Image incremental clustering method and device based on local probability density structure development neural network

Abstract

The application discloses an image incremental clustering method and device based on a local probability density structure development neural network, and the method steps comprise the following steps: constructing a local probability density structure development neural network LPD-SDNN, the network is a two-layer network structure, wherein the first layer is an input layer for acquiring input samples, and the second layer is a competitive layer, and information recorded by each neuron at least comprises a center weight vector and a covariance matrix; acquiring image data samples to be clustered and outputting the image data samples to the LPD-SDNN for clustering; in the clustering process, the competitive layer adjusts parameters of each neuron according to a preset competitive learning rule, so that each neuron continuously updates and adjusts the corresponding center weight vector and the covariance matrix according to the input samples, so as to fit the local probability density information of the input samples, and until a clustering result is output. The application has the advantages of simple implementation method, high clustering accuracy and clustering result credibility, strong flexibility and the like.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to an image incremental clustering method and apparatus based on a local probability density structure developmental neural network. Background Technology

[0002] When intelligent systems operate in open environments, they encounter information they haven't yet learned, necessitating continuous learning capabilities. Incremental image clustering aims to enable intelligent systems to continuously learn new category structures from dynamically changing image data streams while maintaining stable memory of previously learned knowledge. This allows for the continuous accumulation and updating of visual knowledge, finding wide application in open and dynamic scenarios such as video surveillance, autonomous driving, and social media image classification. However, current mainstream machine learning methods struggle to meet the demands of continuous learning. When encountering new classes, they often require inefficient retraining of the entire model or forget previously learned information while adapting to new information, leading to catastrophic forgetting.

[0003] Existing technologies typically employ continuous learning methods based on neural networks to mitigate catastrophic forgetting, such as regularization methods like LwF and EWC, which impose constraints on the updating of neural weights to reduce catastrophic forgetting. However, these methods are almost exclusively used in supervised or reinforcement learning scenarios. New data in open environments is usually unlabeled, and it's difficult to know the number of classes present in the environment a priori. Therefore, continuous unsupervised learning in open environments needs attention. In continuous unsupervised learning, it's challenging for the learning network to balance the plasticity required to adapt to rapidly changing environments with the stability of previously learned knowledge—a phenomenon known as the stability-plasticity dilemma.

[0004] To address the aforementioned issues, existing technologies have proposed several incremental neural networks based on Hebbian competitive learning. The network topology of these networks autonomously evolves with the input data stream to adapt to the current learning data; these are known as structure-developmental neural networks. Most structure-developmental neural networks are inspired by classic self-organizing map neural networks (SOM) and neural gas networks (NG), with the Growing Neural Gas (GNG) network being a representative example. GNG improves the dynamics of the network, but it still cannot track rapidly changing, non-stationary data. The GWR (GrowWhen Required) network has achieved good performance in novelty detection. Self-organizing incremental learning neural networks (SOINN) can improve the plasticity of neural networks, but they cannot provide stable clustering results.

[0005] Some researchers have further proposed a structural developmental neural network with information saturation (SDNNIS, ESDNNIS), capable of handling non-stationary data streams in open environments. It expands the network through neuron splitting with parameter adjustments and information saturation regulation, adapting to dynamic data and providing real-time clustering results for the current input samples, enabling continuous unsupervised learning tasks. SDNNIS mimics human long and short-term memory capabilities by utilizing information saturation and neuron expansion mechanisms. Furthermore, it can identify smaller clusters within larger clusters as the number of learning samples increases, achieving better clustering results than GNG, GWR, and SOINN in various continuous unsupervised learning scenarios with less runtime. However, the neurons in SDNNIS only store the central weight vector (first moment), typically employing a single hyperspherical covering domain. This fails to describe complex distributions with different orientations, scales, or densities (such as strip-shaped, ellipsoidal, or unbalanced distributions). Using a single spherical covering domain leads to inaccurate clustering results, low reliability of clustering results, and neuron redundancy. Summary of the Invention

[0006] The technical problem to be solved by this invention is: in view of the above-mentioned problems existing in the prior art, this invention provides an image incremental clustering method and device based on a local probability density structure developmental neural network, which is simple to implement, has high clustering accuracy and reliability of clustering results, and is highly flexible. It can be applied to various complex distributed datasets to achieve continuous unsupervised learning tasks.

[0007] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0008] An image incremental clustering method based on a local probability density structure developmental neural network includes the following steps:

[0009] A Local Probability Density Structure Developmental Neural Network (LPD-SDNN) is constructed. The LPD-SDNN is a two-layer network structure, wherein the first layer is an input layer for acquiring input samples, and the second layer is a competition layer composed of multiple neurons for learning the distribution characteristics of input samples and recording information by utilizing the competition mechanism between neurons. The information recorded by each neuron includes at least a center weight vector and a covariance matrix of local region data for modeling local data distribution.

[0010] The image data samples to be clustered are obtained and output to the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) for clustering.

[0011] During the clustering process, the competitive layer adjusts the parameters of each neuron according to the preset competitive learning rules, so that each neuron continuously updates and adjusts the corresponding center weight vector and covariance matrix according to the input samples to fit the local probability density information of the input samples until the clustering result is output.

[0012] Furthermore, each neuron in the competition layer is represented as a quintuple. :< , , , , >, among which Indicates the first one neuron The center weight vector, Indicates the first one neuron The covariance matrix, Indicates the first one neuron Number of wins Indicates the first one neuron Information saturation Indicates the first one neuron The parent index is given, and the initial set of neurons is N0 = {n0}, where n0 is the base neuron.

[0013] Furthermore, the competitive learning rule includes detecting the neuron coverage region based on the Mahalanobis distance between the input sample and the neuron according to the following steps:

[0014] Obtain the features of the current sample and use Mahalanobis distance to determine whether each neuron meets the coverage condition: If the condition is met, the corresponding neuron is activated and participates in subsequent competition, where, It has One degree of freedom and confidence level The value of the distribution Equal to the feature dimension of the input sample, Indicates the first Input Samples With the one neuron Mahalanobis distance between them , It is the current set of neurons in the competition layer.

[0015] Furthermore, the competitive learning rule also includes determining the winning neuron according to the following steps:

[0016] During the competition phase, the activated neurons use the recorded covariance matrix to calculate the activation value of the input sample for the neuron through a Gaussian function, and determine the neuron with the largest activation value to win;

[0017] The expression for calculating the activation value of the neuron by the input sample is as follows:

[0018]

[0019] in, Indicates the first Input Samples For the first one neuron activation value, Indicates the first one neuron The covariance matrix, Indicates the first Input Samples With the one neuron Mahalanobis distance between them , It is the current set of neurons in the competition layer.

[0020] Furthermore, in the competitive learning rule, the winning neuron and its corresponding ancestor neuron are recursively updated with the following formula: [Formula omitted for brevity]

[0021]

[0022]

[0023] ;

[0024] in, and These represent the number of wins before and after the current neuron update, respectively. and These represent the center weight vectors of the current neuron before and after the update, respectively. and Let these represent the covariance matrices before and after the current neuron update, respectively. Indicates the first One sample.

[0025] Furthermore, the competitive learning rule also includes the following information saturation update mechanism:

[0026] Winning neurons Regarding the first Input Samples Change in information saturation Based on the input sample and winning neurons Similarity between The calculation yielded:

[0027]

[0028] Where 'a' is the constraint coefficient, taking a value of 1-R, and R is the information saturation reset threshold. Indicates the first k Information saturation of individual neurons, similarity The calculation expression is:

[0029]

[0030] in, Indicates the first Input Samples With winning neurons Mahalanobis distance between them , It is the current set of neurons in the competition layer.

[0031] Furthermore, the competitive learning rule also includes using the recorded covariance matrix to represent the volume covered by the neurons. If two neurons are siblings, both are mature neurons, and meet the merging condition, then neuron merging is performed. The merging condition is that the volume covered by the merged neuron is less than the sum of the volumes covered by the two neurons before merging. The parameters of the two merged neurons are updated using the following formula:

[0032]

[0033]

[0034]

[0035] in, This represents the number of wins for the merged neurons. These represent the number of wins for the two merged neurons, respectively. , These are the center weight vectors of the two merged neurons. The merged neurons are, , These are the covariance matrices of the two merged neurons, respectively. This is the covariance matrix of the merged neurons.

[0036] Furthermore, the competitive learning rule also includes the following information saturation forgetting mechanism:

[0037] The information saturation of mature neurons is adjusted towards the reset threshold according to an adaptive decay function. Attenuation, the adaptive attenuation function is:

[0038]

[0039] in, For the first one neuron Information saturation after decay For the first one neuron Information saturation It has One degree of freedom and confidence level The value of the distribution It is an adjustable parameter.

[0040] Furthermore, the execution steps of the local probability density structure developmental neural network in the clustering process include:

[0041] After receiving the input sample from the input layer, the competition layer performs neuron coverage detection based on the Mahalanobis distance between the input sample and the neuron, so as to activate the local neurons covering the current input sample.

[0042] Entering the competitive learning phase, the activated neurons compete with each other. The activation value of the activated neurons is calculated based on the recorded covariance matrix, and the neuron with the largest activation value is selected as the winning neuron.

[0043] The winning neuron and its corresponding ancestor neuron are adjusted in terms of execution center weight vector and covariance matrix, and the information saturation of the winning neuron is updated based on the similarity between the input sample and the winning neuron.

[0044] If the information saturation of the winning neuron reaches the preset maturity threshold, that is, the winning neuron matures, then it is determined whether the corresponding parent neuron needs to be updated or merged with the mature sibling neuron.

[0045] If the information saturation of the winning neuron reaches a preset saturation threshold, the neuron splits to generate a new neuron, and the information saturation of all mature neurons is decayed according to an information saturation forgetting mechanism. This forgetting mechanism involves the information saturation of mature neurons decreasing towards a reset threshold according to an adaptive decay function. attenuation;

[0046] The category of the current input sample is labeled based on the winning neuron, the layer in the clustering tree where the winning neuron is located, and the preset clustering layer parameters.

[0047] An image incremental clustering apparatus includes a processor and a memory, the memory for storing a computer program and the processor for executing the computer program to perform the method described above.

[0048] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention introduces the concept of local probability density from statistics into neural network design by constructing a structured developmental neural network with local probability density (LPD-SDNN). Second-order statistics are introduced into the neuron model. Each neuron, in addition to recording the first-order central weight vector, also records the second-order moment information (covariance matrix) of the local region data, enabling it to record the shape, orientation, and distribution density of the local region data. This transforms the neuron's coverage area from an isotropic "hypersphere" to a "hyperellipsoid" that can adapt to the data distribution characteristics. Based on this network, by adaptively updating, adding, and fusing neurons, it achieves the ability to incrementally cluster open-world non-steady-state data streams. It can adapt to new classes and learn different coarse-grained categories without prior knowledge or a preset number of neurons. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the implementation principle of a local probability density structure developmental neural network according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the implementation process of the image incremental clustering method based on a local probability density structure developmental neural network according to an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the test results on the impact of the information saturation reset threshold on the number of parameters generated by the LPD-SDNN of this invention under specific instance incremental and class incremental learning scenarios. (a) corresponds to the number of neurons, and (b) corresponds to the clustering performance.

[0052] Figure 4 In a specific application embodiment, the effect diagram of the subdivision function obtained by the LPD-SDNN of the present invention is shown, where (a) to (f) correspond to the subdivision results under different sample numbers (20 to 900). Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0054] This invention targets open environments. Without prior knowledge or a predetermined number of neurons, it constructs a structural developmental neural network with local probability density (LPD-SDNN). This introduces the statistical concept of local probability density into neural network design, incorporating second-order statistics into the neuron model. Each neuron, in addition to recording the first-order central weight vector, simultaneously records the second-order moments (covariance matrix) of the local region data. This allows it to record the shape, orientation, and distribution density of the local data, transforming the neuron's coverage area from an isotropic "hypersphere" to a "hyperellipsoid" capable of adapting to data distribution characteristics. Based on this network, through adaptive updating, adding, and fusing neurons, it achieves incremental clustering of non-stationary data streams in open worlds, enabling adaptation to new classes and learning of categories with varying degrees of coarseness. Furthermore, it utilizes a multivariate Gaussian distribution to model the local data distribution of each neuron, establishing the activation, splitting, fusion, and forgetting mechanisms of neurons based on the response value of the probability density function. This achieves probability-driven topological evolution. Combining information saturation and the Gaussian response function, it can also solve the plasticity-stability balance problem in open environments.

[0055] like Figure 1 As shown, where Figure 1 The left side of the figure shows that in an open environment, the sample feature space has different class structures with different coarse and fine granularities, and new classes will appear. Taking the example in the figure, the coarse-grained class "cat" in the feature space contains the fine-grained classes "gray cat" and "white cat", and as time goes by, a new class "dog" appears. The right side of the figure shows the developmental hierarchical structure of the LPD-SDNN of this invention, which is adapted to the class structure of the sample feature space in the open environment and the human cognitive process. That is, as the number of similar samples learned increases, the model will subdivide the coarse class. The neurons representing the coarse-grained class will split into sub-elements to learn the sub-class distribution. In addition, when a new class appears, the parent element will split into sub-elements to learn the new class distribution.

[0056] like Figure 2 As shown, the steps of the image incremental clustering method based on a local probability density structure developmental neural network in this embodiment include:

[0057] Step S01. Construct a Local Probability Density Structure Developmental Neural Network (LPD-SDNN). The LPD-SDNN is a two-layer network structure. The first layer is an input layer for acquiring input samples, and the second layer is a competition layer composed of multiple neurons for learning the distribution characteristics of the input samples and recording information using the competition mechanism between neurons. The information recorded by each neuron includes at least a center weight vector and a covariance matrix of local region data for modeling the local data distribution.

[0058] Step S02. Obtain image data samples to be clustered and output them to the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) for clustering;

[0059] Step S03. During the clustering process, the competitive layer adjusts the parameters of each neuron according to the preset competitive learning rules, so that each neuron continuously updates and adjusts the corresponding center weight vector and covariance matrix according to the input samples, so as to fit the local probability density information of the input samples until the clustering result is output.

[0060] In this embodiment, the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) inherits and develops the traditional SDNNIS. LPD-SDNN uses a two-layer network. The first layer is the input layer, which acquires input samples. Its size depends on the dimension of the sample features. The second layer is the competition layer, where each neuron learns and records the distribution of the input samples using competition learning rules. Each neuron records not only the center weight vector but also the second-order moment information of its local region data, i.e., the covariance matrix, which enables a more accurate representation of the pattern. Furthermore, the structure of the competition layer is not predefined (such as the number of nodes and the topology) but is adaptively obtained by learning external data incrementally. When the second layer completes sample learning, LPD-SDNN outputs the clustering results of the samples and gives the number of current clusters.

[0061] In this embodiment, a central weight vector is introduced into the Local Probability Density Structure Developmental Neural Network (LPD-SDNN). and covariance matrix Modeling individual neurons means that each neuron includes at least a central weight vector. and covariance matrix By introducing the covariance matrix This enables neurons to perceive the "direction" and "broadening" of data. Specifically, the competition layer can consist of a set of neurons Ni = {n1, ..., nn2}. } indicates that, among which Indicates the sequence number and subscript of the input sample. Indicates processing the first The number of neurons generated after each input sample, element Indicates the first There are *j* generated neurons, where *j* is the encoding of the neuron generation order. Neurons are defined as quintuples. :< , , , , >, among which Indicates the first one neuron The center weight vector, Indicates the first one neuron The covariance matrix, Indicates the first one neuron Number of wins Indicates the first one neuron Information saturation Indicates the first one neuron The parent index.

[0062] Furthermore, the initial set of neurons is defined as N0 = {n0}, where n0 are the base neurons. Setting the base neurons n0 ensures sensitivity to new features and unifies the learning of new classes and subclasses within the neuron splitting mechanism. The central weight vector of the base neurons is the origin of the sample space, and its covariance matrix is ​​denoted as ∞I, meaning the covariance matrix of the initial neuron set is infinite to guarantee the responsiveness of the entire sample space to the input samples. The information saturation is 1, i.e., it is in a saturated state, ensuring that when only base neurons respond (new features appear), new neurons can be instantly split to represent the new features. Specifically, the five-tuple form of the base neurons can be represented as follows: It remains unchanged throughout the entire network development process.

[0063] Preferably, in the competitive learning rule of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment, the neuron coverage region is detected according to the Mahalanobis distance between the input sample and the neuron, following these steps:

[0064] Obtain the features of the current sample and use Mahalanobis distance to determine whether each neuron meets the coverage condition: If the condition is met, the corresponding neuron is activated and participates in subsequent competition, where, It has One degree of freedom and confidence level The value of the distribution Equal to the feature dimension of the input sample, Indicates the first Input Samples With the one neuron Mahalanobis distance between them , It is the current set of neurons in the competition layer.

[0065] This embodiment employs the aforementioned coverage detection method based on Mahalanobis distance and introduces the chi-square distribution ( Using the threshold as the activation criterion allows the coverage area of ​​neurons to be ellipsoidal, and the shape can be automatically adjusted according to the correlation of the data. Compared with the traditional nearest neighbor search method based on Euclidean distance, it can better adapt to long strip or flat data clusters and improve the adaptability of clustering.

[0066] Preferably, in the competitive learning mechanism of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment, the step of determining the winning neuron includes:

[0067] During the competition phase, activated neurons use the recorded covariance matrix to calculate the activation value of the input sample for the neuron through a Gaussian function;

[0068] The neuron with the highest activation value wins.

[0069] Specifically, in the matching phase, when new sample features are input to the input layer, the neurons in the competition layer perform coverage matching. Only neurons whose coverage contains the input pattern will respond and participate in subsequent competition. The matching phase restricts subsequent competitive learning to local neurons, avoiding global computation. Specifically, the matching phase yields a set of metrics for the responding neurons. :

[0070]

[0071] in, It is the i-th input sample. It is the current set of neurons in the competition layer. It has d degrees of freedom and q confidence level. The distribution value, d is equal to the dimension of the input sample, and q can be set according to actual needs, preferably 0.99. Indicates the first Input Samples With the one neuron Mahalanobis distance between them:

[0072]

[0073] in, , It is the current set of neurons in the competition layer.

[0074] During the competition phase, the response neurons calculate the activation value of the input sample to the neuron using a Gaussian function, expressed as:

[0075]

[0076] in, Indicates the first Input Samples For the first one neuron activation value, Indicates the first one neuron The covariance matrix.

[0077] Then, the neuron with the highest activation value is selected as the winner. The winning neuron is labeled with k, and its calculation expression is:

[0078]

[0079] This embodiment employs the aforementioned competition mechanism based on Mahalanobis distance, using Mahalanobis distance for similarity calculation and combining it with the chi-squared distribution to set a dynamic confidence threshold, thereby achieving accurate coverage and identification of non-spherical and non-equilibrium distributed data.

[0080] Preferably, in the competitive learning mechanism of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment, the parameters of each neuron are adjusted using the following steps to achieve the update of the winning neuron:

[0081] If k ≠ 0, the winning neuron and its corresponding ancestor neuron are recursively updated with the following formula: [Formula omitted for brevity]

[0082] (5)

[0083] (6)

[0084] (7)

[0085] in, and These represent the number of wins before and after the current neuron update, respectively. and These represent the center weight vectors of the current neuron before and after the update, respectively. and Let these represent the covariance matrices before and after the current neuron update, respectively. Indicates the first One sample.

[0086] Traditional calculation of the covariance matrix requires batch data. This embodiment achieves the updating of the winning neuron by adopting an online recursive update of the covariance matrix. By using incremental learning, the covariance matrix can be updated using only the current input sample and the old statistics, ensuring that the network can converge to the true statistical features of the data in real time during a single traversal of the data stream, without the need to store historical samples.

[0087] In order to learn new class and subclass structures, the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment adopts a neuron expansion mechanism based on information saturation. Information saturation reflects the amount of information that a neuron has learned. When a neuron carries too much information, it means that the neuron is frequently activated and will generate new neurons to represent possible subclasses.

[0088] Preferably, in the competitive learning mechanism of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment, the following information saturation update mechanism is adopted:

[0089] Winning neurons Regarding the first Input Samples Change in information saturation Based on the input sample and winning neurons Similarity between The calculation yielded:

[0090] (8)

[0091] Where 'a' is the constraint coefficient, taking a value of 1-R, and R is the information saturation reset threshold, which is the value of R that indicates the information saturation after neuron splitting will revert to its original value. Indicates the first k Information saturation of individual neurons, similarity The calculation expression is:

[0092] (9)

[0093] in, Indicates the first Input Samples With winning neurons Mahalanobis distance between them , It is the current set of neurons in the competition layer.

[0094] This embodiment calculates the increment of information saturation based on the response value of the Gaussian radial basis function (RBF). Meanwhile, the response value is determined by Mahalanobis distance, which allows the maturation process of neurons to be directly related to probability density and has stronger statistical significance.

[0095] Neuronal development is divided into two stages based on information saturation: immature and mature. When information saturation is below the reset threshold R, the neuron is in the immature stage. When information saturation exceeds the reset threshold, the neuron enters the mature stage. Neurons exhibit different cluster markers and behaviors at different developmental stages; therefore, it is believed that new neurons in the initial stage can only represent new clusters after being activated a sufficient number of times.

[0096] To mitigate the confusion caused by the randomness of the input order, this embodiment establishes a parent-child relationship update mechanism: when a neuron... When it matures and moves out of the parent element's coverage area, that is... This indicates that the neuron is no longer a child of the cluster represented by its parent neuron. The current parent-child connection is deleted, and the neuron is moved up one level in the parent-child relationship tree. A new parent neuron is then searched for in the higher levels. The new parent is determined by... Determined by whether it is within the coverage area of ​​the neuron. If no new parent is found, it is promoted to another layer until a new parent is found or it becomes a neuron in the first layer.

[0097] Preferably, the competitive learning rule of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment further includes using the recorded covariance matrix to represent the volume covered by the neuron. If two neurons satisfy the conditions of sibling relationship, are both mature neurons, and meet the merging condition, then the neurons are merged. The merging condition is that the volume covered by the merged neuron is less than the sum of the volumes covered by the two neurons before merging. The parameters of the two neurons after merging are updated according to the following formula:

[0098] (10)

[0099] (11)

[0100] (12)

[0101] in, This represents the number of wins for the merged neurons. These represent the number of wins for the two merged neurons, respectively. , These are the center weight vectors of the two merged neurons. The merged neurons are, , These are the covariance matrices of the two merged neurons, respectively. This is the covariance matrix of the merged neurons.

[0102] This embodiment employs a merging criterion based on the covering domain volume, utilizing the determinant of the covariance matrix. The volume representing the coverage of neurons is used, and the merging condition is "the merged volume < the sum of the two volumes before merging". This can avoid wasting computational resources on excessively overlapping neurons, and the merging process can maintain the compactness of the probability distribution.

[0103] Specifically, if the winning neuron is mature, it is checked whether it can merge with its responding sibling neuron to reduce the number of redundant neurons. Specifically, the following three conditions can be met to determine whether two neurons can merge:

[0104] (1) The two neurons are brothers, that is, they have the same parent neuron.

[0105] (2) Both neurons are mature neurons.

[0106] (3) Combined neurons n c Its volume is smaller than that of two neurons and The sum of their volumes, that is:

[0107]

[0108] If node and Merge nodes if the above conditions are met. and And update the parameters according to equations (10) to (12), and combine the neurons n c Inheritance n i and n j The sub-neurons and the patterns they represent.

[0109] Furthermore, if the information saturation reaches the saturation threshold, meaning that neurons are frequently activated and carry enough information to accurately represent local pattern distributions, the neurons will split to generate new neurons to further represent finer-grained clusters. The input sample that triggers the splitting of the winning neuron is used as the center weight vector. and Initialize them all to 0. It is the indicator of the winning neuron. Initialization depends on whether the parent neuron is a base neuron:

[0110]

[0111] Where I is the identity matrix and σ is a small positive number.

[0112] Preferably, in the competitive learning rule of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) in this embodiment, in order to prevent frequent neuron splitting and improve model stability, the following information saturation forgetting mechanism is also included:

[0113] The information saturation of mature neurons is adjusted towards the reset threshold according to an adaptive decay function. Attenuation, the adaptive attenuation function is:

[0114] (15)

[0115] in, For the first one neuron Information saturation after decay For the first one neuron Information saturation It has One degree of freedom and confidence level The value of the distribution It is an adjustable parameter.

[0116] This embodiment employs an information saturation forgetting mechanism based on an adaptive decay function of the chi-square distribution boundary. The decay rate is related to whether the sample falls within the region of "low relevance but still belonging to the class" (affected by...). (Parameter control), therefore by explicitly adjusting It can control the range of samples that may trigger neuron splitting. This mechanism can effectively prevent noisy data from causing neuron oversaturation and meaningless splitting, thus enhancing the robustness of the model during long-term operation.

[0117] In this embodiment, the entire workflow of the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) during the clustering process is detailed as follows:

[0118] Response detection: After receiving the input sample from the input layer, the competitive layer performs neuron coverage detection based on the Mahalanobis distance between the input sample and the neuron, in order to activate the local neurons covering the current input sample. For example, Mahalanobis distance is used to determine whether each neuron satisfies the coverage condition. If the condition is met, the corresponding neuron is activated and participates in subsequent competition;

[0119] Competitive activation: Entering the competitive learning stage, that is, the activated neurons compete with each other. The activation value of the activated neurons is calculated based on the recorded covariance matrix. The calculation expression is shown in equation (3), that is... The neuron with the highest activation value is selected as the winning neuron.

[0120] Parameter update: The weight vector and covariance matrix of the execution center of the winning neuron and its corresponding ancestor neuron are adjusted as shown in Equations (6) and (7), and the information saturation of the winning neuron is updated according to the similarity between the input sample and the winning neuron as shown in Equations (8) and (9);

[0121] Neuron merging: If the information saturation of the winning neuron reaches the preset maturity threshold, that is, the winning neuron matures, then it is detected whether the corresponding parent neuron needs to be updated or merged with the mature sibling neuron. The parameters after merging are shown in equations (10) to (12).

[0122] Neuron splitting: If the information saturation of the winning neuron reaches a preset saturation threshold, the neuron splits to generate a new neuron. The information saturation of all mature neurons is then decayed according to an information saturation forgetting mechanism. This mechanism involves the information saturation of mature neurons decreasing according to an adaptive decay function towards a reset threshold. The attenuation is as shown in equation (15);

[0123] Clustering output: The category of the current input sample is labeled according to the winning neuron, the layer in the clustering tree where the winning neuron is located, and the preset clustering layer parameters.

[0124] This embodiment, by employing the above steps, stores the covariance matrix in LPD-SDNN, enabling the recording of richer cluster information and more accurate representation of patterns in the neighborhood. Supported by statistical theory, it utilizes Mahalanobis distance and sensing threshold to make the neuron coverage more interpretable. Furthermore, by merging mature sibling neurons, redundant neurons can be removed. The neuron splitting mechanism based on information saturation allows the network to learn new classes and recognize subclass structures like human cognitive processes, thereby fully leveraging the advantages of probability density models and SDNNIS. By establishing the activation, splitting, fusion, and forgetting mechanisms of neurons on the response value of the probability density function, and by combining information saturation with the Gaussian response function, it can also effectively solve the plasticity-stability balance problem in open environments.

[0125] This embodiment, based on LPD-SDNN, effectively combines prototype-based clustering and hierarchical clustering. Specifically, the central weight vector of each neuron can serve as a prototype in the cluster, and the number of prototypes does not need to be pre-specified but is dynamically adjusted based on the input samples. A tree-like hierarchical structure is established by dividing the parent-child relationships formed by neurons, where child neurons further subdivide the classes represented by parent neurons, and the user can set the clustering layer parameter CL according to the desired degree of class subdivision.

[0126] To verify the effectiveness of this invention, experiments were conducted on artificial and real datasets using the method of this invention, and comparisons were made with typical existing neural network structures such as GNG, GWR, SOINN, and SDNNIS, as detailed below:

[0127] 1. Dataset: This embodiment uses artificial datasets (as shown in Table 1) and real-world datasets to verify the LPD-SDNN of this invention. Considering the most common distributions in visualization and nature, this embodiment constructs a two-dimensional normal distribution dataset with different distribution characteristics, wherein the real-world dataset is obtained from the AT&T (ORL Face Database) face database.

[0128] Table 1: List of Artificial Datasets

[0129]

[0130] 2. Evaluation Metrics: The clustering performance is evaluated using the external and internal evaluation metrics, Accuracy (ACC) and Davies-Bouldin Index (DBI). Accuracy (ACC) represents the proportion of correctly clustered data, and its calculation formula is as follows:

[0131]

[0132] in and These represent the predicted labels and the true labels, respectively. n is the total number of data points, and `map()` is the Hungarian algorithm function used to achieve optimal label assignment. The ACC value ranges from 0 to 1. A higher ACC value indicates a higher proportion of correctly clustered results.

[0133] The Davies-Bouldin index (DBI) is defined as follows for a cluster partition C = {C1, C2, ..., Ck}:

[0134]

[0135] Where avg(C) corresponds to the average distance between samples in cluster C, and Corresponding to clustering Centroid and Clustering The distance between them. Obviously, the smaller the DBI value, the better the clustering result.

[0136] 3. Comparison Methods: In addition to comparison with SDNNIS, this embodiment also uses structural developmental neural networks GNG, GWR, and SOINN for comparison and verification. All three networks involve a hyperparameter: agemax, which is used as an age threshold for connection edges. In the following comparison experiments, this hyperparameter was uniformly set to the optimal value agemax = 5. Furthermore, the periodicity parameter λ of neurons in GNG and SOINN was slightly adjusted based on the amount of experimental data.

[0137] 4. Hyperparameter R Analysis: The information saturation reset threshold R reflects the winning frequency required for neurons to reach maturity and defines the callback value of information saturation after neuron splitting. When R is large, the central weight vector of the neuron is relatively stable in the maturity stage. However, since R affects both the amount of information required for mature neurons to split and the rate of increase of information saturation, its impact on the network growth rate is relatively complex.

[0138] The number of neurons and the DBI value of the clustering results generated by the LPD-SDNN of this invention were tested on DS2 with different values ​​of R, which to some extent reflects the influence of R on the LPD-SDNN of this invention. According to the meaning of the information saturation reset threshold, neurons at the information saturation reset threshold should be in a relatively stable state. Therefore, this embodiment tested the influence of R ∈ [0.50, 0.95] (within the interval 0.02) on the number of neurons and the DBI value of the clustering results of the LPD-SDNN under two learning scenarios: instance increment and class increment. Since the LPD-SDNN of this invention is sensitive to the sample input order, each test was performed 30 times according to different sample input orders, and the final result is the average of the 30 test results.

[0139] like Figure 3 As shown, by Figure 3 Figure (a) reflects the effect of the reset threshold on the number of neurons generated by the network. In the two different learning scenarios, the number of neurons decreases as the reset threshold increases, reflecting that the constraint effect of the reset threshold on the growth of information saturation is dominant. That is, the larger R is, the slower the growth of information saturation and the more difficult it is for neurons to split. Figure 3 (b) reflects that R has little impact on the DBI value of the clustering results.

[0140] Considering the network's operating efficiency, plasticity stability, and clustering performance, this embodiment sets the hyperparameter R to 0.8 to maintain a good DBI level and a low number of neurons.

[0141] 5. Network Development Process: The LPD-SDNN of this invention inherits the segmentation function of SDNNIS. The segmentation capability of the LPD-SDNN of this invention under different sample numbers (No. of samples) is as follows: Figure 4 As shown, dots represent input samples, stars represent neuron center vectors, circles represent the boundaries of neuron coverage areas, and different colors represent different categories. The test dataset consists of DS2 clusters of 6 clusters following a normal distribution, with some clusters overlapping each other. Overlapping clusters can be considered as one large cluster, so the dataset can be viewed as two large clusters, each consisting of 3 sub-clusters. Figure 4 This demonstrates that as the number of input samples increases, the LPD-SDNN of this invention gradually identifies sub-clusters. That is, as the number of learning samples increases, the neurons in the LPD-SDNN of this invention further subdivide the coarse-grained classes, simulating the human cognitive process.

[0142] 6. Incremental unsupervised learning on artificial datasets;

[0143] The LPD-SDNN of this invention can adapt to class-incremental unsupervised learning problems in open environments, discovering new classes and subdividing recurring classes through structural evolution. This embodiment verifies the clustering performance of the LPD-SDNN on artificial datasets in instance-incremental and class-incremental learning scenarios, and compares it with typical structural evolutionary neural networks such as GNG, GWR, SOINN, and SDNNIS.

[0144] This embodiment first uses the same artificial dataset DS1 used in the SDNNIS test, which contains five Gaussian distributions with the same variance and sample size, and the distributions do not overlap. On DS1, SDNNIS achieves high clustering accuracy in the incremental learning scenario, but generates too many redundant neurons. This embodiment tests LPD-SDNN on DS1. The parameters of LPD-SDNN are set to R = 0.8 and CL = 1. The LPD-SDNN of this invention achieves higher clustering accuracy in the incremental learning scenario while generating far fewer neurons.

[0145] This embodiment then uses the artificial dataset DS2 to test the LPD-SDNN of the present invention (CL = 2). DS2 consists of 6 clusters. Each cluster has a different size and density, and there is overlap between clusters, which can better characterize the complexity of sample distribution in open environments. On this more complex dataset, the LPD-SDNN of the present invention achieves better clustering results while generating fewer neurons in instance and class incremental learning scenarios.

[0146] This embodiment also tested three other structural developmental neural networks, GNG (λ = 20), GWR, and SOINN (λ = 30), on these datasets for comparison. The estimated clustering results of each network and the number of neurons generated in these networks are shown in Table 2.

[0147] Table 2: Comparison of clustering accuracy and final number of neurons of structural developmental neural networks on artificial datasets of different complexities under different incremental learning scenarios.

[0148]

[0149] As can be seen from Table 2, the LPD-SDNN of this invention achieves almost the best clustering results and generates the fewest neurons on these datasets in both instance increment and class increment learning scenarios. This proves that the LPD-SDNN of this invention is capable of undertaking continuous unsupervised learning tasks on complex distributed datasets.

[0150] 7. Continuous unsupervised learning on real-world datasets;

[0151] This embodiment tests the clustering performance of LPD-SDNN in incremental task scenarios on the AT&T face database and compares it with SOINN (λ = 50) and SDNNIS (CL = 1).

[0152] The AT&T face database contains 40 different subjects, with 10 different images for each subject. For some subjects, these images were taken at different times, with variations in lighting, facial expressions (eyes open / closed, smiling / not smiling), and facial details (wearing glasses / not wearing glasses). All images were taken against a dark, uniform background, with the subjects in an upright, frontal position (allowing some lateral movement). Each image is 92 x 112 pixels in size, with 256 gray levels per pixel.

[0153] In this embodiment, 15 subjects were selected from the database and divided into two groups, A and B. Group A contained 10 subjects, and group B contained 5 subjects. As the number of training samples increases, the neurons in the LPD-SDNN of this invention further subdivide the coarse-grained classes, simulating the human cognitive process. The purpose of dividing the subjects into two groups is to simulate two different environments in an open environment, representing two tasks, thereby testing the clustering performance of the proposed structural developmental neural network under incremental task conditions. The image feature vector extraction is as follows: the image size is adjusted to 23x28 using bilinear interpolation, and then the adjusted image is downsampled using the PCA method to obtain a 10-dimensional feature vector.

[0154] The test consists of two tasks, each with a learning and evaluation phase. The training set for the learning phase of the first task is group A images (10 classes, 10 samples per class), with class-based incrementing. In the evaluation phase, images of group A subjects with added Gaussian noise are used as the test set, and the input order is randomized. The training set for the learning phase of the second task is group B images (5 classes, 10 samples per class), with instance-based incrementing. The second task in the evaluation phase tests not only the clustering results of the LPD-SDNN on the group B data but also whether the learning results of the first task are corrupted after learning from the group B data. Therefore, the test set for the evaluation phase of the second task is images of both groups of subjects with added Gaussian noise, and the input order is randomized. Because the structural developmental neural network is sensitive to the sample input order, the test is repeated 30 times with different input orders. The average clustering accuracy is used as the performance evaluation metric.

[0155] The LPD-SDNN of this invention achieves a 100% accuracy rate in face clustering. Table 3 shows the average ACC and the specific number of neurons generated by each network. As shown in Table 3, the LPD-SDNN of this invention achieves optimal accuracy using the smallest network size, meaning it can perform continuous unsupervised learning in incremental task scenarios with relatively high-dimensional datasets. Furthermore, Table 3 shows that after learning the second task, the clustering accuracy of the LPD-SDNN for group A is equal to that in the first task; therefore, the learning results of the LPD-SDNN on the first task are not compromised by the second task. Additionally, GNG and GWR are not suitable for this task because they cannot separate subjects. Experimental results demonstrate that the LPD-SDNN of this invention is capable of performing incremental unsupervised learning on face databases.

[0156] Table 3: Clustering accuracy and final network size of structural developmental neural networks on face datasets in incremental learning scenarios.

[0157]

[0158] In summary, this invention addresses incremental image clustering in open environments by employing a structural developmental neural network with local probability density. Each neuron adjusts its center weight vector and covariance matrix based on input samples to fit local probability density information, and the response range is determined using statistically supported thresholds. A neuron splitting mechanism based on information saturation enables the network to learn new classes and recognize subclass structures, adapting to continuous learning in open environments. Experiments demonstrate that the LPD-SDNN of this invention exhibits excellent incremental clustering performance for both artificial and real data.

[0159] This embodiment further provides an image incremental clustering apparatus, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the method as described above.

[0160] Those skilled in the art will understand that the above embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce implementations of the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0161] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. An image incremental clustering method based on a local probability density structure developmental neural network, characterized by the following steps: include: A Local Probability Density Structure Developmental Neural Network (LPD-SDNN) is constructed. The LPD-SDNN has a two-layer network structure: the first layer is the input layer for acquiring input samples, and the second layer is a competition layer composed of multiple neurons. This competition layer is used to learn the distribution characteristics of the input samples and record information. Each neuron records at least a center weight vector and a covariance matrix of the local region data used to model the local data distribution. The execution steps of the LPD-SDNN in the clustering process include: After receiving the input sample from the input layer, the competition layer performs neuron coverage detection based on the Mahalanobis distance between the input sample and the neuron, so as to activate the local neurons covering the current input sample. Entering the competitive learning phase, the activated neurons compete with each other. The activation value of the activated neurons is calculated based on the recorded covariance matrix, and the neuron with the largest activation value is selected as the winning neuron. The winning neuron and its corresponding ancestor neuron are adjusted in terms of execution center weight vector and covariance matrix, and the information saturation of the winning neuron is updated based on the similarity between the input sample and the winning neuron. If the information saturation of the winning neuron reaches the preset maturity threshold, that is, the winning neuron matures, then it is determined whether the corresponding parent neuron needs to be updated or merged with the mature sibling neuron. If the information saturation of the winning neuron reaches a preset saturation threshold, the neuron splits to generate a new neuron, and the information saturation of all mature neurons is decayed according to an information saturation forgetting mechanism. This forgetting mechanism involves the information saturation of mature neurons decreasing according to an adaptive decay function towards a reset threshold. attenuation; The category of the current input sample is labeled based on the winning neuron, the layer in the clustering tree where the winning neuron is located, and the preset clustering layer parameters; The image data samples to be clustered are obtained and output to the Local Probability Density Structure Developmental Neural Network (LPD-SDNN) for clustering. During the clustering process, the competitive layer adjusts the parameters of each neuron according to the preset competitive learning rules, so that each neuron continuously updates and adjusts the corresponding center weight vector and covariance matrix according to the input samples to fit the local probability density information of the input samples until the clustering result is output.

2. The image incremental clustering method based on a local probability density structure developmental neural network according to claim 1, characterized in that, Each neuron in the competition layer is represented as a quintuple. :< , , , , >, among which Indicates the first one neuron The center weight vector, Indicates the first one neuron The covariance matrix, Indicates the first one neuron Number of wins Indicates the first one neuron Information saturation Indicates the first one neuron The parent index is given, and the initial set of neurons is N0 = {n0}, where n0 is the base neuron.

3. The image incremental clustering method based on a local probability density structure developmental neural network according to claim 1, characterized in that, The competitive learning rule includes detecting the neuron coverage region based on the Mahalanobis distance between the input sample and the neuron, following these steps: Obtain the features of the current sample and use Mahalanobis distance to determine whether each neuron meets the coverage condition: If the condition is met, the corresponding neuron is activated and participates in subsequent competition, where, It has One degree of freedom and confidence level The value of the distribution Equal to the feature dimension of the input sample, Indicates the first Input Samples With the one neuron Mahalanobis distance between them , It is the current set of neurons in the competition layer.

4. The image incremental clustering method based on a local probability density structure developmental neural network according to claim 1, characterized in that, The competitive learning rule also includes determining the winning neuron according to the following steps: During the competition phase, the activated neurons use the recorded covariance matrix to calculate the activation value of the input sample for the neuron through a Gaussian function, and determine the neuron with the largest activation value to win; The expression for calculating the activation value of the neuron by the input sample is as follows: in, Indicates the first Input Samples For the one neuron activation value, Indicates the first one neuron The covariance matrix, Indicates the first Input Samples With the one neuron Mahalanobis distance between them , It is the current set of neurons in the competition layer.

5. The image incremental clustering method based on a local probability density structure developmental neural network according to claim 1, characterized in that, In the competitive learning rule, the winning neuron and its corresponding ancestor neuron are recursively updated with the winning count, center weight vector, and covariance matrix according to the following formula: ； in, and These represent the number of wins before and after the current neuron update, respectively. and These represent the center weight vectors of the current neuron before and after the update, respectively. and Let these represent the covariance matrices before and after the current neuron update, respectively. Indicates the first One sample.

6. The image incremental clustering method based on a local probability density structure developmental neural network according to claim 1, characterized in that, The competitive learning rule also includes the following information saturation update mechanism: Winning neurons Regarding the first Input Samples Change in information saturation Based on the input sample and winning neurons Similarity between The calculation yielded: in, This is the constraint coefficient, with a value of 1-R, where R is the information saturation reset threshold. Indicates the first k Information saturation of individual neurons, similarity The calculation expression is: in, Indicates the first Input Samples With winning neurons Mahalanobis distance between them , It is the current set of neurons in the competition layer.

7. The image incremental clustering method based on a local probability density structure developmental neural network according to any one of claims 1 to 6, characterized in that, The competitive learning rule also includes using the recorded covariance matrix to represent the volume covered by neurons. If two neurons are siblings, both are mature neurons, and meet the merging condition, then neuron merging is performed. The merging condition is that the volume covered by the merged neuron is less than the sum of the volumes covered by the two neurons before merging. The parameters of the two neurons after merging are updated according to the following formula: in, This represents the number of wins for the merged neurons. These represent the number of wins for the two merged neurons, respectively. , These are the center weight vectors of the two merged neurons. This is the center weight vector of the merged neurons. , These are the covariance matrices of the two merged neurons, respectively. This is the covariance matrix of the merged neurons.

8. The image incremental clustering method based on a local probability density structure developmental neural network according to any one of claims 1 to 6, characterized in that, The competitive learning rule also includes the following information saturation forgetting mechanism: The information saturation of mature neurons is adjusted towards the reset threshold according to an adaptive decay function. Attenuation, the adaptive attenuation function is: in, For the first one neuron Information saturation after decay For the first one neuron Information saturation It has One degree of freedom and confidence level The value of the distribution It is an adjustable parameter. It is the constraint coefficient.

9. An image incremental clustering device, comprising a processor and a memory, the memory being used to store a computer program, characterized in that, The processor is used to execute the computer program to perform the method as described in any one of claims 1 to 8.

Images