Incomplete underwater hyperspectral image clustering method based on multilayer tensor
By recovering incomplete data from underwater hyperspectral images using a multilayer tensor method, the problems of missing data and noise interference in underwater images are solved, and high-precision unsupervised clustering is achieved.
Patent Information
- Application Number
- CN202511281601.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-09
- Publication Date
- 2025-12-23
AI Technical Summary
Existing technologies struggle to effectively process incomplete data in underwater hyperspectral images, leading to data loss and noise interference, which affects classification accuracy.
We employ a multi-level tensor-based approach, extracting latent consensus representations through orthogonal projection operators, and combining adaptive weights and multi-level tensor learning to recover high-dimensional spectral structures and perform unsupervised clustering.
In complex underwater environments, it can effectively recover missing information, improve the robustness and stability of clustering results, and enhance the modeling and generalization capabilities for complex multi-source information.
Smart Images
Figure CN121190797A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater image processing technology, and more specifically to an incomplete underwater hyperspectral image clustering method based on multilayer tensors. Background Technology
[0002] Underwater hyperspectral images are a special type of image acquired by underwater robots or platforms equipped with hyperspectral imaging equipment. Compared to ordinary images, they have more spectral channels and higher spectral resolution, enabling continuous imaging within a specific range. Therefore, they contain rich spectral information and spatial structural features of underwater targets. These images have wide applications in marine resource surveys, underwater ecological monitoring, aquaculture, and marine engineering.
[0003] Underwater hyperspectral image classification aims to accurately determine and categorize each pixel in an image based on acquired high-dimensional spectral features. Traditional classification methods typically rely on the differences in the spectral response of underwater targets across different regions. However, due to the strong absorption and scattering of light by water, underwater imaging quality is easily limited, leading to significant spectral degradation. To improve classification accuracy, various machine learning algorithms, such as support vector machines, sparse representation methods, and convolutional neural networks, have been introduced into this task in recent years. However, in practical applications, due to complex imaging conditions and environmental interference, hyperspectral data often becomes incomplete during acquisition and transmission, making it difficult to capture multi-layered data structures, achieve information exchange across multiple regions, and handle redundant information. Therefore, how to achieve underwater hyperspectral clustering of incomplete data and, in the process, reasonably recover missing spectral information to further improve overall classification performance has become a challenging and valuable research problem. Summary of the Invention
[0004] To address the aforementioned technical challenges of data loss in underwater environments due to poor visibility and sensor failure, this invention provides a clustering method for incomplete underwater hyperspectral images based on multi-level tensors. This invention first extracts latent consensus representations from incomplete intervals using orthogonal projection operators, then dynamically compensates for the credibility of missing views through adaptive weights. Subsequently, it learns high-order complementary relationships across intervals using multi-level tensor decomposition, and uniformly optimizes similarity graphs under nuclear norm constraints. This achieves robust recovery of high-dimensional spectral structure, mining of hierarchical features, and unsupervised clustering even under conditions of missing views, noise interference, and poor underwater visibility.
[0005] The technical means employed in this invention are as follows: A method for incomplete underwater hyperspectral image clustering based on multilayer tensors includes the following steps: A process is initiated by acquiring an image to be processed and obtaining incomplete spectral data of the image in different intervals. A clustering model based on spectral clustering learning is constructed, and the incomplete spectral data is processed based on this model. The clustering model includes: a projection operator information extraction module, a multi-level graph learning module, an adaptive weighting module, a multi-level tensor learning module, and a nuclear norm constraint module. The projection operator information extraction module processes the incomplete spectral data in different intervals using a projection matrix and an index matrix to obtain latent consensus representation matrices for different intervals. The multi-level graph learning module performs highest-level learning based on these latent consensus representation matrices to obtain the highest-level similarity. On the other hand, by learning the dynamic collaboration in the incomplete spectral data, the lowest and intermediate layer similarity maps are obtained. The adaptive weight module adaptively updates the weights based on the importance of the lowest and intermediate layer similarity maps to obtain updated lowest and intermediate layer similarity maps. The multi-level tensor learning module stacks the highest layer similarity map and the updated lowest and intermediate layer similarity maps into a tensor and rotates it to obtain a multi-level tensor. The nuclear norm constraint module constrains the rotated tensor to obtain a fused tensor. The hyperspectral clustering model is iterated based on an alternating variable update strategy until the iteration stopping condition is met, and the clustering result is output based on spectral clustering.
[0006] Furthermore, the workflow of the projection operator information extraction module is as follows: Incomplete spectral data in different intervals are processed using a projection matrix to obtain projected data; an index matrix is used to complete the accurate samples of the projected data to obtain a latent consensus representation matrix. The latent consensus representation matrix is calculated using the following formula:
[0007] in, The data is after projection processing. The matrix represents the potential consensus. For orthogonal projection operators, For the first v Incomplete spectral data for each interval, For index matrix, The index matrix, representing the highest-level first noise matrix, is as follows:
[0008] in, For elements in the index matrix, Let be the element in the incomplete spectral data of the v-th interval.
[0009] Furthermore, the multi-layer graph learning module includes a highest similarity graph learning unit and a dynamic cross-view learning unit. The highest similarity graph learning unit calculates the highest-level similarity graph according to the following formula:
[0010] in, The matrix represents the potential consensus. This is the highest-level similarity graph. The second noise matrix is the highest layer; the dynamic cross-view learning unit and the adaptive weight module calculate the similarity graphs of the lowest and middle layers according to the following formula:
[0011] in, For the first v Incomplete spectral data for each interval, For index matrix, This is the lowest level similarity graph. This is the lowest level noise matrix. This is an intermediate-level similarity graph. This is the second minimized noise matrix. For the first v The weight of incomplete spectral data in each interval.
[0012] Furthermore, the multi-level tensor module is configured as follows:
[0013] in, These are similarity diagrams representing the lowest, middle, and highest levels, respectively. To stack multi-level similar graphs into a tensor and rotate them, It is a multi-level tensor.
[0014] Furthermore, the underwater multi-view clustering model based on spectral clustering learning is set as follows:
[0015] in, For multi-level tensors, This is the first noise matrix at the highest level. This is the second noise matrix at the highest level. This is the lowest level noise matrix. This is the noise matrix of the intermediate layer. As the first penalty parameter, This is the second penalty parameter.
[0016] Furthermore, the clustering results output based on spectral clustering include: After reducing the unified similarity graph information learned from the multi-level tensor through spectral clustering, clustering is performed to obtain the classification result. The method for calculating the unified similarity graph is as follows: using the spectral clustering model to constrain the affinity graph in low rank, the unified similarity graph is obtained.
[0017] in, To unify similar graphs, The number of spectral intervals. These are similarity diagrams representing the lowest, middle, and highest levels, respectively. For the first A range.
[0018] Compared with the prior art, the present invention has the following advantages: 1. This invention recovers incomplete data between different multispectral datasets and introduces a multi-level structure-guided tensor to explore the multi-level structure of each interval, thus freeing the data representation from being limited to a planar structure. This multi-level structure representation not only improves the modeling ability for complex multi-source information but also enhances the robustness and stability of clustering results in the face of incomplete data, noise interference, and diverse scenarios.
[0019] 2. In the representation learning process for recovering missing instances, this invention introduces an orthogonal projection operator to effectively separate out interference components in the original high-dimensional feature space, retaining only the core information related to data distribution and potential subspaces. This not only improves the purity and discriminative power of the feature representation but also ensures the model's ability to capture useful signals even when faced with missing data or noise interference, thereby enhancing the stability and robustness of the clustering results.
[0020] 3. This invention introduces dynamic cross-view learning, which on the one hand maintains the consistency of different intervals in the shared structure, ensuring the uniformity of global semantics; on the other hand, it fully preserves the personalized features of each interval, avoiding the loss of important difference information. This mechanism effectively strengthens the interaction and complementarity between different views, enabling multi-source information to collaboratively model, thereby further improving the overall performance and generalization ability of clustering.
[0021] 4. In this invention, the introduction of an adaptive weighting strategy ensures that clustering is guided by the importance exhibited by each interval. The weight allocation of each interval is dynamically adjusted based on its contribution to the clustering task, giving greater attention to high-quality intervals and weakening low-quality intervals. This flexible weighting method not only improves the model's expressive power but also avoids clustering degradation caused by excessive noise in a single interval, ensuring that the final result better reflects the essential structure of the data.
[0022] 5. This invention constructs a multi-level tensor model to dynamically learn a unified similarity graph among all pixels. This not only effectively addresses the problems of missing views and incomplete observations, but also fully explores the complementarity between multiple views and the hierarchical structural features of underwater targets. It is suitable for unsupervised clustering tasks in complex underwater hyperspectral data.
[0023] Based on the above reasons, this invention can be widely applied in fields such as underwater image processing. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a flowchart of an incomplete underwater hyperspectral image clustering method based on multilayer tensors in an embodiment of the present invention.
[0026] Figure 2 This is a schematic diagram illustrating the process of using an incomplete underwater hyperspectral image clustering method based on multilayer tensors in an embodiment of the present invention. Detailed Implementation
[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0029] like Figure 1 and Figure 2 As shown, this invention provides a method for incomplete underwater hyperspectral image clustering based on multilayer tensors, the specific steps of which are as follows: S1. Obtain the image to be processed. The spectral range itself contains multiple ranges, and each range is regarded as a view. The number of ranges within a range is the feature number of that view. Obtain incomplete spectral data of the image to be processed in different ranges. .in, Indicates the first Data for each interval.
[0030] A spectral range itself contains multiple bands, and each range can be considered a view. The number of bands within a range is the feature number of that view. The band division criteria are determined by the wavelength intervals specified in the physical design of the imaging sensor.
[0031] S2. Construct a clustering model based on spectral clustering learning, and process incomplete spectral data based on the clustering model. The clustering model includes: a projection operator information extraction module, a multi-layer graph learning module, an adaptive weight module, a multi-level tensor learning module, and a nuclear norm constraint module.
[0032] The projection operator information extraction module processes incomplete spectral data in different intervals using projection and index matrices to obtain potential consensus representation matrices for different intervals.
[0033] Specifically, the workflow of the projection operator information extraction module is as follows: Incomplete spectral data from different intervals are processed using a projection matrix to obtain projected data. An index matrix is then used to complete the projected data with accurate samples, resulting in a latent consensus representation matrix. This latent consensus representation matrix is calculated using the following formula:
[0034] in, The data is after projection processing. Let k be the latent consensus representation matrix, k be the feature dimension of the latent consensus representation, and n be the number of data samples. For orthogonal projection operators, For the first v Incomplete spectral data for each interval, For index matrix, This is the highest-level, first-order noise matrix. Because the data contains noise and outliers, we extract [the noise matrix] from the spectrum. The noise matrix is modeled to minimize its impact by learning it through the alternating direction multiplier method. Indicates the first There are intervals, where the data dimension is... Observable One sample.
[0035] In this application, the purpose of setting up an index matrix is to store the locations of missing data. The index matrix is represented as follows:
[0036] in, For elements in the index matrix, Let be the element in the incomplete spectral data of the v-th interval.
[0037] The multi-layer graph learning module performs top-level learning based on the latent consensus representation matrix to obtain the top-level similar graph. On the other hand, it obtains the bottom-level and intermediate-level similar graphs by learning dynamic collaborations in incomplete spectral data.
[0038] The multi-layer graph learning module includes a highest similarity graph learning unit and a dynamic cross-view learning unit. The highest similarity graph unit calculates the highest-level similarity graph according to the following formula:
[0039] in, The matrix represents the potential consensus. This is the highest-level similarity graph. This is the second noise matrix at the highest level.
[0040] To fully explore the consistency and complementarity information across different views, dynamic cross-view learning is employed. To address the issue of quality differences between intervals, an adaptive weighting module is introduced. This module adaptively updates the weights based on the importance of the bottom-level and intermediate-level similarity maps, obtaining updated bottom-level and intermediate-level similarity maps. The dynamic cross-view learning unit and the adaptive weighting module calculate the bottom-level and intermediate-level similarity maps using the following formula:
[0041] in, For the first v Incomplete spectral data for each interval, For index matrix, This is the lowest level similarity graph. This is the lowest level noise matrix. This is an intermediate-level similarity graph. The purpose of the intermediate-level similarity graph is to fully explore the similarities and complementarities between multiple views and promote cooperation among the views. This is the noise matrix of the intermediate layer. represents the weight of the incomplete spectral data in the v-th interval, and the weights of all intervals are summed to 1.
[0042] The multilevel tensor learning module stacks the highest-level similarity graph and the updated lowest-level and intermediate-level similarity graphs into a tensor and rotates it to obtain a multilevel tensor.
[0043] Specifically, the multi-level tensor module is set up as follows:
[0044] in, These are similarity diagrams representing the lowest, middle, and highest levels, respectively. To stack multi-level similar graphs into a tensor and rotate them, For multi-level tensors, , where n is the number of data samples and m is the number of spectral intervals.
[0045] The nuclear norm constraint module constrains the rotated tensor, captures higher-order information, promotes the fusion of multi-level data, and obtains the fused tensor.
[0046] Finally, the underwater multi-view clustering model based on spectral clustering learning was set as follows:
[0047] in, For multi-level tensors, This is the first noise matrix at the highest level. This is the second noise matrix at the highest level. This is the lowest level noise matrix. This is the noise matrix of the intermediate layer. As the first penalty parameter, This is the second penalty parameter.
[0048] S3. Iterate the hyperspectral clustering model based on the alternating variable update strategy until the iteration stopping condition is met, and then output the clustering results based on spectral clustering.
[0049] Specifically, an iterative update strategy is used to alternately update the target variable while keeping other variables fixed until convergence. The alternating direction multiplier method is applied to an underwater multi-view clustering model.
[0050] S31 Update The weights of each interval are changed according to the adaptive strategy, and the update form is as follows:
[0051] in, As a penalty item, The trace of a matrix is the sum of all elements on its main diagonal.
[0052] S32 Update Given that all other variables are constant, we can obtain:
[0053] in, Update the auxiliary matrix for projection. .
[0054] pass Update by decomposing SVD .
[0055] S33 Update Based on the operational rules of the matrix trace and the iterative solution method, the final result will be... Write it in the following form:
[0056] in, It is an identity matrix.
[0057] S34 Update Given that the other variables are constant, we can obtain:
[0058] in, Introduced for ease of calculation Alternatives It is the third Lagrange multiplier. It is the fifth Lagrange multiplier. It is the seventh Lagrange multiplier.
[0059] Based on the operational rules of the matrix trace, the above equation is derived into a Sylvester form for solution:
[0060] in, This is the first coefficient matrix. , This is the second coefficient matrix. , This is the third coefficient matrix. .
[0061] S35. Update the noise matrix Given that all other variables are constant, we can obtain:
[0062] Solve using the following formula:
[0063] in, This is the intermediate reference matrix for the noise matrix. for The List, This is the soft threshold shrinkage parameter.
[0064] S36. Update Based on the operational rules of the matrix trace and the iterative solution method, the final result will be... Write it in the following form:
[0065] in, Matrices introduced to facilitate calculations The first alternative, For the second Lagrange multiplier, It is the fourth Lagrange multiplier.
[0066] S37 Update Based on the operational rules of the matrix trace and the iterative solution method, the final result will be... Write it in the following form:
[0067] in, Matrices introduced to facilitate calculations The second alternative, It is the sixth Lagrange multiplier.
[0068] S38. Update Tensor Given that all other variables are constant, we can obtain:
[0069] Tensor updates are achieved through threshold shrinkage.
[0070] S39. Update Lagrange multipliers and penalty terms:
[0071] in, For the updated Lagrange multipliers, For Lagrange multipliers, The difference between the replacement term and the original matrix. As a penalty item, The maximum value of the penalty item. This is the multiple by which the penalty item increases.
[0072] Specifically,
[0073]
[0074]
[0075]
[0076]
[0077]
[0078] S310. Output clustering results based on spectral clustering, including: After reducing the unified similarity graph information learned from multi-level tensors by spectral clustering, clustering is performed to obtain the classification result: Using spectral clustering models to constrain affinity graphs in low rank, a unified similarity graph is obtained:
[0079] in, To unify the similarity graph, spectral clustering is performed on the unified similarity graph to obtain the clustering results. The number of spectral intervals. These are similarity diagrams representing the lowest, middle, and highest levels, respectively. For the first A range.
[0080] In summary, this invention discloses an incomplete underwater hyperspectral image clustering method based on multi-level tensors. First, a complete latent consensus matrix is obtained by introducing an index matrix and a projection operator. Then, similarity graphs corresponding to the latent consensus representations are learned, thus obtaining the highest-level graph of the tensor. Through dynamic cross-view learning, similarity graphs of the lowest and middle layers are learned to explore consistency and complementarity. Next, the learned three-layer graphs are stacked into a tensor, constrained using weighted sp-norms to achieve information exchange between layers. Adaptive weight updates are used to assign corresponding weights to different intervals. An iterative update strategy is used, fixing one variable while alternately updating other variables until convergence. Finally, spectral clustering is performed on the unified similarity graph to obtain the final classification result. This invention explores the multi-level structural relationships between intervals using multi-level tensors, making the data no longer limited to a single level.
[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for incomplete underwater hyperspectral image clustering based on multilayer tensors, characterized in that, Includes the following steps: Acquire the image to be processed, and acquire incomplete spectral data of the image to be processed in different intervals; A clustering model based on spectral clustering learning is constructed, and the incomplete spectral data is processed based on the clustering model. The clustering model includes: a projection operator information extraction module, a multi-layer graph learning module, an adaptive weight module, a multi-level tensor learning module, and a nuclear norm constraint module. The projection operator information extraction module processes the incomplete spectral data of different intervals using a projection matrix and an index matrix to obtain potential consensus representation matrices for different intervals. The multi-layer graph learning module performs top-level learning based on the latent consensus representation matrix to obtain a top-level similarity graph. Simultaneously, it learns dynamic collaborations within the incomplete spectral data to obtain bottom-level and intermediate-level similarity graphs. The adaptive weighting module adaptively updates the weights based on the importance of the bottom-level and intermediate-level similarity graphs, thereby obtaining the updated bottom-level and intermediate-level similarity graphs. The multi-level tensor learning module stacks the highest-level similarity graph and the updated lowest-level and intermediate-level similarity graphs into a tensor and rotates it to obtain a multi-level tensor. The nuclear norm constraint module constrains the rotated tensor to obtain the fused tensor. The hyperspectral clustering model is iterated based on an alternating variable update strategy until the iteration stopping condition is met, at which point the clustering result is output based on spectral clustering.
2. The incomplete underwater hyperspectral image clustering method based on multilayer tensors according to claim 1, characterized in that, The workflow of the projection operator information extraction module is as follows: Incomplete spectral data from different intervals are processed using a projection matrix to obtain projected data. An index matrix is then used to complete the accurate samples of the projected data, resulting in a latent consensus representation matrix. This latent consensus representation matrix is calculated using the following formula: in, The data is after projection processing. The matrix represents the potential consensus. For orthogonal projection operators, For the first v Incomplete spectral data for each interval, For index matrix, This is the first noise matrix at the highest level. The index matrix is represented as follows: in, For elements in the index matrix, Let be the element in the incomplete spectral data of the v-th interval.
3. The incomplete underwater hyperspectral image clustering method based on multilayer tensors according to claim 1, characterized in that, The multi-layer graph learning module includes a highest similarity graph learning unit and a dynamic cross-view learning unit. The highest similarity graph learning unit calculates the highest-level similarity graph according to the following formula: in, The matrix represents the potential consensus. This is the highest-level similarity graph. This is the second noise matrix at the highest level; The dynamic cross-view learning unit and the adaptive weight module calculate the similarity graphs of the bottom layer and the intermediate layer according to the following formula: in, For the first v Incomplete spectral data for each interval, For index matrix, This is the lowest level similarity graph. This is the lowest level noise matrix. This is an intermediate-level similarity graph. This is the noise matrix of the intermediate layer. For the first v The weight of incomplete spectral data in each interval.
4. The incomplete underwater hyperspectral image clustering method based on multilayer tensors according to claim 1, characterized in that, The multi-level tensor module is configured as follows: in, These are similarity diagrams representing the lowest, intermediate, and highest levels, respectively. To stack multi-level similar graphs into a tensor and rotate them, It is a multi-level tensor.
5. The incomplete underwater hyperspectral image clustering method based on multilayer tensors according to claim 1, characterized in that, The underwater multi-view clustering model based on spectral clustering learning is set as follows: in, For multi-level tensors, This is the first noise matrix at the highest level. This is the second noise matrix at the highest level. This is the lowest level noise matrix. Intermediate layer noise matrix As the first penalty parameter, This is the second penalty parameter.
6. The incomplete underwater hyperspectral image clustering method based on multilayer tensors according to claim 1, characterized in that, The clustering results output based on spectral clustering include: After reducing the unified similarity graph information learned from multi-level tensors through spectral clustering, clustering is performed to obtain the classification result. The calculation method of the unified similarity graph is as follows: Using spectral clustering models to constrain affinity graphs in low rank, a unified similarity graph is obtained: in, To unify similar graphs, The number of spectral intervals. These are similarity diagrams representing the lowest, intermediate, and highest levels, respectively. For the first A range.