) algorithm, optimizes parameters and clustering distribution of the deep neural network simultaneously, and provides a powerful tool for clustering research.
At present, in the process of carrying out cluster analysis on large-scale data, the influence of potential feature representation on a clustering effect is fully considered by a deep clustering algorithm, so that the problem of dimension disaster is effectively solved. However, the existing deep clustering algorithm lacks consideration on the influence of the whole clustering environment, especially the adjacent area environment on the clustering effect, the similarity of the same kind is enhanced more and more along with the increase of the iteration times, and the heterogeneous difference is not obvious, so that the clustering of the partial image input points is fuzzy. In the process of interaction with the environment, reinforcement learning learns the strategy by acquiring rewards and guiding behaviors to maximize return, so that the accuracy of a clustering result is improved by guiding the walking direction of a prototype by adopting a reinforcement learning idea on the premise of not neglecting potential feature representation on the basis of deep clustering.
In summary, the invention provides a depth-enhanced image clustering method, which mainly considers the use problem of all clustering information in the iterative process of clustering analysis, and utilizes the clustering information after each iteration to adjust a clustering prototype and guide the clustering process.
Disclosure of Invention
The invention provides a depth-enhanced image clustering method. Firstly, in order to solve the dimensionality disaster, the method selects a deep self-encoder to perform dimensionality reduction on original image data, and obtains deep semantic features contained in the data. Secondly, the invention designs a Bernoulli-logistic unit to represent the clustering prototype, and effectively utilizes various available information to adjust the influence of the adjacent area on the clustering prototype. Finally, the invention adopts a reward strategy of reinforcement learning, distributes rewards for each cluster after the available information of each cluster is obtained, guides the behavior action of the clustering prototype in the clustering environment and obtains more accurate clustering results. In summary, the invention provides a depth-enhanced image clustering method, which adopts a learning mode of reward regression to learn the potential features of images from large-scale unlabeled data and perform clustering division so as to improve the Accuracy (ACC) of the clustering method, adjust the lander index (ARI) and standardize mutual information (NMI).
In order to achieve the purpose, the technical scheme adopted by the depth-enhanced image clustering method comprises the following steps:
step 1, pre-training an encoding and decoding network, and learning potential features of an image;
step 2, mining a clustering prototype in the potential feature space by adopting a K-means method, and distributing a Bernoulli-logistic unit for the clustering prototype;
step 3, randomly selecting a sample xiCalculating the logic between the point and the clustering prototypeRegression probability and bernoulli distribution parameters;
step 4, dynamically distributing temporary rewards by using a reward regression strategy, and calculating motion tracks of various prototypes by combining Bernoulli distribution;
step 5, calculating weight, iteratively optimizing clustering prototypes until convergence conditions are met, and completing a deep enhanced clustering process;
the invention has the beneficial effects that: the invention designs a deep enhanced clustering method aiming at image data, considers the interaction problem of the clustering environment and the prototype wandering direction in the clustering process, designs a Bernoulli-logistic unit and dynamically updates clustering information. Meanwhile, the method adjusts the clustering prototype by using a reward regression strategy based on the latent characteristics of the image and by using the idea of reinforcement learning, fully applies all clustering information, especially the clustering information of the adjacent areas to image clustering analysis, and effectively solves the problem of clustering blur. The method disclosed by the invention is used for carrying out experiments on evaluation indexes ACC, ARI and NMI commonly used by the clustering method, and proves that the method can effectively improve the clustering accuracy.
Drawings
FIG. 1 is a frame diagram of a depth-enhanced image clustering method;
FIG. 2 is a flow chart of a method of the present invention.
Detailed Description
The following further describes embodiments of the present invention with reference to the drawings.
FIG. 1 is a frame diagram of a deep enhanced clustering method. Firstly, a latent feature representation of data is extracted by a deep self-encoder, and high-dimensional original image data is mapped to a low-dimensional feature space, so that the problem of dimension disaster of the high-dimensional data is solved. Secondly, mining the clustering mass center of the data by using a K-means method, initializing clustering prototypes, distributing Bernoulli-logistic units to the clustering prototypes, and storing clustering environment information in the iteration process. Then, the similarity between the data points and the clustering prototypes in the characteristic space is measured by using the Euclidean distance, and the logistic regression parameters of the clustering and the Bernoulli distribution with high confidence level are updated. Secondly, dynamically rewarding and punishing each prototype by using a reward regression strategy, updating the motion track of the clustering prototype by combining Bernoulli distribution, ensuring the use of all clustering environments, particularly environment information of adjacent areas, and finishing the interaction of the current clustering environment and the input point behaviors. Finally, the process is repeated using a reinforcement learning algorithm by jointly using bernoulli distribution and reward penalty until the convergence condition is satisfied.
The method comprises the following specific steps:
step 1, pre-training an encoding and decoding network, and learning potential features of an image;
the original image data can provide more abundant and detailed information due to the characteristic of higher dimensionality, but at the same time, the understandability and usability of the data are greatly reduced due to the improvement of the data dimensionality. In order to solve the problem of the disaster of the dimension of the image data, the invention adopts a deep self-encoder model, minimizes the reconstruction error in an unsupervised mode for training, extracts the high-order features of the input data layer by layer in the process, reduces the dimension of the input data, and converts the complex input image data into a simple low-dimensional feature space.
The deep layer self-encoder network is formed by stacking a noise reduction self-encoder network, the noise reduction self-encoder network is formed by an encoding layer and a decoding layer, in the training process, the network randomly destroys the input of each noise reduction self-encoder, and then reconstructs the original input as the output to obtain the potential representation of the input data. The network can be defined as the following process:
Figure GDA0003224831300000031
Figure GDA0003224831300000032
Figure GDA0003224831300000033
Figure GDA0003224831300000034
setting image data x as the input of a noise reduction self-encoder; dropout () sets the partial input to 0 as a random mapping function; f. ofe、fdAs a mapping function of the coding layer and the decoding layer, θ ═ We,be,Wd,bdAre parameters of the network model; the potential features h output by the encoding layer will be the input to the decoding layer. Meanwhile, in order to ensure that the reconstructed image data x' is consistent with the original image data x as much as possible, a minimum square loss function is adopted
Figure GDA0003224831300000035
The method of (3) optimizes the model.
Specifically, in order to make the reconstructed feature vector contain all the information of the original feature vector, the activation functions of the coding layer and the decoding layer of the first noise reduction self-encoder are set as identity functions, and the activation functions of the coding layer and the decoding layer of the other noise reduction self-encoders are set as ReLu functions. After the construction of each noise reduction self-encoder network is completed, parameters of an encoding layer and a decoding layer are initialized by random distribution, then training is carried out by adopting a random gradient descent (SGD) back propagation mode, then the encoding layer and the decoding layer of each noise reduction self-encoder are disassembled, and an integral model frame of a deep self-encoder is combined according to input and output dimensions. Next, the SGD is adopted again to train the network model, so that the reconstruction loss is reduced to the maximum extent, a good deep self-encoder network model is generated, and the encoding part is selected as the original feature space to the potential feature space (f)θX → H), where θ is a parameter of the input point, X is an original feature representation of the input point, and H is a potential feature representation of the input point.
In the process of pre-training the deep self-encoder, the iteration number is set to 300, the number of samples in each training is set to 256, and experiments prove that the above-mentioned super-parameters are set, so that more effective potential feature representation H of the original image data can be obtained.
Step 2, initializing a clustering prototype, and distributing Bernoulli-logistic units;
generating an input point x by using a trained deep self-coder modeliIs represented by the potential feature of (a)iThe composition set H ═ Hi|hi＝fθ(xi),xiE.g. X, i is 1,2
Figure GDA0003224831300000041
And then, aggregating the prototypes on the potential feature representation H through a K-means clustering algorithm
Figure GDA0003224831300000042
And updating to obtain K initialized clustering prototypes.
Specifically, the purpose of the K-means clustering algorithm is to find K clustering prototypes by optimizing the following objective functions:
Figure GDA0003224831300000043
wherein d (c)k,hi) Is the input point hiAnd clustering prototype ckThe distance between the two adjacent channels is Euclidean distance, and the calculation process is as follows:
Figure GDA0003224831300000044
wherein n is the input point hiAnd clustering prototype ckOf (c) is calculated.
In order to effectively solve the process, a proper K value is selected according to the priori knowledge, and then a heuristic iteration method is adopted to select a proper clustering prototype. Specifically, K samples are randomly selected from a sample set as initial prototypes, the distance between each sample and each prototype is calculated according to the formula (6), and each sample is allocated to the nearest clustering prototype to obtain an initial clustering division result; then updating the prototype to obtain a new clustering division result; repeating the above process until the prototype is no longer usedThe change occurs, and the final clustering result is obtained as the clustering prototype set of the potential feature representation
Figure GDA0003224831300000045
Clustering prototypes obtained by using K-means
Figure GDA0003224831300000046
And constructing a Bernoulli-logistic unit BLlist ═ w, p, dw, fx } which contains clustering information of the clustering prototype under the current environment as the clustering prototype, wherein w is the weight of the clustering prototype under the current environment, p is the Bernoulli distribution coefficient of the clustering prototype, dw is the Euclidean distance of the prototype, and fx is the logistic regression coefficient of the prototype. The initial weight of the clustering prototype can be obtained through a K-means clustering algorithm, and the initial values of the other parameters are set to be 0 values.
Step 3, strengthening clustering;
after the latent feature extraction and cluster prototype initialization are completed, the main part of the invention, namely the initial non-linear mapping f to the original image data, will be performedθAnd clustering prototypes in Bernoulli-logistic units
Figure GDA0003224831300000047
The strengthening process of (1). The strengthening process mainly comprises the following two steps. And 3-1, calculating the potential feature representation of the input points, the logistic regression parameters of the clustering prototypes and the Bernoulli distribution parameters, and dynamically updating the parameters into the Bernoulli-logistic unit. And 3-2, rewarding or punishing each clustering prototype in the current clustering environment by adopting a reward regression strategy, and learning clustering loss by combining parameters in a Bernoulli-logistic stetty unit. Obtaining the probability p corresponding to the clustering prototype closest to the potential feature of the input image by using the step 3-1 and the step 3-2kIndicating variable yiPrize value riAnd determining the motion trail of the prototype.
1) Bernoulli-logistic distribution
According to the method, the Bernoulli-logistic distribution is selected as the auxiliary target distribution to be used for the activation unit, namely, the similarity between the potential features of the input image and the clustering prototype is measured through the Bernoulli-logistic distribution, the clustering accuracy is improved, and the confidence coefficient of clustering distribution is improved.
Firstly, randomly selecting potential features h of an input image to interact with each Bernoulli-logistic unit, and calculating the Euclidean distance s between the potential features h of the input image and each Bernoulli-logistic unitk＝d(h,wk) Then, a logistic function is used to measure the similarity between the latent features h of the input image and the prototype to which each cell belongs, and the function is as follows:
Figure GDA0003224831300000051
after obtaining the probability distribution of the current point, estimating by using an auxiliary cost function, wherein the calculation formula is as follows:
pk＝h(sk)＝2×(1-F(sk)) (8)
wherein p iskBelonging to a cluster prototype c for the potential feature of the input imagekThe probability of the cluster is, when the potential feature of the input image is closer to a unit, the probability p of the corresponding unitkThe larger the size, the smaller the size otherwise.
Because the influence of each clustering unit on the potential characteristics of the input image is not uniform due to the uncertainty of probability distribution under Bernoulli-logistic distribution, a formula (9) is designed to generate random seeds p, the random seeds p are compared with the probability obtained in the logical regression, an indication variable y is obtained, and the influence effect of each clustering prototype on the whole clustering result is balanced. The calculation formula is as follows:
Figure GDA0003224831300000052
the Bernoulli-logistic distribution obtained above is dynamically updated to the clustered Bernoulli-logistic unit along with an iterative process, and is used for iteratively updating the weight information of the clustering prototype to which the unit belongs.
2) Reward regression strategy
In order to fully utilize the influence of the clustering environment on the clustering result and to prominently consider the positive or negative effect of each front clustering on the back clustering, the invention adopts the idea of reinforcement learning, and selects a proper reward strategy to further clarify the learning direction for the set clustering prototype so as to seek the optimal result. For this, each clustering prototype c is calculatedkOf (a) metric value ykAnd then, allocating an evaluation decision for each prototype by using an incentive regression strategy to dynamically update the behavior generated after interaction between the input point and each clustering prototype, rewarding the effective clustering prototype and performing punishment operation on the adjacent invalid region, thereby solving the problem that the consideration of all clusters, particularly the adjacent regions, in a clustering algorithm is insufficient. The specific strategy scheme is as follows:
Figure GDA0003224831300000061
wherein the prototype is a closer prototype, i.e. a more compact prototype
Figure GDA0003224831300000062
When the current clustering prototype needs to be matched with a more active action scheme, namely a reward signal is sent to the current clustering prototype, the current clustering prototype needs to be matched with a more active action scheme
Figure GDA0003224831300000063
Conversely, when the prototype is a distant prototype, i.e. the prototype is a distant prototype
Figure GDA0003224831300000064
When the prototype is described as an error prototype, a penalty signal is sent to the prototype, i.e. the prototype is interpreted as an error prototype
Figure GDA0003224831300000065
In doing so, the weight of the clustering prototype is not affected.
Step 4, updating the weight and optimizing a clustering prototype;
after the strengthening task of the clustering prototype is completed, the method adopts a strategy gradient algorithm to update the weight parameter of the prototype k corresponding to the input point x, and the initial updating formula is as follows:
Figure GDA0003224831300000066
wherein a > 0 is the learning rate, r is the enhancement signal obtained during the enhancement process, bikA reinforcement baseline; while
Figure GDA0003224831300000067
Is the unit weight wikDegree of transformation of corresponding features, the value being subject to probability density function g under continuous distributionik(yk；wik,hi) Is influenced by variations in, i.e. by, the latent features h of the input imageiAnd a weight wikThe determined current prototype indicates a variable y in the current environmentkThe influence of (a);
according to the result of the enhanced clustering task, combining the indication variable ykWith allocation results of reward policies, and allocation of reinforcement baselines bikThe available final weight update formula is as follows:
Δwik＝ark(yk-pk)(-fx/(1-pk))(wik-hi) (12)
clustering prototype
Figure GDA0003224831300000068
The updating is performed by equation (12). And when the iteration times reach the preset maximum training times 60000 times, the whole clustering task is completed.
The method comprises the following steps:
the whole process of the invention is divided into three parts: a characteristic preprocessing process, a clustering prototype initialization process and a reinforced clustering process. Firstly, a deep self-encoder model is built, a decoder and an encoder of a noise reduction self-encoder are adopted to pre-train a network, and then the built encoding layer is utilized to map original high-dimensional image data to a low-dimensional potential feature space, so as to obtain potential feature representation of an image. Secondly, based on potential feature representation extracted from data in the feature preprocessing process, a traditional K-means algorithm is adopted to initialize clustering prototypes, and the clustering prototypes are stored in a Bernoulli-logistic unit mode. And finally, acquiring reward signals of each unit by adopting a reward regression strategy in reinforcement learning, and dynamically optimizing a clustering result by combining Bernoulli-logistic distribution until a clustering completion condition is met, wherein the specific flow is shown in figure 2.
And (4) verification result:
in the experiments of the present method, two general image data sets were selected: the MNIST writes the digital data set and the fast-MNIST data set to verify the effectiveness of the method, and the detailed information of the data sets is shown in Table 1.
MNIST handwritten digit data set: consisting of 70000 handwritten digits of 28 x 28 pixel size. The present invention reconstructs each image into a 784-dimensional vector.
The fast-MNIST dataset: consisting of 70000 apparel images of 28 x 28 pixel size. The present invention reconstructs each image into a 784-dimensional vector.
Table 1 details of the data set
Data set Number of samples Sample dimension Number of categories
MNIST 70000 784 10
Fashion-MNIST 70000 784 10
The method uses the traditional clustering evaluation criteria: clustering Accuracy (ACC), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI).
To verify the performance of the invention, 3 general typical clustering methods were chosen: the traditional unsupervised clustering method K means (K-means), the deep clustering method (AE + K-means) and the deep embedded clustering method (DEC) are compared.
The results of comparing the performance of ACC, ARI and NMI on MNIST and Fashion-MNIST data sets by the method proposed by the present invention are shown in tables 3 and 4.
Table 2 comparison of results on MNIST dataset for each experiment
Experiments ACC ARI NMI
K-means 0.5319 0.3633 0.4971
AE+K-means 0.8184 0.7421 0.7790
DEC 0.8430 0.8181 0.8437
The invention 0.9292 0.8493 0.8438
TABLE 3 comparison of results on the Fashinon-MNIST dataset for each experiment
Experiments ACC ARI NMI
K-means 0.4758 0.3485 0.5122
AE+K-means 0.5713 0.4259 0.5764
DEC 0.5829 0.4823 0.6404
The invention 0.6166 0.4871 0.6002
From tables 2 and 3, it can be observed that the method provided by the invention is superior to the comparative baseline method in both the evaluation indexes ACC and ARI of the data sets of MNIST and fast-MNIST, which proves the effectiveness of the invention. In particular, compared with the K-means method, the method has the advantages that the potential features of the image data can be extracted through the deep self-encoder network, and the clustering effect is improved. Compared with the AE + K-means method, the method has the advantages that the Bernoulli-logistic unit is adopted, the clustering information in the clustering unit is used for adjusting the cluster center, and the clustering performance is improved. Compared with the DEC method, the method has the advantages that the reward regression strategy is adopted, the valid clustering units are rewarded, meanwhile, invalid adjacent clustering units are punished, the influence of all clusters, particularly adjacent areas, on the clustering effect is fully considered, and the clustering performance is improved.

Claims (2)

1. A depth-enhanced image clustering method is characterized by comprising the following steps:
step 1, pre-training an encoding and decoding network, and learning potential features of an image;
the deep layer self-encoder network is formed by stacking a noise reduction self-encoder network, the noise reduction self-encoder network consists of an encoding layer and a decoding layer, in the training process, the network randomly destroys the input of each noise reduction self-encoder, and then reconstructs the original input as the output to obtain the potential representation of the input data; the denoising self-encoder network is defined as the following process:
Figure FDA0003224831290000011
Figure FDA0003224831290000012
Figure FDA0003224831290000013
Figure FDA0003224831290000014
setting image data x as the input of a noise reduction self-encoder; dropout () sets the partial input to 0 as a random mapping function; f. ofe、fdAs a mapping function of the coding layer and the decoding layer, θ ═ We,be,Wd,bdAre parameters of the network model; the potential characteristics h output by the coding layer are used as the input of the decoding layer; meanwhile, in order to ensure that the reconstructed image data x' is consistent with the original image data x as much as possible, a minimum square loss function is adopted
Figure FDA0003224831290000015
The method of (3) optimizes the model;
setting the activation functions of the coding layer and the decoding layer of the first noise reduction self-encoder as identity functions, and setting the activation functions of the coding layer and the decoding layer of the other noise reduction self-encoders as ReLu functions; after the network of each noise reduction self-encoder is built, parameters of a coding layer and a decoding layer are initialized by random distribution, then training is carried out by adopting a random gradient descent back propagation mode, then the coding layer and the decoding layer of each noise reduction self-encoder are disassembled, and an integral model frame of a deep self-encoder is combined according to input and output dimensions;
and training the network model by adopting the SGD again to reduce the reconstruction loss to the maximum extent, generating a good deep self-encoder network model, and selecting a coding part as a mapping tool f from an original characteristic space to a potential characteristic spaceθX → H, where θ is a parameter of the input point, X is an original feature representation of the input point, and H is a potential feature representation of the input point, and finally a potential feature representation H of the original image data is obtained;
step 2, initializing a clustering prototype, and distributing Bernoulli-logistic units;
generating an input point x by adopting the deep self-coder model trained in the step 1iIs represented by the potential feature of (a)iThe composition set H ═ Hi|hi＝fθ(xi),xiE.g. X, i is 1,2
Figure FDA0003224831290000016
And then, aggregating the prototypes on the potential feature representation H through a K-means clustering algorithm
Figure FDA0003224831290000017
Updating to obtain K initialized clustering prototypes;
clustering prototypes obtained by using K-means
Figure FDA0003224831290000018
As a clustering prototype, constructing a bernoulli-logistic unit BLlist ═ { w, p, dw, fx }, which contains clustering information of the clustering prototype in the current environment, wherein w is the weight of the clustering prototype in the current environment, p is the bernoulli distribution coefficient of the clustering prototype, dw is the euclidean distance of the prototype, fx is the prototypeA logistic regression coefficient; obtaining the initial weight of a clustering prototype through a K-means clustering algorithm, and setting the initial values of the other parameters as 0 values;
step 3, strengthening clustering;
step 3-1, calculating the potential feature representation of the input points, the logistic regression parameters of the clustering prototypes and the Bernoulli distribution parameters, and dynamically updating the parameters into a Bernoulli-logistic unit;
firstly, randomly selecting potential features h of an input image to interact with each Bernoulli-logistic unit, and calculating the Euclidean distance s between the potential features h of the input image and each Bernoulli-logistic unitk＝d(h,wk) Then, a logistic function is used to measure the similarity between the latent features h of the input image and the prototype to which each cell belongs, and the function is as follows:
Figure FDA0003224831290000021
after obtaining the probability distribution of the current point, estimating by using an auxiliary cost function, wherein the calculation formula is as follows:
pk＝h(sk)＝2×(1-F(sk)) (8)
wherein p iskBelonging to a cluster prototype c for the potential feature of the input imagekThe probability of the cluster is, when the potential feature of the input image is closer to a unit, the probability p of the corresponding unitkThe larger the size, the smaller the size otherwise;
meanwhile, a random seed p is generated by the formula (9), and is compared with the probability obtained in the logical regression to obtain an indication variable y, so that the influence effect of each clustering prototype on the whole clustering result is balanced; the calculation formula is as follows:
Figure FDA0003224831290000022
the Bernoulli-logistic distribution obtained in the above way is dynamically updated to a Bernoulli-logistic unit of the clustering prototype along with the iterative process, and is used for iteratively updating the weight information of the clustering prototype to which the unit belongs;
step 3-2, reward or punishment is carried out on each clustering prototype under the current clustering environment by adopting a reward regression strategy, and clustering loss is learned by combining parameters in a Bernoulli-logistic stetty unit;
in calculating each clustering prototype ckOf (a) metric value ykThen, an evaluation decision is distributed to each prototype by using an incentive regression strategy to dynamically update behaviors generated after interaction between the input points and each clustering prototype, and punishment operation is carried out on adjacent invalid regions while an effective clustering prototype is rewarded, so that the problem that the consideration of all clusters, particularly the adjacent regions, in a clustering algorithm is insufficient is solved; the specific strategy scheme is as follows:
Figure FDA0003224831290000023
wherein the prototype is a closer prototype, i.e. a more compact prototype
Figure FDA0003224831290000024
When the current clustering prototype needs to be matched with a more active action scheme, namely a reward signal is sent to the current clustering prototype, the current clustering prototype needs to be matched with a more active action scheme
Figure FDA0003224831290000037
Conversely, when the prototype is a distant prototype, i.e. the prototype is a distant prototype
Figure FDA0003224831290000038
When the prototype is described as an error prototype, a penalty signal is sent to the prototype, i.e. the prototype is interpreted as an error prototype
Figure FDA0003224831290000039
In the process of carrying out the strategy, the weight of the clustering prototype is not influenced;
determining a prototype motion track by using the step 3-1 and the step 3-2;
step 4, updating the weight and optimizing a clustering prototype;
updating the weight parameters of the prototype k corresponding to the input point x by adopting a strategy gradient algorithm, wherein an initial updating formula is as follows:
Figure FDA0003224831290000031
wherein a > 0 is the learning rate, r is the enhancement signal obtained during the enhancement process, bikA reinforcement baseline; while
Figure FDA0003224831290000032
Is the unit weight wikDegree of transformation of corresponding features, the value being subject to probability density function g under continuous distributionik(yk；wik,hi) Is influenced by variations in, i.e. by, the latent features h of the input imageiAnd a weight wikThe determined current prototype indicates a variable y in the current environmentkThe influence of (a);
according to the result of the enhanced clustering task, combining the indication variable ykWith allocation results of reward policies, and allocation of reinforcement baselines bikWhen it is 0, the final weight update formula is as follows:
Δwik＝ark(yk-pk)(-fx/(1-pk))(wik-hi) (12)
clustering prototype
Figure FDA0003224831290000033
Updating by formula (12); and when the iteration times reach the preset maximum training times, finishing the whole clustering task.
2. The method for depth-enhanced image clustering according to claim 1, wherein the K-means clustering algorithm in step 1 aims to find K clustering prototypes by optimizing the following objective functions:
Figure FDA0003224831290000034
wherein d (c)k,hi) Is the input point hiAnd clustering prototype ckThe Euclidean distance is adopted as the distance between the two, and the calculation process is as follows:
Figure FDA0003224831290000035
wherein n is the input point hiAnd clustering prototype ckDimension (d);
firstly, randomly selecting K samples from a sample set as initial prototypes, calculating the distance between the samples and each prototype according to the formula (6), and distributing each sample to the nearest clustering prototype to obtain an initial clustering division result; then updating the prototype to obtain a new clustering division result; repeating the above process until the prototype is not changed any more, and obtaining the final clustering result as the clustering prototype set of the potential feature representation
Figure FDA0003224831290000036
CN202011343296.9A 2020-11-26 2020-11-26 Depth-enhanced image clustering method Active CN112464005B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011343296.9A CN112464005B (en) 2020-11-26 2020-11-26 Depth-enhanced image clustering method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011343296.9A CN112464005B (en) 2020-11-26 2020-11-26 Depth-enhanced image clustering method

Publications (2)

Publication Number Publication Date
CN112464005A CN112464005A (en) 2021-03-09
CN112464005B true CN112464005B (en) 2021-12-03

Family

ID=74808392

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011343296.9A Active CN112464005B (en) 2020-11-26 2020-11-26 Depth-enhanced image clustering method

Country Status (1)

Country Link
CN (1) CN112464005B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TWI795787B (en) * 2021-05-24 2023-03-11 鴻海精密工業股份有限公司 Method for training autoencoder, method for detecting a number of cells and related equipment
CN113435546B (en) * 2021-08-26 2021-12-24 山东力聚机器人科技股份有限公司 Migratable image recognition method and system based on differentiation confidence level
TWI796837B (en) * 2021-11-17 2023-03-21 宏碁股份有限公司 Noise reduction convolution auto-encoding device and noise reduction convolution self-encoding method
CN114462548B (en) * 2022-02-23 2023-07-18 曲阜师范大学 Method for improving accuracy of single-cell deep clustering algorithm
CN116682128A (en) * 2023-06-02 2023-09-01 中央民族大学 Method, device, equipment and medium for constructing and identifying data set of water book single word
CN117528085B (en) * 2024-01-08 2024-03-19 中国矿业大学 Video compression coding method based on intelligent feature clustering
CN117934890B (en) * 2024-03-21 2024-06-11 南京信息工程大学 Prototype comparison image clustering method and system based on local and global neighbor alignment

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016359A (en) * 2017-03-24 2017-08-04 同济大学 A kind of fast face recognition method being distributed under complex environment based on t
CN109919223A (en) * 2019-03-05 2019-06-21 北京联合大学 Object detection method and device based on deep neural network
CN111259979A (en) * 2020-02-10 2020-06-09 大连理工大学 Deep semi-supervised image clustering method based on label self-adaptive strategy
CN111291181A (en) * 2018-12-10 2020-06-16 百度（美国）有限责任公司 Representation learning for input classification via topic sparse autoencoder and entity embedding

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20180077691A (en) * 2016-12-29 2018-07-09 주식회사 엔씨소프트 Apparatus and method for sentence abstraction

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107016359A (en) * 2017-03-24 2017-08-04 同济大学 A kind of fast face recognition method being distributed under complex environment based on t
CN111291181A (en) * 2018-12-10 2020-06-16 百度（美国）有限责任公司 Representation learning for input classification via topic sparse autoencoder and entity embedding
CN109919223A (en) * 2019-03-05 2019-06-21 北京联合大学 Object detection method and device based on deep neural network
CN111259979A (en) * 2020-02-10 2020-06-09 大连理工大学 Deep semi-supervised image clustering method based on label self-adaptive strategy

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
A Survey on Deep Learning for Multimodal Data Fusion;Jing Gao等;《IEEE》;20200415;第829-864页 *
Deep clustering: On the link between discriminative models and K-means;Mohammed Jabi等;《 https://arxiv.org/pdf/1810.04246.pdf》;20191231;第1-11页 *
大规模机器学习理论研究与应用;张利军;《万方数据》;20131008;第1-144页 *
深度卷积自编码图像聚类算法;谢娟英等;《计算机科学与探索》;20180629;第586-594页 *
面向大数据特征学习的深度卷积计算模型研究;李朋;《中国博士学位论文全文数据库 信息科技辑》;20200815;第I138-6页 *

Also Published As

Publication number Publication date
CN112464005A (en) 2021-03-09

Similar Documents

Publication Publication Date Title
CN112464005B (en) Depth-enhanced image clustering method
Jia et al. Subspace clustering of categorical and numerical data with an unknown number of clusters
Al-Jabery et al. Computational learning approaches to data analytics in biomedical applications
Zhang et al. Clustering-guided particle swarm feature selection algorithm for high-dimensional imbalanced data with missing values
Wu et al. Effective hierarchical clustering based on structural similarities in nearest neighbor graphs
Bi et al. A survey on evolutionary computation for computer vision and image analysis: Past, present, and future trends
CN112906770A (en) Cross-modal fusion-based deep clustering method and system
Zhang et al. Hybrid fuzzy clustering method based on FCM and enhanced logarithmical PSO (ELPSO)
CN108171012B (en) Gene classification method and device
CN101968853A (en) Improved immune algorithm based expression recognition method for optimizing support vector machine parameters
Tripoliti et al. Modifications of the construction and voting mechanisms of the random forests algorithm
CN112464004A (en) Multi-view depth generation image clustering method
Kianmehr et al. Fuzzy clustering-based discretization for gene expression classification
CN103593674A (en) Cervical lymph node ultrasonoscopy feature selection method
Ouadfel et al. A multi-objective gradient optimizer approach-based weighted multi-view clustering
CN113610139A (en) Multi-view-angle intensified image clustering method
CN116187835A (en) Data-driven-based method and system for estimating theoretical line loss interval of transformer area
CN109409434A (en) The method of liver diseases data classification Rule Extraction based on random forest
Huang et al. Self-supervised graph attention networks for deep weighted multi-view clustering
CN111583194A (en) High-dimensional feature selection algorithm based on Bayesian rough set and cuckoo algorithm
Huang et al. Deep embedded multi-view clustering via jointly learning latent representations and graphs
CN117059284A (en) Diabetes parallel attribute reduction method based on co-evolution discrete particle swarm optimization
CN111353525A (en) Modeling and missing value filling method for unbalanced incomplete data set
CN116842354A (en) Feature selection method based on quantum artificial jellyfish search mechanism
CN110837853A (en) Rapid classification model construction method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant