Hyperspectral image contrast clustering method and system based on uncertainty perception learning

Abstract

The invention discloses a hyperspectral image contrast clustering method and system based on uncertainty perception learning, and relates to the field of remote sensing image processing, the hyperspectral image contrast clustering method based on uncertainty perception learning mainly comprises the following steps: using an entropy rate superpixel segmentation method to obtain a high correlation sample set; constructing an uncertain sensing network, training the uncertain sensing network and a reverse network based on a high-correlation sample set, and obtaining a potential feature H; constructing a comparative clustering network, and training the comparative clustering network by using the high-correlation sample set and the potential features H to obtain a trained comparative clustering network; and clustering the hyperspectral image to be processed by using the trained contrast clustering network to obtain a clustering result. By implementing the hyperspectral image contrast clustering method and system based on uncertainty perception learning provided by the invention, the accuracy of hyperspectral image clustering can be improved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image processing, and more specifically, to a hyperspectral image contrast clustering method and system based on uncertainty-aware learning. Background Art

[0002] Since hyperspectral images are quite different from traditional images, especially when the data distribution is relatively complex, the reliability of self-supervised labels constructed through proxy tasks is still questionable. For example, it is difficult to ensure that the samples in the same superpixel sub-region are completely pure during the process of superpixel segmentation. Directly using such self-supervised labels to guide the clustering task usually may result in a decrease in clustering accuracy.

[0003] Since the clustering task has no label guidance, there is no better solution to this situation. The standard contrast learning framework only uses different feature augmentations from the same sample as positive sample pairs to avoid this problem and enables the encoder to learn this pattern. However, although this approach avoids the generation of wrong labels, it also ignores the association between samples and cannot reasonably utilize the rich spatial information in hyperspectral images, making the learning process always at the instance level and unable to solve the dilemmas faced in the hyperspectral clustering process. Therefore, how to measure the credibility of the labels learned from the data during the self-supervised learning process is the key to solving this problem.

[0004] Uncertainty learning aims to address uncertainties in data and models, including problems such as noise, sample bias, and environmental changes. In the field of deep learning, uncertainty mainly studies how to quantify and handle the uncertainty of deep learning models during prediction. According to the source, it can be divided into two types of uncertainties: model uncertainty and data uncertainty. Model uncertainty stems from the uncertainty of model parameters, such as overfitting and insufficient training data. Data uncertainty comes from the noise and unpredictability of the data itself, such as label errors, measurement errors, or the inherent randomness of the data. Common methods for solving data uncertainty include data augmentation, outlier detection and filtering, adding noise perturbations using adversarial training, and using variational autoencoders, etc. However, these methods usually directly target the original data and some methods require the support of sample labels for learning, and cannot meet the proposed requirements.

[0005] The above content is only used to assist in understanding the technical solution of the present invention, and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The purpose of the present invention is to provide a hyperspectral image contrast clustering method and system based on uncertainty-aware learning, which can improve the accuracy of hyperspectral image clustering.

[0007] The present invention provides a hyperspectral image contrast clustering method based on uncertainty-aware learning, comprising the following steps:

[0008] S1: Using the entropy rate superpixel segmentation method to obtain a highly correlated sample set;

[0009] S2: Constructing a contrast clustering network, and training the contrast clustering network using the highly correlated sample set to obtain a trained contrast clustering network;

[0010] S3: Using the trained contrast clustering network to cluster the hyperspectral image to be processed to obtain a clustering result.

[0011] Further, the contrast clustering network includes an uncertainty-aware feature fusion network and a siamese contrast clustering network; the uncertainty-aware feature fusion network is used to take the superpixel segmentation result as a prior condition, combine highly correlated samples to provide a learning target for each pixel, remove the noise samples in the self-supervised labels provided by the superpixel segmentation method, and generate a feature representation helpful for the clustering task under unsupervised conditions; the siamese contrast clustering network is used to learn a representation helpful for clustering and output the cluster probability distribution of the samples through a projection head.

[0012] Further, the uncertainty-aware feature fusion network learns to reconstruct the feature representations of multiple highly correlated samples to different extents, as shown in the formula:

[0013]

[0014]

[0015] where represents certainty, is the variance used to express the uncertainty of the nth highly correlated sample; is the highly correlated sample of the ith sample; represents the feature of the ith sample in the highly correlated sample set; ∈ represents a hyperparameter used to control the fault tolerance or uncertainty of the model; represents the given feature h i After that, the sample belongs to the probability of a certain category; h i represents the latent feature representation of the ith sample; represents the reconstructed output feature; represents the output of the uncertainty-aware network for calculating the reliability or uncertainty of the feature; represents the loss function of a single reconstruction target; represents the input data of the ith sample; Denote the total loss function of the uncertainty-aware feature fusion network; M represents the total number of training samples; N represents the number of highly correlated samples.

[0016] Further, for the Siamese contrastive clustering network, the encoder is trained to extract spatial-spectral information suitable for the clustering task, as shown in the formula:

[0017]

[0018] where is the first objective function; C represents the number of clusters to be clustered; P pq represents the similarity between cluster p and cluster q; λ represents the balance coefficient; p, q represent the indices of the clusters; represents the contrastive loss in the clustering objective, calculating the similarity between the samples in cluster p and their positive samples; represents the sample y ·p and the prediction between the similarity measure; y ·p represents the probability distribution of the sample on cluster p; represents the probability distribution of the sample predicted by the model on cluster p; τ represents the temperature coefficient, used to adjust the sensitivity of the similarity between samples; s(y ·p , y ·q ) represents the similarity between the samples in cluster p and the samples in cluster q; y ·q represents the samples in cluster q; represents the samples in cluster p and the predicted samples in cluster q; represents the predicted samples in cluster q; represents the joint objective function; represents the other part of the contrastive loss in the objective, calculating the similarity between the samples in cluster p and their negative samples; H(Y) represents the entropy, usually used to measure the uncertainty of the clustering result or the chaos degree of the distribution; represents the final objective function of the Siamese contrastive clustering network, and α is the balance coefficient, used to adjust the weight of the cluster-level loss.

[0019] The present invention also provides a hyperspectral image contrastive clustering system based on uncertainty-aware learning, and the system includes the following modules:

[0020] Sample set acquisition module, configured to: use the entropy rate superpixel segmentation method to obtain a highly correlated sample set;

[0021] Contrastive clustering network construction and training module, configured to: construct a contrastive clustering network, and use the highly correlated sample set to train the contrastive clustering network to obtain a trained contrastive clustering network;

[0022] The hyperspectral image clustering module is configured to: cluster the hyperspectral image to be processed by using the trained contrast clustering network, and obtain a clustering result.

[0023] Further, the contrast clustering network includes an uncertainty-aware feature fusion network and a siamese contrast clustering network; the uncertainty-aware feature fusion network is used to use the superpixel segmentation result as a prior condition, combine highly correlated samples to provide a learning target for each pixel, remove the noise samples in the self-supervised labels provided by the superpixel segmentation method, and generate a feature representation helpful for the clustering task under unsupervised conditions; the siamese contrast clustering network is used to learn a representation helpful for clustering, and output the cluster probability distribution of the samples through a projection head.

[0024] Further, the uncertainty-aware feature fusion network learns a feature representation that can reconstruct the features of multiple highly correlated samples to different degrees, such as the formula:

[0025]

[0026]

[0027] where, represents certainty, is the variance used to express the uncertainty of the nth highly correlated sample; is the highly correlated sample of the ith sample; represents the feature of the ith sample in the highly correlated sample set; ∈ represents a hyperparameter used to control the fault tolerance or uncertainty of the model; represents the probability that the sample i belongs to a certain category after the given feature h ; h i represents the latent feature representation of the ith sample; represents the reconstructed output feature; represents the output of the uncertainty-aware network calculating the reliability or uncertainty of the feature; represents the loss function of a single reconstruction target; represents the input data of the ith sample; represents the total loss function of the uncertainty-aware feature fusion network; M represents the total number of training samples; N represents the number of highly correlated samples.

[0028] Further, the siamese contrast clustering network trains an encoder to extract spatial-spectral information applicable to the clustering task, such as the formula:

[0029]

[0030]

[0031] Among them, is the first objective function; C represents the number of clusters to be clustered; P pq represents the similarity between cluster p and cluster q; λ represents the balance coefficient; p, q represent the indices of the clusters; represents the contrast loss in the clustering objective, calculating the similarity between the samples in cluster p and their positive samples; represents the sample y ·p and the prediction between the similarity measure; y ·p represents the probability distribution of the sample on cluster p; represents the probability distribution of the sample predicted by the model on cluster p; τ represents the temperature coefficient, used to adjust the sensitivity of the similarity between samples; s(y ·p , y ·q ) represents the similarity between the samples in cluster p and the samples in cluster q; y ·q represents the samples in cluster q; represents the samples in cluster p and the samples predicted in cluster q; represents the samples predicted in cluster q; represents the joint objective function; represents another part of the contrast loss in the objective, calculating the similarity between the samples in cluster p and their negative samples; H(Y) represents entropy, usually used to measure the uncertainty of the clustering result or the chaos degree of the distribution; represents the final objective function of the siamese contrast clustering network, and α is the balance coefficient, used to adjust the weight of the cluster-level loss.

[0032] The present invention also provides a computer program product, including a computer program, which realizes the steps of the above-mentioned hyperspectral image contrast clustering method based on uncertainty-aware learning when executed by a processor.

[0033] Implementing the hyperspectral image contrast clustering method and system based on uncertainty-aware learning provided by the present invention has the following beneficial effects:

[0034] The present invention proposes a deep clustering framework for processing hyperspectral data noise from a generative perspective, which can evaluate the noise from complex hyperspectral data to reduce the influence of noise data on clustering; the present invention first introduces uncertainty learning into the self-supervised task to evaluate the self-supervised labels generated by the proxy task, providing support for the application and development of self-supervised learning; at the same time, this framework is an online clustering model and can also effectively handle large-scale and complex hyperspectral image clustering scenarios; the present invention is superior to the more popular advanced methods in the same field in various clustering evaluation indicators, and through testing the feature extraction ability of UFF, its performance is superior to the more popular feature extraction methods in the same field. Description of the Drawings

[0035] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. In the accompanying drawings:

[0036] Figure 1 is a flowchart of the hyperspectral image contrast clustering method based on uncertainty-aware learning provided by the present invention;

[0037] Figure 2 is the SSUC network structure diagram provided by the present invention;

[0038] Figure 3 is the uncertainty-aware feature fusion network structure diagram provided by the present invention;

[0039] Figure 4 is a schematic diagram of the principle of the uncertainty-aware feature fusion network provided by the present invention. Specific Embodiments

[0040] For a clearer understanding of the technical features, objectives, and effects of the present invention, the specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] Figure 1 shows a schematic diagram of the hyperspectral image contrast clustering method based on uncertainty-aware learning in this embodiment. In this embodiment, the hyperspectral image contrast clustering method based on uncertainty-aware learning includes the following steps:

[0042] S1: Using the entropy rate superpixel segmentation method, obtain a highly correlated sample set;

[0043] S2: Construct a contrast clustering network, and use the highly correlated sample set to train the contrast clustering network to obtain a trained contrast clustering network;

[0044] In an exemplary embodiment, the contrast clustering network includes an uncertainty-aware feature fusion network and a siamese contrast clustering network; the uncertainty-aware feature fusion network is used to take the superpixel segmentation result as a prior condition, combine highly correlated samples to provide a learning target for each pixel, remove the noise samples in the self-supervised labels provided by the superpixel segmentation method, and generate a feature representation helpful for the clustering task under unsupervised conditions; the siamese contrast clustering network is used to learn a representation helpful for clustering and output the cluster probability distribution of the samples through a projection head.

[0045] In an exemplary embodiment, step S2 specifically includes: S2-1: Construct an uncertainty-aware network, use the highly correlated sample set to train the uncertainty-aware network and the inverse network, and obtain the latent feature H; S2-2: Construct a contrast clustering network, and use the highly correlated sample set and the latent feature H to train the contrast clustering network to obtain a trained contrast clustering network;

[0046] In an exemplary embodiment, the uncertainty-aware feature fusion network learns to reconstruct the feature representations of multiple highly correlated samples to varying degrees, as shown in the formula:

[0047]

[0048] where, represents certainty, is the variance used to express the uncertainty of the nth highly correlated sample; is the highly correlated sample of the ith sample; represents the feature of the ith sample in the high-correlation sample set; ∈ represents a hyperparameter used to control the fault tolerance or uncertainty of the model; represents the given feature h i after which the sample belongs to a certain class; h i represents the latent feature representation of the ith sample; represents the reconstructed output feature; represents the output of the uncertainty-aware network calculating the reliability or uncertainty of the feature; represents the loss function of a single reconstruction target; represents the input data of the ith sample; represents the total loss function of the uncertainty-aware feature fusion network; M represents the total number of training samples; N represents the number of high-correlation samples;

[0049] In an exemplary embodiment, the siamese contrastive clustering network trains an encoder to extract spatial-spectral information suitable for the clustering task, as shown in the formula:

[0050]

[0051]

[0052] where, is the first objective function; C represents the number of clusters to be clustered; P pq represents the similarity between cluster p and cluster q; λ represents a balance coefficient; p, q represent the indices of the clusters; represents the contrastive loss in the clustering objective, calculating the similarity between the samples in cluster p and their positive samples; represents the sample y ·p and the prediction between; y ·p represents the probability distribution of the sample on cluster p; represents the probability distribution of the sample predicted by the model on cluster p; τ represents the temperature coefficient used to adjust the sensitivity of the similarity between samples; s(y·p , y ·q ) represents the similarity between samples in cluster p and samples in cluster q; y ·q represents the samples in cluster q; represents the samples in cluster p and the predicted samples in cluster q; represents the predicted samples in cluster q; represents the combined objective function; represents another part of the contrast loss in the objective, which calculates the similarity between samples in cluster p and their negative samples; H(Y) represents entropy, which is usually used to measure the uncertainty of the clustering result or the chaos degree of the distribution; represents the final objective function of the twin contrast clustering network, where α is a balance coefficient used to adjust the weight of the cluster-level loss;

[0053] S3: Use the trained contrast clustering network to cluster the hyperspectral image to be processed, and obtain a clustering result.

[0054] In some embodiments, the above-mentioned hyperspectral image contrast clustering method based on uncertainty-aware learning can also be implemented in the following manner.

[0055] In this embodiment, the hyperspectral image contrast clustering method based on uncertainty-aware learning includes: using the entropy rate superpixel segmentation method to obtain a highly correlated sample set; constructing a contrast clustering network, and using the highly correlated sample set to train the contrast clustering network to obtain a trained contrast clustering network; using the trained contrast clustering network to cluster the hyperspectral image to be processed, and obtain a clustering result.

[0056] In this embodiment, the contrast clustering network (Self-Supervised Contrastive Clustering, SSUC) model based on uncertainty-aware learning consists of two modules, which are the uncertainty-aware feature fusion network (Uncertainty-Aware Feature Fusion, UFF) and the twin contrast clustering network (Twin Contrastive Clustering, TCC). UFF takes the superpixel segmentation result as a prior condition and combines highly correlated samples to provide a learning target for each pixel. Its unique uncertainty-aware mechanism can effectively remove the noise samples in the self-supervised labels provided by the superpixel segmentation method and generate feature representations helpful for the clustering task under unsupervised conditions. The TCC module is the backbone of SSUC, and this module aims to train an encoder e ψ(·) Learn a representation that helps clustering. TCC enforces the encoder features to be consistent with the features generated by UFF and outputs the cluster probability distribution of the samples through the projection head. The model adopts an alternating training mode, and the feature vector H generated by UFF is always trainable. Its network structure is as Figure 2 shown. The pseudo-code of SSUC is shown in Algorithm 1 in Table 1.

[0057] Table 1: Pseudo-code table of SUCC

[0058]

[0059]

[0060] The main task of Uncertainty-Aware Feature Fusion (UFF) is to evaluate each highly correlated sample separately and merge them together. This process can be divided into two parts. First, UFF needs to search for a set of highly correlated samples for each sample in the dataset. Generally, pixels belonging to the same object usually come from the same cluster. Based on the spatial-spectral characteristics of hyperspectral images, the superpixel segmentation method performs well in dividing homogeneous regions. Its designed architecture is as Figure 3 shown.

[0061] UFF uses Entropy Rate Superpixel Segmentation (ERS) to generate S sub-regions for the entire dataset. Then, based on the superpixel labels given by ERS and the Gaussian distance, N highly correlated samples are selected for each sample. For a sample x i , there exists a set of highly correlated samples S i = {s i 1 ,..., s i N+1}. Then, for the entire dataset, there exists S = {s i 1 ,..., s i N+1} M i＝1 .

[0062] The second part is the fusion that includes uncertainty awareness. The highly correlated samples are used to assist in representing the target sample and finally generate the feature vector. UFF uses a reverse network similar to the one in to learn the feature representation. As Figure 4 shown, the reverse network is similar to a decoder. It consists of trainable features H = {h1,..., h M} and multiple sub-networks f() = {f1(),..., f N+1()} It consists of... (Different from traditional autoencoders, the inverse network reconstructs samples by treating the hidden layer representation as a condition.)

[0063] The goal of UFF is to learn a feature representation that can reconstruct N highly correlated samples to different degrees. Let s (n) i be the highly correlated sample of the i-th sample. To dynamically evaluate each highly correlated sample, UFF assumes that highly correlated samples are sampled from different Gaussian distributions, i.e., s i (v) ~N(μ i (n) ,(σ i (n) ) 2 ). Then we can get:

[0064]

[0065] where μ i (n) represents determinacy, and σ i (n) is the variance used to express the uncertainty of the n-th highly correlated sample. UFF hopes to learn a latent representation H that can dynamically reconstruct all highly correlated samples. Based on Bayes' theorem, the joint distribution of h i and multiple highly correlated samples s i (n) can be decomposed into a prior distribution p (hi) and a likelihood p(s i (n) |h i ).

[0066] Then we can get:

[0067]

[0068] The likelihood describes the probability of observing s i under the condition h i (n) . For unsupervised tasks, it is difficult to obtain prior knowledge p(h i ). Therefore, UFF mainly focuses on the likelihood p(s i (1) ,···,s i (n) |h i ). By assuming that the observations of each highly correlated sample are conditionally independent under the condition h i , this likelihood estimate can be expressed as the product of the views of each highly correlated sample p(s i (n) |h i ), that is:

[0069]

[0070] For each highly relevant sample observation s i (n) , the corresponding latent feature h i and a network f θn (·) can be used to reconstruct it. Similarly, in order to specifically evaluate the uncertainty of each highly relevant sample, UFF also uses a neural network to fit the variance. According to the above assumptions, under the condition of h i , the Gaussian distribution probability of s i n can be expressed as:

[0071]

[0072] where θ n and are the trainable parameters of the network. Essentially, the network evaluates the confidence of each highly relevant sample by learning the variance as an indicator. The advantage of doing this is that the network can provide different uncertainty evaluations for different samples instead of outputting a definite value. Considering the Gaussian distribution, UFF takes s i (n) as the reconstruction target, and its likelihood is:

[0073]

[0074] The loss function of a single reconstruction target is to minimize:

[0075]

[0076] where μ i (n) = f θn (h i ), According to common sense, it is difficult to reconstruct noisy samples with the same latent representation during the training process. If so, the corresponding reconstruction loss and variance will increase. This means that when the variance increases, the uncertainty of the observation will also be higher. The second term of the formula is a regularization term used to limit the excessive growth of uncertainty. (In this process, the constant term is removed from the formula to simplify the function.) In addition, UFF uses an inverse network so that the features of each sample can be encoded in H to different degrees. Finally, in order to calculate the loss of all samples, UFF sums up the losses of each sample to obtain the total loss function as follows:

[0077]

[0078] During this process, UFF uses N + 1 networks to obtain the reconstruction loss of all samples. This method can be regarded as using multiple highly correlated samples to obtain the latent representation of a single sample. In addition, the weights of each strongly correlated sample are dynamically adjusted according to its uncertainty. Therefore, during the process of learning the latent representation, UFF can make full use of high-quality correlated samples while reducing the influence of noisy samples brought by superpixel self-supervised labels.

[0079] Siamese Contrastive Clustering Network: Due to the influence of "false" negative sample pairs, the traditional instance-level contrastive loss (InfoNCE) is affected in clustering tasks. Therefore, SSUC improves the contrastive clustering framework accordingly. In the Siamese Contrastive Clustering Network, TCC attempts to train an encoder to extract spatio-spectral information suitable for clustering tasks. Specifically, TCC hopes that the information extracted by the encoder is consistent with the features generated by the UFF module. To some extent, this is similar to the knowledge distillation process. TCC uses a dual-network architecture composed of an encoder e ψ (·) and a projection head g φ (·). Given a mini-batch of sample set O = {x k} k＝1 B (B is the batch size) and its uncertainty-aware fusion feature set H = {hk} k＝1 B , the encoder e ψ (·) takes x k as input, and its output result is fed into g φ (·). Then through TCC, y k = g φ (e ψ (x k )) can be obtained, where the projection head g φ (·) outputs y ∈ R C to predict the corresponding label representation.

[0080] To prevent the encoder from only learning fixed patterns, TCC randomly selects a sample s k from the strongly correlated sample set and fuses it with h k . Similarly, e ψ (·) and g φ (·) also participate in the fusion process to obtain ^y k ∈ R C , and this process can be expressed as ^y k = g φ (e ψ (s k ) + h k ). For a batch O, there is Y = {y k} k＝1B , ^Y = {^y k} k＝1 B and their matrix forms Y ∈ R B×C , ^Y ∈ R B×C .

[0081] The objective function of TCC consists of two parts: L BAR and L CLU . To avoid the interference caused by "false" negative sample pairs, TCC adopts the Barlow Twins loss function L BAR that only considers positive sample pairs and can prevent trivial solutions. In the ideal case, Y and ^Y should be the same, and each column y ·p and ^y ·p can be regarded as the marginal probability distribution of batch samples on the p-th cluster. TCC calculates the cosine similarity along the row dimension to generate a cross-correlation matrix P ∈ R C×C . For each element in P, there is:

[0082]

[0083] To keep Y and ^Y consistent, P should be a diagonal matrix. Therefore, the objective function of L BAR is:

[0084]

[0085] The second term of the loss is used to perform the disassociation operation between different clusters, where λ represents the balance coefficient. For the second objective function L CLU , TCC adopts the cluster-level contrast loss in contrastive clustering. The InfoNCE loss is calculated on the column dimension of Y and ^Y. For a column y ·p in Y, the cluster-level contrast objective function can be expressed as:

[0086]

[0087] For all columns in Y and ^Y, the joint objective function is:

[0088]

[0089] The final objective function of TCC is defined as the sum of LBAR and LCLU:

[0090]

[0091] where α is a balance coefficient used to adjust the weight of the cluster-level loss.

[0092] This embodiment provides a hyperspectral image contrast clustering system based on uncertainty-aware learning. The system includes the following modules:

[0093] A sample set acquisition module, configured to: use the entropy rate superpixel segmentation method to obtain a highly correlated sample set;

[0094] A contrast clustering network construction and training module, configured to: construct a contrast clustering network, and use the highly correlated sample set to train the contrast clustering network to obtain a trained contrast clustering network;

[0095] A hyperspectral image clustering module, configured to: use the trained contrast clustering network to cluster the hyperspectral image to be processed to obtain a clustering result.

[0096] In an exemplary embodiment, the hyperspectral image contrast clustering system based on uncertainty-aware learning includes the following modules: A sample set acquisition module, configured to: use the entropy rate superpixel segmentation method to obtain a highly correlated sample set;

[0097] An uncertainty-aware network construction and training module, configured to: construct an uncertainty-aware network, use the highly correlated sample set to train the uncertainty-aware network and the inverse network to obtain the latent feature H; A contrast clustering network construction and training module, configured to: construct a contrast clustering network, and use the highly correlated sample set and the latent feature H to train the contrast clustering network to obtain a trained contrast clustering network; A hyperspectral image clustering module, configured to: use the trained contrast clustering network to cluster the hyperspectral image to be processed to obtain a clustering result.

[0098] In an exemplary embodiment, the contrast clustering network includes an uncertainty-aware feature fusion network and a siamese contrast clustering network; the uncertainty-aware feature fusion network is used to use the superpixel segmentation result as a prior condition, combine highly correlated samples to provide a learning target for each pixel, remove the noise samples in the self-supervised labels provided by the superpixel segmentation method, and generate a feature representation helpful for the clustering task under unsupervised conditions; the siamese contrast clustering network is used to learn a representation helpful for clustering and output the cluster probability distribution of the samples through a projection head.

[0099] In an exemplary embodiment, the uncertainty-aware feature fusion network learns to reconstruct the feature representations of multiple highly correlated samples to different degrees, such as the formula:

[0100]

[0101]

[0102] Where represents certainty The variance is used to express the uncertainty of the nth highly correlated sample; is the highly correlated sample of the ith sample; represents the features of the ith sample in the highly correlated sample set; ∈ represents a hyperparameter used to control the fault tolerance or uncertainty of the model; represents the given feature h i after which the sample belongs to a certain class; h i represents the latent feature representation of the ith sample; represents the output feature of the reconstruction; represents the output of the uncertainty-aware network for calculating the reliability or uncertainty of features; represents the loss function for a single reconstruction target; represents the input data of the ith sample; represents the total loss function of the uncertainty-aware feature fusion network; M represents the total number of training samples; N represents the number of highly correlated samples.

[0103] In an exemplary embodiment, the siamese contrastive clustering network trains an encoder to extract spatio-spectral information suitable for the clustering task, as shown in the formula:

[0104]

[0105] where is the first objective function; C represents the number of clusters to be clustered; P pq represents the similarity between cluster p and cluster q; λ represents a balance coefficient; p, q represent the indices of the clusters; represents the contrastive loss in the clustering objective, calculating the similarity between the samples in cluster p and their positive samples; represents the sample y ·p and the prediction between; y ·p represents the probability distribution of the sample on cluster p; represents the probability distribution of the sample predicted by the model on cluster p; τ represents the temperature coefficient used to adjust the sensitivity of the similarity between samples; s(y ·p , y ·q ) represents the similarity between the samples in cluster p and the samples in cluster q; y ·q represents the sample in cluster q; represents the samples in cluster p and the predicted samples in cluster q; represents the predicted samples in cluster q; represents the joint objective function; Represents another part of the contrast loss in the target, which calculates the similarity between samples in cluster p and their negative samples; H(Y) represents entropy, which is usually used to measure the uncertainty of the clustering result or the chaos of the distribution; Represents the final objective function of the Siamese contrast clustering network. α is a balancing coefficient used to adjust the weight of the cluster-level loss.

[0106] This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described hyperspectral image contrast clustering method based on uncertainty-aware learning.

[0107] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the claims of the present invention. These all fall within the protection scope of the present invention.

Claims

1. A hyperspectral image contrast clustering method based on uncertainty-aware learning, characterized in that It includes the following steps: S1: Using the entropy rate superpixel segmentation method, obtain a highly correlated sample set; S2: Construct a contrast clustering network, and use the highly correlated sample set to train the contrast clustering network to obtain a trained contrast clustering network; S3: Use the trained contrast clustering network to cluster the hyperspectral image to be processed to obtain a clustering result.

2. The hyperspectral image contrast clustering method based on uncertainty-aware learning according to claim 1, wherein The contrast clustering network includes an uncertainty-aware feature fusion network and a siamese contrast clustering network; the uncertainty-aware feature fusion network is used to take the superpixel segmentation result as a prior condition, combine highly correlated samples to provide a learning target for each pixel, remove the noise samples in the self-supervised labels provided by the superpixel segmentation method, and generate a feature representation helpful for the clustering task under unsupervised conditions; the siamese contrast clustering network is used to learn the representation helpful for clustering and output the cluster probability distribution of the samples through a projection head.

3. The hyperspectral image contrast clustering method based on uncertainty-aware learning according to claim 2, wherein, The uncertainty-aware feature fusion network learns to reconstruct the feature representations of multiple highly correlated samples to different degrees, as shown in the formula: Among them, represents certainty, and the variance is used to express the uncertainty of the nth highly correlated sample; is the highly correlated sample of the ith sample; represents the feature of the ith sample in the highly correlated sample set; ∈ represents a hyperparameter used to control the fault tolerance or uncertainty of the model; represents the given feature h i after which the sample belongs to a certain category with probability; h i represents the latent feature representation of the ith sample; represents the output feature of the reconstruction; represents the output of the uncertainty-aware network for calculating the reliability or uncertainty of the feature; represents the loss function for a single reconstruction target; represents the input data of the ith sample; represents the total loss function of the uncertainty-aware feature fusion network; M represents the total number of training samples; N represents the number of highly correlated samples.

4. The hyperspectral image contrast clustering method based on uncertainty-aware learning according to claim 2, wherein The siamese contrast clustering network trains an encoder to extract the spatial-spectral information suitable for the clustering task, as shown in the formula: Among them, The first objective function is; C represents the number of clusters to be clustered; P pq represents the similarity between cluster p and cluster q; λ represents the balance coefficient; p, q represent the indices of the clusters; represents the contrast loss in the clustering objective, calculating the similarity between the samples in cluster p and their positive samples; represents the sample y ·p and the prediction between the similarity measure; y ·p represents the probability distribution of the sample on cluster p; represents the probability distribution of the sample predicted by the model on cluster p; τ represents the temperature coefficient, used to adjust the sensitivity of the similarity between samples; s(y ·p , y ·p ) represents the similarity between the samples in cluster p and the samples in cluster q; y ·q represents the samples in cluster q; represents the samples in cluster p and the samples in the predicted cluster q; represents the samples in the predicted cluster q; represents the joint objective function; represents another part of the contrast loss in the objective, calculating the similarity between the samples in cluster p and their negative samples; H(Y) represents entropy, usually used to measure the uncertainty of the clustering result or the chaos degree of the distribution; represents the final objective function of the siamese contrast clustering network, and α is the balance coefficient, used to adjust the weight of the cluster-level loss.

5. A hyperspectral image contrast clustering system based on uncertainty-aware learning, characterized in that, The system includes the following modules: A sample set acquisition module, configured to: using the entropy rate superpixel segmentation method, obtain a highly correlated sample set; A contrast clustering network construction and training module, configured to: construct a contrast clustering network, and use the highly correlated sample set to train the contrast clustering network to obtain a trained contrast clustering network; A hyperspectral image clustering module, configured to: use the trained contrast clustering network to cluster the hyperspectral image to be processed to obtain a clustering result.

6. The hyperspectral image contrast clustering system based on uncertainty-aware learning according to claim 5, characterized in that, The contrast clustering network includes an uncertainty-aware feature fusion network and a siamese contrast clustering network; the uncertainty-aware feature fusion network is used to take the superpixel segmentation result as a prior condition, combine highly correlated samples to provide a learning target for each pixel, remove the noise samples in the self-supervised labels provided by the superpixel segmentation method, and generate a feature representation helpful for the clustering task under unsupervised conditions; the siamese contrast clustering network is used to learn the representation helpful for clustering and output the cluster probability distribution of the samples through a projection head.

7. The hyperspectral image contrast clustering system based on uncertainty-aware learning according to claim 6, wherein The uncertainty-aware feature fusion network learns to reconstruct the feature representations of multiple highly correlated samples to different degrees, as shown in the formula: Among them, represents certainty, and the variance is used to express the uncertainty of the nth highly correlated sample; is the highly correlated sample of the ith sample; represents the feature of the ith sample in the highly correlated sample set; ∈ represents a hyperparameter used to control the fault tolerance or uncertainty of the model; represents the probability that the sample i belongs to a certain category after the given feature h ; h i represents the latent feature representation of the ith sample; represents the output feature of the reconstruction; represents the output of the uncertainty-aware network for calculating the reliability or uncertainty of the feature; represents the loss function of a single reconstruction target; represents the input data of the ith sample; represents the total loss function of the uncertainty-aware feature fusion network; M represents the total number of training samples; N represents the number of highly correlated samples.

8. The hyperspectral image contrast clustering system based on uncertainty-aware learning according to claim 6, characterized in that The siamese contrast clustering network trains an encoder to extract the spatial-spectral information suitable for the clustering task, as shown in the formula: Among them, The first objective function is; C represents the number of clusters to be clustered; P pq represents the similarity between cluster p and cluster q; λ represents the balance coefficient; p, q represent the indices of the clusters; represents the contrast loss in the clustering objective, calculating the similarity between the samples in cluster p and their positive samples; represents the sample y ·p and the prediction between the similarity measure; y ·p represents the probability distribution of the sample on cluster p; represents the probability distribution of the sample predicted by the model on cluster p; τ represents the temperature coefficient, used to adjust the sensitivity of the similarity between samples; s(y ·p , y ·q ) represents the similarity between the samples in cluster p and the samples in cluster q; y ·q represents the samples in cluster q; represents the samples in cluster p and the samples in the predicted cluster q; represents the samples in the predicted cluster q; represents the joint objective function; represents another part of the contrast loss in the objective, calculating the similarity between the samples in cluster p and their negative samples; H(Y) represents the entropy, usually used to measure the uncertainty of the clustering result or the chaos of the distribution; represents the final objective function of the siamese contrast clustering network, and α is the balance coefficient, used to adjust the weight of the cluster-level loss.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the hyperspectral image contrast clustering method based on uncertainty-aware learning according to any one of claims 1-4.

Images