Image clustering method and system based on deep subspace clustering sequential ensemble

Abstract

The application discloses an image clustering method and system based on deep subspace clustering sequential integration, and the method comprises the following steps: obtaining an initial image encoder network and an image decoder network based on input image data and reconstructed image data; obtaining an initial self-expression coefficient matrix based on a random masking visual feature vector and a self-expression visual feature vector; fixing the self-expression coefficient matrix, and iteratively updating the image encoder network and the image decoder network; fixing the image encoder network and the image decoder network, and iteratively updating the self-expression coefficient matrix; iteratively updating an image similarity matrix and an image integrated similarity matrix based on the self-expression coefficient matrix; and taking the finally obtained image integrated similarity matrix as the input of a spectral clustering algorithm to obtain an image clustering result. The self-expression coefficient matrix is sequentially generated and directly integrated in a sequential mode, so that the performance of the deep subspace clustering network for image clustering is improved.

Description

Technical Field

[0001] This invention relates to the field of image clustering technology, and in particular to an image clustering method and system based on deep subspace clustering sequence integration. Background Technology

[0002] Deep learning performs multi-level learning on image data representations, enabling it to learn more abstract and advanced feature representations from naive raw features. The non-linear feature learning capability of deep networks can solve the problem of insufficient performance of traditional clustering models on non-linear datasets. Therefore, the fusion framework of deep networks and clustering models has become the focus of image clustering research. Image clustering models combining deep learning can be roughly divided into two categories: (1) step-by-step learning strategy. An alternating iterative update method is used to update the feature representations learned by the deep autoencoder network and the partitioning results obtained by clustering modules (such as spectral clustering, K-means, etc.); (2) joint learning strategy. The main idea of ​​this strategy is to integrate feature representation learning and data clustering into the same objective function. By optimizing the objective function, feature representations and clustering partitions can be learned simultaneously. In terms of performance, both of these learning strategies have a significant improvement over traditional clustering models. Zhan et al. found that the training strategy based on joint learning is better than the training strategy based on step-by-step learning. Because the training method of alternating parameter updates leads to unstable latent feature representations, the clustering performance is poor.

[0003] However, when processing real-world image data, the fusion of deep networks and clustering models can be affected by various factors, leading to decreased robustness and performance. For example, network weight initialization is one of the main problems in neural network training; using random initialization introduces randomness into the learned representations. Different network architectures (such as the number and width of layers) can cause the network to learn different information from the original representations, resulting in different feature representations. Therefore, the performance of deep clustering methods is often heavily dependent on specific network structures and hyperparameter choices. Deep subspace clustering that relies on autoencoders is similarly susceptible to the influence of the autoencoder's network structure and hyperparameter settings.

[0004] Therefore, combining deep learning and ensemble learning is a worthwhile area of ​​exploration. Ensemble methods for deep clustering typically utilize hyperparameter settings (such as random parameter initialization and network framework design) to train multiple deep networks in parallel (i.e., obtain multiple base learners), and then fuse them using a specific merging method. For example, Affeldt et al. proposed Spectral Clustering Based on Ensemble Deep Autoencoder Learning (SC-EDAE), whose core idea is to first obtain multiple differential latent feature representations using random parameter settings (such as different network frameworks and different parameter initializations), then merge them into a similarity matrix, and finally use spectral clustering methods to obtain the final partition. Han proposed a Deep Ensemble Clustering (DCE) model, whose core idea is to pre-train multiple autoencoder networks using different initialization methods, select appropriate autoencoder networks using a certain evaluation metric, and finally combine the selected high-quality autoencoder networks to train the ensemble model, obtaining the final clustering result.

[0005] Sequential ensemble means that a base learner is generated in each iteration and then fused together until the iteration ends to obtain the final ensemble result. However, deep clustering based on stepwise learning leads to instability in the learned latent feature representations due to the alternating updates between network parameter updates and representation clustering. Summary of the Invention

[0006] The main objective of this invention is to overcome the problems of existing technologies and propose an image clustering method and system based on deep subspace clustering sequential integration. This method fully utilizes the instability of feature representation in alternating training to learn diverse self-expression coefficient matrices, that is, by iteratively updating the parameters of the autoencoder network and the self-expression layer through alternating iterations. This step-by-step training strategy, combined with a random masking strategy, can sequentially generate diverse self-expression coefficient matrices and similarity matrices, and then integrate them step by step until the iteration ends and the final integration result is obtained.

[0007] The present invention adopts the following technical solution:

[0008] On the one hand, an image clustering method based on deep subspace clustering order integration includes the following steps:

[0009] Step 1: Construct an image reconstruction loss function based on the input image data and the reconstructed image data; obtain the initial image encoder network and image decoder network by minimizing the image reconstruction loss function.

[0010] Step 2: Construct a first self-expression loss function based on the error between the random occlusion visual feature vector and the self-expression visual feature vector, and obtain an initial self-expression coefficient matrix by minimizing the first self-expression loss function; wherein, the random occlusion visual feature vector is obtained based on the input image, the image encoder network, and the random occlusion matrix;

[0011] Step 3: Fix the self-expression coefficient matrix, and update the image encoder network and image decoder network based on the joint loss function of image reconstruction and self-expression; repeat this step until the first iteration is reached;

[0012] Step 4: Fix the image encoder network and image decoder network, and update the self-expression coefficient matrix based on the second self-expression loss function; repeat this step until the second iteration number is reached;

[0013] Step 5: Update the image similarity matrix and the image ensemble similarity matrix based on the self-expression coefficient matrix;

[0014] Step 6: Repeat steps 3 to 5 until the third iteration is reached;

[0015] Step 7: Use the final image integration similarity matrix as input to the spectral clustering algorithm to obtain the image clustering results.

[0016] Preferably, the reconstruction loss function is expressed as follows:

[0017]

[0018] Where L0 represents the reconstruction loss function; X represents the input image data; This represents the reconstructed image data; It represents the square of the F-norm of the matrix.

[0019] Preferably, the reconstructed image The reconstructed visual feature vector Z′C from the self-expression is represented as follows:

[0020]

[0021] Z′ represents the following:

[0022] Z′=Z⊙M

[0023] in, denoted as the image decoder network; C represents the self-expression coefficient matrix; Z represents the compressed visual feature vector of the extracted input image data; Z′ represents the visual feature vector after random occlusion; M represents the random occlusion matrix; ⊙ represents the Hadamard product;

[0024] Z represents the following:

[0025]

[0026] in, This represents an image encoder network.

[0027] Preferably, the first self-expression loss function is expressed as follows:

[0028]

[0029] Among them, L exp1 Represents the first self-expression loss function; ||C|| p λ represents the regularization loss; λ represents the weighting coefficient of the error between the visual feature vector Z′ and the self-expressed visual feature vector Z′C.

[0030] Preferably, in step 3, the joint loss function for image reconstruction and self-representation is expressed as follows:

[0031]

[0032] Where λ1 represents the weight of the self-expression loss component.

[0033] Preferably, in step 4, the second self-expression loss function is expressed as follows:

[0034]

[0035] Among them, L exp2 λ represents the second self-expression loss function; λ2 represents the weight of the self-expression loss component.

[0036] Preferably, in step 5, the image similarity matrix and the image ensemble similarity matrix are updated based on the self-expression coefficient matrix, as follows:

[0037] The image similarity matrix W is represented as follows:

[0038]

[0039] Image ensemble similarity matrix F t It is expressed as follows:

[0040]

[0041] Among them, F t F represents the ensemble similarity matrix of the images generated at the current time. t-1 W represents the image ensemble similarity matrix from the previous time step. t This represents the image similarity matrix generated at the current time.

[0042] On the other hand, an image clustering system based on deep subspace clustering sequence integration includes: a base learner generation module and an integration module;

[0043] The base learner generation module includes a feature compression module, a feature reconstruction module, a random occlusion module, and a self-representation module. The feature compression and feature reconstruction modules are used to learn visual feature representations, and both modules consist of multiple convolutional network layers. The base learner generation module sequentially generates a self-representation coefficient matrix C, which in turn generates an image similarity matrix. The integration module is used to successively merge the generated image similarity matrices, and the image integration similarity matrix F generated in each iteration is... t The similarity matrix F is integrated from the images at the previous time step. t-1 The result is obtained by adding the image similarity matrix W generated at the current time, as follows:

[0044]

[0045] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0046] (1) This invention fully utilizes the instability of feature representation in alternating training to learn diverse self-expression coefficient matrices, that is, by alternating iteratively updating the parameters of the autoencoder network and the parameters of the self-expression layer; this step-by-step training strategy combined with the random masking strategy can sequentially generate diverse and more robust self-expression coefficient matrices and similarity matrices, and then integrate them one by one until the iteration ends to obtain the final integration result, thereby improving the final clustering performance.

[0047] (2) Since only the self-expression coefficient matrix and similarity matrix of the current iteration need to be saved each time, this method can also save memory overhead. Attached Figure Description

[0048] Figure 1 This is a flowchart of the image clustering method based on deep subspace clustering sequence integration of the present invention;

[0049] Figure 2 This is a structural block diagram of the image clustering system based on deep subspace clustering sequence integration according to the present invention;

[0050] Figure 3 This is a partial example of the dataset (input image data) in an embodiment of the present invention; wherein, (a) represents ORL; (b) represents COIL20; (c) represents EYaleB; and (d) represents COIL100. Detailed Implementation

[0051] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0052] See Figure 1 As shown, the image clustering method based on deep subspace clustering sequence integration of the present invention includes the following steps:

[0053] Step 1: Construct an image reconstruction loss function based on the input image data and the reconstructed image data; obtain the initial image encoder network and image decoder network by minimizing the image reconstruction loss function.

[0054] Step 2: Construct a first self-expression loss function based on the error between the random occlusion visual feature vector and the self-expression visual feature vector, and obtain an initial self-expression coefficient matrix by minimizing the first self-expression loss function; wherein, the random occlusion visual feature vector is obtained based on the input image, the image encoder network, and the random occlusion matrix;

[0055] Step 3: Fix the self-expression coefficient matrix, and update the image encoder network and image decoder network based on the joint loss function of image reconstruction and self-expression; repeat this step until the first iteration is reached;

[0056] Step 4: Fix the image encoder network and image decoder network, and update the self-expression coefficient matrix based on the second self-expression loss function; repeat this step until the second iteration number is reached;

[0057] Step 5: Update the image similarity matrix and the image ensemble similarity matrix based on the self-expression coefficient matrix;

[0058] Step 6: Repeat steps 3 to 5 until the third iteration is reached;

[0059] Step 7: Use the final image integration similarity matrix as input to the spectral clustering algorithm to obtain the image clustering results.

[0060] Specifically, the reconstruction loss function is expressed as follows:

[0061]

[0062] Where L0 represents the reconstruction loss function; X represents the input image data; This represents the reconstructed image data; It represents the square of the F-norm of the matrix.

[0063] Reconstructed image The reconstructed visual feature vector Z′C from the self-expression is represented as follows:

[0064]

[0065] Z′ represents the following:

[0066] Z′=Z⊙M

[0067] in, denoted as the image decoder network; C represents the self-expression coefficient matrix; Z represents the compressed visual feature vector of the extracted input image data; Z′ represents the visual feature vector after random occlusion; M represents the random occlusion matrix; ⊙ represents the Hadamard product;

[0068] Z represents the following:

[0069]

[0070] in, This represents an image encoder network.

[0071] The first self-expression loss function is expressed as follows:

[0072]

[0073] Among them, L exp1 Represents the first self-expression loss function; ||C|| p λ represents the regularization loss; λ represents the weighting coefficient of the error between the visual feature vector Z′ and the self-represented visual feature vector Z′C. It represents the square of the F-norm of the matrix.

[0074] In step 3, the joint loss function for image reconstruction and self-representation is expressed as follows:

[0075]

[0076] Where λ1 represents the weight of the self-expression loss component.

[0077] In step 4, the second self-expression loss function is expressed as follows:

[0078]

[0079] Among them, L exp2 λ represents the second self-expression loss function; λ2 represents the weight of the self-expression loss component.

[0080] In step 5, the image similarity matrix and the image ensemble similarity matrix are updated based on the self-expression coefficient matrix, as follows:

[0081] The image similarity matrix W is represented as follows:

[0082]

[0083] Image ensemble similarity matrix F t It is expressed as follows:

[0084]

[0085] Among them, F t F represents the ensemble similarity matrix of the images generated at the current time. t-1 W represents the image ensemble similarity matrix from the previous time step. t This represents the image similarity matrix generated at the current time.

[0086] Sequential ensemble methods generate base learners sequentially. For example, adaptive boosting in supervised learning trains the next base learner by boosting the weights of misclassified samples from the current base learner, repeatedly generating multiple base learners before weighted fusion. This invention focuses on clustering methods, therefore adaptive boosting cannot be directly used to generate base learners. Ensuring the accuracy and diversity of base learners is crucial in unsupervised sequential ensemble methods. This invention utilizes self-representation constraints in deep subspace clustering networks to guarantee the quality of the self-representation coefficient matrix. It also generates an unstable self-representation layer by iteratively updating the autoencoder network and the self-representation layer, thus generating a base learner set that combines diversity and accuracy. Finally, a sequential ensemble method is used to improve overall performance.

[0087] See Figure 2 As shown, an image clustering system based on deep subspace clustering sequential ensemble includes a base learner generation module and an ensemble module. The core idea is to train the parameters of a deep autoencoder network and a self-expression layer network using an alternating iterative update method, thereby successively generating different self-expression coefficient matrices. The ensemble module then fuses these self-expression coefficient matrices one by one to generate a more robust self-expression coefficient matrix, thus improving the final clustering performance. Specifically, the optimization objective of the subspace clustering network based on self-expression characteristics ensures the accuracy of the self-expression coefficient matrix, while the random masking module and the alternating iterative update method provide diversity to the self-expression coefficient matrix.

[0088] The base learner generation module includes a feature compression module, a feature reconstruction module, a random masking module, and a self-representation module. The feature compression module and the feature reconstruction module are used to learn visual feature representations with rich information, and both modules are composed of multiple convolutional layers.

[0089] Specifically, based on the input image set X, the visual feature compression vector Z of each image is extracted through a two-dimensional convolution operation f(·);

[0090]

[0091] After obtaining the compressed visual feature vector Z of the image, a random occlusion module is first used to erase some features in the visual feature vector Z:

[0092] Z′=Z⊙M (2)

[0093] Where M represents the random masking matrix; ⊙ represents the Hadamard product.

[0094] Then, the randomly occluded visual feature vector Z′ is input into a self-expression layer, that is, the self-expression visual feature vector Z′C is obtained through the mutual expression between features (by multiplying the randomly occluded visual feature vector Z′ with the self-expression coefficient matrix C).

[0095] Finally, the obtained self-expressive visual feature vector Z′C is processed through a two-dimensional deconvolution operation. Reconstruct the input image

[0096]

[0097] Furthermore, the base learner generation module can successively generate the self-expression coefficient matrix C, thereby generating the image similarity matrix. The function of the integration module is to successively merge the generated image similarity matrices, and the image integration similarity matrix F generated in each iteration is... t The similarity matrix F is integrated from the images at the previous time step. t-1 Adding this to the image similarity matrix W generated at the current time yields:

[0098]

[0099] The steps for constructing the loss function for network optimization are as follows:

[0100] (1) First self-expression loss function

[0101] First self-expression loss function L exp1 The goal is to enable the encoder network to learn visual feature vectors suitable for subspace clustering. By minimizing the error between the randomly occluded visual feature vector Z′ and the self-represented visual feature vector Z′C, the encoder model is encouraged to learn visual feature vectors more suitable for subspace clustering. Furthermore, a regularization loss ‖C‖ is added to this approach. pThis encourages the model to learn a sparser self-expression matrix. The formula for the first self-expression loss function is as follows:

[0102]

[0103] Where λ is the weighting coefficient of the error between the visual feature vector Z′ and the self-expressed visual feature vector Z′C.

[0104] (2) Image reconstruction loss function

[0105] The goal of the image reconstruction loss function L0 is to encourage the encoder network to learn more discriminative visual feature vectors. These feature vectors should contain most of the information from the original image; therefore, the image obtained by reconstructing the visual feature vectors should be as similar to the original image as possible. Based on this, the image reconstruction loss function is constructed using the error between the original image and the reconstructed image, as shown in the following formula:

[0106]

[0107] The model training process of the image clustering system based on deep subspace clustering order integration is as follows.

[0108] To obtain a more effective base learner, this paper uses an alternating iterative training method, which can be divided into three parts:

[0109] 1) Pre-training.

[0110] First, the feature compression module and feature reconstruction module without a self-expression layer are pre-trained using a given image set. The initial image encoder network is obtained by minimizing equation (6). and image decoder network

[0111] Secondly, a deep subspace clustering network is pre-trained, and the loss functions of formulas (5) and (6) are jointly optimized to provide the ensemble module with a good initial self-representation matrix. During the pre-training phase, the random masking module does not function, i.e., it does not perform the erasure operation.

[0112] 2) Update the weights of the depthwise convolutional autoencoder.

[0113] To learn an effective self-expression coefficient matrix, making the similarity matrix between images more accurate, a good visual feature vector Z′ after random occlusion is first needed. The network is then encouraged to learn how to generate a better visual feature vector Z′ by jointly optimizing the image reconstruction loss and the self-expression loss.

[0114]

[0115] Where λ1 is the weight of the self-expression loss component; L cae This represents the joint loss function for image reconstruction and self-expression.

[0116] 3) Update the weights of the self-expression layer.

[0117] To ensure the accuracy and diversity of the learned self-expression coefficient matrix, the error between the random occlusion visual feature vector Z′ and the self-expression visual feature vector Z′C, as well as the regularization loss of the self-expression coefficient matrix ‖C‖, are minimized. p :

[0118]

[0119] Where λ2 is the weight of the self-expression loss component; L exp2 This represents the second self-expression loss function.

[0120] The accuracy of the self-expression coefficient matrix is ​​ensured by minimizing formula (8), and then a more accurate image integration similarity matrix F is obtained by formula (4).

[0121] The final image integration similarity matrix F is obtained by alternately executing steps 2) and 3). t .

[0122] The final image is integrated into a similarity matrix F. t As input to the spectral clustering algorithm, the final image clustering result is obtained.

[0123] Specifically, the model algorithm implementation is shown in Table 1 below.

[0124] Table 1

[0125]

[0126] The following comparison will be conducted from five aspects: experimental data, comparison methods, evaluation indicators, experimental design, and experimental results.

[0127] (1) Description of the dataset (input image data)

[0128] See Figure 3 As shown, some examples of the four image datasets used in the experiment are listed, including ORL, COIL20, Extended Yale B, and COIL100.

[0129] As shown in Figure (a), the ORL dataset consists of 10 photos taken of each of 40 volunteers under normal lighting conditions, but with different facial angles and expressions. During the experiment, the original images, with a length of 112 and a width of 92, were uniformly downsampled to a length of 32 and a width of 32.

[0130] As shown in Figure (c), the Extended Yale B dataset consists of 38 volunteers with the same facial expressions and angles, but each volunteer was photographed 64 times under 64 different lighting intensities (i.e., one photo was taken for each lighting intensity). In the experiment, the original images with a length of 192 and a width of 168 were uniformly downsampled to a length of 48 and a width of 42.

[0131] As shown in Figures (b) and (d), the COIL20 and COIL100 datasets are both toy image datasets, representing 20 and 100 different toys, respectively, such as lucky cats, cups, and other toys. The datasets are created by photographing toys at different rotation angles, with 72 images taken for each toy. In the experiments, the original length and width were uniformly downsampled from 64 to 32. Detailed information about the datasets is summarized in Table 2.

[0132] Table 2 Basic Information of the Dataset

[0133] Dataset Feature number Sample size Number of categories ORL 32×32 400 40 EYaleB 48×42 2432 38 COIL20 32×32 1440 20 COIL100 32×32 7200 100

[0134] (2) Comparison Method

[0135] The comparison methods are mainly divided into two categories: (1) Traditional subspace clustering methods without incorporating deep network frameworks: including low-rank representation (LRR), low-rank subspace clustering (LRSC), sparse subspace clustering (SSC), kernel sparse subspace clustering (KSSC), sparse subspace clustering + orthogonal matching (SSC-OMP) and efficient dense subspace clustering (EDSC); (2) Subspace clustering methods incorporating deep networks: pre-trained convolutional autoencoder network + sparse subspace clustering (AE+SSC), pre-trained convolutional autoencoder network + efficient dense subspace clustering (AE+EDSC) and deep subspace clustering network (DSC-Nets).

[0136] (3) Evaluation indicators

[0137] For the final ensemble clustering results, accuracy (acc) was used as the evaluation metric to assess the algorithm's performance. Accuracy is generally used to represent the probability that a sample is correctly classified by the algorithm, and can be specifically expressed as:

[0138]

[0139] Error rate (Error, err) can be expressed as:

[0140] err=1-acc (10)

[0141] Where δ represents the indicator function (i.e., 1 for the same, 0 for different), x i The predicted label is l i Its actual label is y i map represents a transformation function (that is, it can transform the predicted label l) i Mapping to a category maximizes the similarity between the predicted result and the true label.

[0142] (4) Experimental Design

[0143] Table 3 Environmental Configuration

[0144] project Configuration CPU INTEL i7-7800X RAM 96.0GB GPU 2*NVIDIA GTX 1080Ti (2*11GB GRAM) OS Win10

[0145] The model building, training, and algorithm performance evaluation involved in the experiment were all implemented in a Win10 system environment, as detailed in Table 3. The network model was built and optimized using the Tensorflow framework, and the implementation code was written in Python.

[0146] In the experimental design, convolutional autoencoder networks were used for all four image datasets because convolutional layers have excellent feature extraction capabilities for image data. Furthermore, the network framework design maintained the same deep convolutional autoencoder structure as the Deep Subspace Clustering Network (DSC-Nets) to ensure the rationality of the experimental design. The network framework parameter settings used on each dataset are shown in Table 4, specifically including the self-expression coefficient matrix C and the convolutional layer representation "kernel size # number of channels" in the autoencoder network. The value of λ in this invention is consistent with that of DSC-Nets, and the hyperparameters set for other parameters on each dataset are shown in Table 5.

[0147] Table 4 Network Framework Parameter Settings

[0148]

[0149] Table 5 shows the hyperparameter settings for each dataset.

[0150]

[0151]

[0152] (5) Experimental Results

[0153] This invention conducted experiments on four publicly available image datasets and compared them with other clustering methods to verify the effectiveness and robustness of SeqEn-DSC for image clustering. The experiments were divided into two parts. Experiment 1 was used to verify the superiority of the proposed SeqEn-DSC over other subspace clustering models. The comparison methods included Low Rank Representation (LRR), Low Rank Subspace Clustering (LRSC), Sparse Subspace Clustering (SSC), Kernel Sparse Subspace Clustering (KSSC), Sparse Subspace Clustering + Orthogonal Matching (SSC-OMP), and Efficient Dense Subspace Clustering (EDSC); (2) Subspace clustering methods that integrate deep networks: Pre-trained convolutional autoencoder network + Sparse Subspace Clustering (AE+SSC), Pre-trained convolutional autoencoder network + Efficient Dense Subspace Clustering (AE+EDSC), and Deep Subspace Clustering Network (DSC-Nets). Experiment 2 was used to verify that the SeqEn-DSC model has higher performance and lower variability than DSC-Nets under different network structures.

[0154] (a) Experiment 1: Image clustering performance of SeqEn-DSC compared to other subspace clustering models

[0155] The experiment was conducted on four image datasets to verify that SeqEn-DSC can achieve better image clustering results than other subspace clustering models on different image datasets.

[0156] (b) Experiment 2: Clustering performance of SeqEn-DSC model compared to DSC-Nets model under different network structures

[0157] SeqEn-DSC is an ensemble model, thus exhibiting higher robustness. Experiment 2 aims to verify the robustness of the SeqEn-DSC model to changes in network structure. This experiment used the ORL face image dataset and performed clustering using SeqEn-DSC and DSC-Nets models with different encoder-decoder structures. As shown in Tables 6 and 7, the clustering performance of the DSC-Nets model fluctuated drastically with changes in network structure, but the method of this invention still maintained high performance.

[0158] Table 6. Clustering error rates (%) on ORL, EYaleB, COIL20, and COIL100.

[0159]

[0160]

[0161] Table 7. Clustering accuracy (%) for different network structures on ORL

[0162] DSC-Nets SeqEn-DSC 3-3-3～3-3-3 80.00 86.50 5-3-3～3-3-5 78.75 87.00 3-5-3～3-5-3 81.75 86.00 3-3-5～5-3-3 86.00 88.75 5-5-3～3-5-5 80.25 83.50 5-3-5～5-3-5 84.25 88.50 3-5-5～5-5-3 81.25 84.75 5-5-5～5-5-5 50.25 83.50 mean ± standard deviation 77.81±11.38 86.06±2.04

[0163] This invention addresses the performance issues of deep subspace clustering networks by employing the concept of sequential ensemble, proposing a sequential ensemble (SeqEn-DSC) for image-oriented deep subspace clustering. Specifically, SeqEn-DSC utilizes an alternating iterative training method and a random masking module to generate a highly dissimilar self-expression coefficient matrix, and leverages the self-expression properties to ensure the accuracy of the self-expression coefficient matrix, thereby improving the final ensemble result.

[0164] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. An image clustering method based on deep subspace clustering sequence ensemble, characterized in that, Includes the following steps: Step 1: Construct an image reconstruction loss function based on the input image data and the reconstructed image data; obtain the initial image encoder network and image decoder network by minimizing the image reconstruction loss function. Step 2: Construct a first self-expression loss function based on the error between the random occlusion visual feature vector and the self-expression visual feature vector, and obtain an initial self-expression coefficient matrix by minimizing the first self-expression loss function; wherein, the random occlusion visual feature vector is obtained based on the input image, the image encoder network, and the random occlusion matrix; Step 3: Fix the self-expression coefficient matrix, and update the image encoder network and image decoder network based on the joint loss function of image reconstruction and self-expression; repeat this step until the first iteration is reached; Step 4: Fix the image encoder network and image decoder network, and update the self-expression coefficient matrix based on the second self-expression loss function; repeat this step until the second iteration number is reached; Step 5: Update the image similarity matrix and the image ensemble similarity matrix based on the self-expression coefficient matrix; Step 6: Repeat steps 3 to 5 until the third iteration is reached; Step 7: Use the final image integration similarity matrix as input to the spectral clustering algorithm to obtain the image clustering results; Reconstructing image data Visual feature vectors for self-expression The reconstructed image obtained after deconvolution is as follows: ； It is expressed as follows: ； in, Represents the image decoder network; Represents the self-expression coefficient matrix; This represents the compressed vector of visual features extracted from the input image data; Represents the visual feature vector after random occlusion; Represents a random masking matrix; It represents the Hadamardi (or Hadama) stack; Z represents the following: ； in, Represents an image encoder network; In step 5, the image similarity matrix and the image ensemble similarity matrix are updated based on the self-expression coefficient matrix, as follows: Image similarity matrix It is expressed as follows: ； Image Integrated Similarity Matrix It is expressed as follows: ； Among them, F t This represents the integrated similarity matrix of the images generated at the current time. This represents the image ensemble similarity matrix from the previous time step. This represents the image similarity matrix generated at the current time.

2. The image clustering method based on deep subspace clustering sequence ensemble according to claim 1, characterized in that, The reconstruction loss function is expressed as follows: ； in, Represents the reconstruction loss function; This represents the input image data; This represents the reconstructed image data; It represents the square of the F-norm of the matrix.

3. The image clustering method based on deep subspace clustering sequence ensemble according to claim 1, characterized in that, The first self-expression loss function is expressed as follows: ； in, This represents the first self-expression loss function; Indicates the regularization loss; Represents visual feature vectors Visual feature vectors of self-expression The weighting coefficients of the errors between them.

4. The image clustering method based on deep subspace clustering sequence ensemble according to claim 3, characterized in that, In step 3, the joint loss function for image reconstruction and self-representation is expressed as follows: ； in, This represents the weight of the portion of self-expression loss.

5. The image clustering method based on deep subspace clustering sequence ensemble according to claim 4, characterized in that, In step 4, the second self-expression loss function is expressed as follows: ； in, This represents the second self-expression loss function; This represents the weight of the portion of self-expression loss.

6. An image clustering system based on deep subspace clustering sequence ensemble, characterized in that, The method based on any one of claims 1 to 5 includes: a base learner generation module and an integration module; The base learner generation module includes a feature compression module, a feature reconstruction module, a random occlusion module, and a self-representation module. The feature compression module and the feature reconstruction module are used to learn visual feature representations, and both modules are composed of multiple convolutional network layers. The base learner generation module sequentially generates the self-representation coefficient matrix. This leads to the generation of an image similarity matrix. ; The integration module is used to successively merge the generated image similarity matrices, and the image integration similarity matrix F generated in each iteration is... t The similarity matrix F is integrated from the images at the previous time step. t-1 The result is obtained by adding the image similarity matrix W generated at the current time, as follows: 。

Images