Unsupervised hyperspectral image defogging method and system based on mixed prior constraints

Abstract

The invention discloses an unsupervised hyperspectral image defogging method and system based on mixed prior constraints, and relates to the technical field of hyperspectral images. The method mainly comprises the following steps: inputting a foggy hyperspectral image and carrying out normalization processing; constructing an unsupervised defogging network based on mixed prior constraints; performing unsupervised iteration restoration on the foggy image by using the defogging network; an output image when the model converges is the optimal defogged image. According to the method, physical prior and deep learning technologies are creatively fused, so that the core problems that a physical model method in the prior art depends on a specific spectral band excessively, and generalization performance is insufficient due to the fact that a deep learning method lacks supervision data and the like are effectively solved; the method can be widely applied to hyperspectral data defogging tasks in the fields of remote sensing monitoring, geological exploration, environment monitoring, precision agriculture, national defense safety and the like.

Description

Technical Field

[0001] The present invention relates to the field of hyperspectral imaging technology, and in particular to an unsupervised hyperspectral image defogging method and system based on mixed prior constraints. Background Art

[0002] Hyperspectral imagery, with its nearly continuous spectral curve, enables precise classification of surface features and is widely used in land surveying, pathology testing, resource exploration, and other fields. However, due to limitations in the hardware performance of aerospace sensors and the complex atmospheric environment, acquired hyperspectral imagery is often contaminated by haze, resulting in a loss of surface information and increasing the difficulty of interpreting this information. Therefore, developing efficient and robust hyperspectral dehazing methods is crucial.

[0003] Currently, hyperspectral dehazing methods are primarily divided into two categories: physical model-based methods and deep learning-based methods. The former achieves dehazing through atmospheric scattering models, which have clear physical meaning and strong generalization capabilities. However, they rely on modeling the spectral characteristics of the visible and near-infrared bands, making them difficult to handle scenarios lacking relevant band data, severely limiting their breadth of practical application. The latter utilizes deep neural networks to learn end-to-end dehazing mappings from large-scale data. While this overcomes the band-dependency limitations of physical models, it suffers from generalization performance degradation due to the difficulty and high cost of collecting paired hyperspectral data (with / without fog) in real-world scenarios and their susceptibility to the distribution of training data. While these two approaches clearly complement each other in terms of theoretical completeness and data-driven nature, the key challenge in this field remains the development of a collaborative mechanism that deeply integrates physical priors with data features to achieve breakthroughs in both model interpretability and deep representation capabilities. Summary of the Invention

[0004] The purpose of this invention is to propose an unsupervised hyperspectral image dehazing method based on hybrid prior constraints. By innovatively integrating physical priors with deep learning technology, it effectively solves the core problems of the existing technology, such as the excessive dependence of physical model methods on specific spectral bands and the insufficient generalization performance of deep learning methods due to the lack of supervised data.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: an unsupervised hyperspectral image dehazing method based on mixed prior constraints, the specific steps of which are as follows: Step S1, obtaining a hyperspectral image polluted by haze and performing normalization processing as the input of the dehazing network; Step S2: construct an unsupervised dehazing network based on hybrid prior constraints, which includes four key components: low-rank generation module, spectral grouping module, spectral grouping feature fusion module, and residual module; The normalized hyperspectral image first undergoes low-rank matrix decomposition through a low-rank generation module. In the spectral grouping module, an adaptive grouping mechanism is constructed based on spectral correlation to divide the continuous band into multiple strongly complementary spectral subsets. Each spectral subset is treated as a spectral branch. After processing through the spectral grouping feature fusion module and the residual module, the output features of each spectral branch are finally spliced and fused across channels. The residuals are then superimposed with the hyperspectral image before spectral grouping through jump connections to complete end-to-end hyperspectral dehazing image reconstruction. Step S3: Iteratively restore the haze image using the dehazing network and continuously optimize the model parameters through the back propagation algorithm until the model converges; In step S4, the output image when the model converges is the optimal defogging image.

[0006] Furthermore, the hyperspectral image polluted by haze in step S1 is recorded as ,in H is the number of image rows, W is the number of image columns, B is the number of spectral bands. The pixel value normalization operation is performed on the image band by band, and the pixel value range is linearly mapped to the [0,1] interval. The specific normalization method is:

[0007] in, represents the normalized image pixel value, Represents the original pixel value of the image, Indicates the minimum value of the band image pixel value. Indicates the maximum value of the pixel value of the band image.

[0008] Furthermore, in step S2, the low-rank generation module consists of two U-shaped generation networks, named two-dimensional spatial network and one-dimensional spectral network, and the output of the two-dimensional spatial network is the left eigenvector of the matrix decomposition V , the output of the one-dimensional spectral network is the right eigenvector of the matrix decomposition U , the low-rank matrix decomposition process formula is expressed as:

[0009] in, represents the normalized hyperspectral data, where H is the number of image rows, W is the number of image columns, B is the number of spectral bands; Represents the abundance matrix, i.e., the right eigenvector, where each row represents the linear combination coefficient of different end members in a certain pixel. Represents the endmember matrix, i.e., the left eigenvector, where each row represents the spectral eigenvector of an endmember.r Indicates the number of end members.

[0010] Furthermore, in the spectral grouping module, the three-dimensional hyperspectral data is divided into multiple spectral groups according to the correlation between spectral bands. The correlation calculation formula is:

[0011] in, Indicates the i Band and j The correlation coefficient between the bands, Indicates the i bands, represents the covariance, Represents variance.

[0012] Furthermore, in the spectral grouping feature fusion module, for two adjacent spectral groups, parallel convolution layers are first used to extract multi-scale features, and the feature maps output by convolutions of different sizes are spliced and fused; then the multi-scale features of the two spectral groups are aggregated element by element, and global average pooling is used to compress the global spatial information into 1×1× C To enhance the feature response, the channel dimension is first compressed to C / r To capture key information, r is the dimensionality reduction ratio, and then restored to the original dimension C , and finally the multi-scale features are dynamically recalibrated through the channel attention mechanism.

[0013] Furthermore, the parallel convolutional layer includes three convolutional layers, and the convolution kernel sizes are 3×3, 5×5, and 7×7 respectively.

[0014] Furthermore, in the residual module, the input feature map is split into three parallel paths after 1×1 convolution, and the inputs of the three paths are recorded as , where the input of the first path is Directly retain the original features and pass them to the output end, recorded as , the input of the second path After extracting local details through 3×3 convolution, it further bifurcates into two branches, one of which flows to the output To pass the mid-level semantic information, the other branch is connected with the input of the third path Perform feature fusion so that deep features can inherit the multi-granularity information of shallow layers. After the fused features are processed by a 3×3 convolutional layer, the output is By superimposing the cascade operation of small-sized convolution kernels, a large receptive field is constructed equivalently; finally, the multi-scale features of the three paths are After splicing, they are deeply fused through 1×1 convolution to achieve adaptive integration of cross-layer features and information complementarity.

[0015] Furthermore, in step S3, the hyperspectral image acquired in step S1 is iteratively optimized using an unsupervised dehazing network based on hybrid prior constraints. The Adam optimization algorithm is used to drive the back propagation process, and the network weights are dynamically updated by minimizing the mean absolute error loss function, where the loss function is defined as the mean pixel-level absolute difference between the network output image and the input image. Its mathematical expression is:

[0016] in, represents the original foggy input image, represents the dehazed image output by the network, N Indicates the number of spectral bands.

[0017] Furthermore, each spectral branch is processed as follows: first, basic features are extracted through the convolution layer, and then the output features are divided into two branches in parallel. The main branch inputs the current branch features and the adjacent spectral branch features into the spectral grouping feature fusion module for cross-branch fusion interaction, and the auxiliary branch performs feature aggregation together with the fused features through the convolution layer; the aggregated features are subjected to deep feature extraction by the residual module; the output of the residual module is also divided into two branches, and the feature fusion interaction and aggregation operations are repeated to form a progressive feature optimization structure.

[0018] The present invention also provides an unsupervised hyperspectral image defogging system based on mixed prior constraints, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, an unsupervised hyperspectral image defogging method based on mixed prior constraints as described in the above technical solution is implemented.

[0019] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: The unsupervised hyperspectral image dehazing method based on mixed prior constraints in the present invention can achieve accurate restoration of hyperspectral foggy images under unsupervised conditions compared with the existing technology. It not only breaks through the limitations of traditional methods on the number and range of spectral bands of specific sensors, but also has stronger advantages in spectral information preservation and spatial detail recovery. It can be widely used in hyperspectral data dehazing tasks in remote sensing monitoring, geological exploration, environmental monitoring, precision agriculture, national defense and security and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Flowchart of the present invention; Figure 2 This is a diagram of the unsupervised dehazing network structure based on mixed prior constraints in an embodiment of the present invention.

[0021] Figure 3 These are the visual comparison results of the method according to the embodiment of the present invention on three datasets: (a) haze image, (b) clean image, (c) DCP, (d) DHIM, (e) ADM, (f) SLP, (g) RLDP, (h) Defog, and (i) HPC-UDN. DETAILED DESCRIPTION

[0022] In order to provide a better understanding and recognition of the technical solution of the present invention, the following is further described with reference to preferred embodiments and accompanying drawings: like Figure 1 As shown, the unsupervised hyperspectral image dehazing method based on mixed prior constraints described in the present invention includes the following steps: Step S1: Obtain the hyperspectral image polluted by haze and perform normalization processing as the input of the proposed model; Step S2: Construct an unsupervised dehazing network based on hybrid prior constraints, which includes four key components: low-rank generation module, spectral grouping module, spectral grouping feature fusion module and residual module. The network structure is as follows: Figure 2 As shown; The normalized hyperspectral image first undergoes low-rank matrix factorization through a low-rank generative module. A two-branch deep generative network adaptively learns the spectral basis matrix and coefficient matrix to reveal the data's intrinsic structure. An adaptive grouping mechanism is then constructed based on spectral correlation, partitioning continuous bands into strongly complementary spectral subsets. Within each spectral branch, the feature extraction process employs a progressive processing architecture: Basic features are first extracted through a 3×3 convolutional layer. The output features are then split into two branches in parallel. The main branch combines the current branch's features with those of adjacent spectral branches for cross-branch interaction. The auxiliary branch aggregates the fused features through a 3×3 convolutional layer. The aggregated features are then aggregated through a residual module for deep feature extraction, enhancing feature representation while effectively expanding the receptive field. The output of the residual module is similarly split into two branches, and the feature fusion and aggregation operations are repeated, forming a progressive feature optimization architecture. Finally, the output features of each spectral branch are concatenated and fused across channels. The residuals are then superimposed with the pre-grouped hyperspectral image via skip connections, completing end-to-end hyperspectral dehazed image reconstruction.

[0023] Step S3: Use the dehazing network to iteratively restore the haze image and continuously optimize the model parameters through the back propagation algorithm until the model converges; Step S4: The output image when the model converges is the optimal dehazed image.

[0024] The hyperspectral image polluted by haze in step S1 is recorded as ,in H is the number of image rows, W is the number of image columns, B is the number of spectral bands. The pixel values of the image are normalized band by band, and the pixel value range is linearly mapped to the interval [0,1] to improve the convergence speed and stability of the model. The specific normalization method is:

[0025] in, represents the normalized image pixel value, Represents the original pixel value of the image, Indicates the minimum value of the pixel value of the band image. Indicates the maximum value of the pixel value of the image in this band.

[0026] In this embodiment, the image acquired by the Gaofen-5 satellite is selected as the original foggy hyperspectral image, which has a width of 512, a length of 512, and a number of bands of 305.

[0027] In step S2, in order to utilize the low-rank characteristics of hyperspectral images and reduce the computational burden, the low-rank generation module consists of two "U-shaped" generation networks, named two-dimensional spatial network and one-dimensional spectral network. The output of the two-dimensional spatial network is the left eigenvector of the matrix decomposition V , the output of the one-dimensional spectral network is the right eigenvector of the matrix decomposition U The former aims to capture the spatial characteristics of the image, and the latter aims to model the spectral correlation of the image. The low-rank matrix decomposition process formula is expressed as:

[0028] in, represents the normalized hyperspectral data, where H is the number of image rows, W is the number of image columns, B is the number of spectral bands; Represents the abundance matrix, i.e., the right eigenvector, where each row represents the linear combination coefficient of different end members in a certain pixel. Represents the endmember matrix, i.e., the left eigenvector, where each row represents the spectral eigenvector of an endmember. r Indicates the number of end members.

[0029] In this embodiment, the pre-processed foggy hyperspectral image When performing matrix decomposition, the decomposition rank is set to 10. Among them, the dimension of the abundance matrix (right eigenvector) is , the dimension of the end member matrix (left eigenvector) is .

[0030] Hyperspectral images have redundant and highly correlated spectral information. To improve computational efficiency, in the spectral grouping module, the three-dimensional hyperspectral data is divided into multiple spectral groups based on the correlation between spectral bands. The correlation calculation formula is:

[0031] in, Indicates the i Band and j The correlation coefficient between the bands, Indicates the i bands, represents the covariance, Represents variance.

[0032] In this embodiment, five spectrum groups are divided.

[0033] The spectral grouping feature fusion module transfers spectral features and spatial information from the longwave band to the shortwave band through information exchange between adjacent spectral groups. For two adjacent spectral groups, parallel convolutional layers (with kernel sizes of 3×3, 5×5, and 7×7, respectively) are first used to extract multi-scale features. The feature maps output by the convolutions of different sizes are then concatenated and fused. The multi-scale features of the two spectral groups are then aggregated element-by-element, and global average pooling is used to compress the global spatial information into a 1×1×128 channel descriptor. To enhance feature response, the channel dimension is first compressed to 128 / 8 via a 1×1 convolutional layer to capture key information, and then restored to its original dimension of 128. Finally, a channel attention mechanism is used to dynamically recalibrate the multi-scale features, thereby improving the model's discriminative representation of spatial-spectral features.

[0034] In the residual module, multi-scale layered processing and structured residual connection design are used to reduce the number of model parameters while expanding the network receptive field: the input feature map is split into three parallel paths after 1×1 convolution, and the inputs of the three paths are recorded as , where the input of the first path is Directly retain the original features and pass them to the output end, recorded as , the input of the second path After extracting local details through 3×3 convolution, it further bifurcates into two branches, one of which flows to the output To pass the mid-level semantic information, the other branch is connected with the input of the third path Perform feature fusion so that deep features can inherit the multi-granularity information of shallow layers. After the fused features are processed by a 3×3 convolutional layer, the output is By superimposing the cascade operation of small-sized convolution kernels, a large receptive field is constructed equivalently; finally, the multi-scale features of the three paths are After splicing, they are deeply fused through 1×1 convolution to achieve adaptive integration of cross-layer features and information complementarity.

[0035] After the features extracted from each spectral group are spliced and fused, they are combined with the original foggy image through residual connection. X Overlay to generate defogging image .

[0036] In step S3, the hazy hyperspectral image is defogged based on the constructed unsupervised defogging network. Iterative optimization is performed, using the Adam optimization algorithm to drive the back propagation process, and dynamically updating the network weights by minimizing the mean absolute error loss function. The loss function is defined as the network output image With input image The mean absolute difference of pixel levels between the two is expressed as follows:

[0037] in, represents the original foggy input image, represents the dehazed image output by the network, N Indicates the number of spectral bands.

[0038] In this embodiment, the number of spectral bands N It is 305.

[0039] In step S4, when the model converges, the output image at this time is the optimal defogging image.

[0040] The effects of the present invention are further described below in conjunction with simulation experiments: 1. Simulation Experiment Conditions: The computer hardware environment for the experiments in this paper is an Intel(R) Xeon(R) Gold6230R CPU @ 2.10GHz, a GeForce RTX 3090 GPU, and the software environment is Linux. The deep learning framework used is PyTorch, and all subsequent experiments are based on this platform. During the iteration process, the initial learning rate is set to 0.001. The present invention is validated on three sets of remote sensing image datasets: AVIRIS, GF-5, and Chikusei. In addition, the present invention uses the correlation coefficient (CC), the universal image quality index (UIQI), the root mean square error (RMSE), and the spectral angle mapping (SAM) as evaluation indicators.

[0041] 2. Simulation experiment content and result analysis: In order to verify the stability and superiority of the present invention (HPC-UDN) in multiple scenarios, this example selects three public hyperspectral datasets to test the performance of the present invention. The quantitative results on the three datasets are listed in Table 1, Table 2 and Table 3. The best performing method in each column is marked in bold, and the suboptimal results are underlined. It should be noted that due to the limitation of the principle, the Defog method is only applicable to the spectrum containing both visible light bands (0.36-0.65 ) and near infrared band (2.01-2.36 ) image data, making it unsuitable for the Chikusei dataset. A comprehensive comparison of the three datasets shows that the proposed method (HPC-UDN) performs best among all methods, effectively removing haze while significantly improving the spectral fidelity of the image.

[0042] Table 1 Quantitative comparison results on the AVIRIS dataset

[0043] Table 2 Quantitative comparison results on the GF-5 dataset

[0044] Table 3 Quantitative comparison results on the Chikusei dataset

[0045] The qualitative results on the three datasets are as follows Figure 3As shown in the figure, regions 1 to 5 are pseudo-color images from the AVIRIS dataset (bands 29, 20, and 12), regions 6 to 10 are pseudo-color images from the Gaofen-5 dataset (bands 59, 38, and 20), and regions 11 to 12 are pseudo-color images from the Chikusei dataset (bands 60, 40, and 20). To enhance the comprehensiveness and diversity of the experiment, the simulated dataset includes a variety of land cover types, including buildings, farmland, rocks, vegetation, and urban landscapes. The pseudo-color images clearly show that a significant amount of haze remains in the dehazing results of the DCP, DHIM, ADM, SLP, and RLDP algorithms. This is because these methods were developed for three-band natural or remote sensing imagery and therefore struggle to dehaze hyperspectral images. Furthermore, due to the high brightness of haze, the images processed by these algorithms generally exhibit low contrast. Although the Defog method has achieved some success in dehazing, significantly improving haze obscuration, it sacrifices spectral accuracy, resulting in significant spectral distortion. In contrast, HPC-UDN not only effectively removes haze and significantly enhances the visibility of images, but also maximizes the preservation of color authenticity and fidelity, achieving the best balance between dehazing effect and spectral information preservation.

[0046] On the other hand, an embodiment of the present invention also provides an unsupervised hyperspectral image defogging system based on mixed prior constraints, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements an unsupervised hyperspectral image defogging method based on mixed prior constraints as described in the above technical solution.

[0047] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An unsupervised hyperspectral image dehazing method based on mixed prior constraints, characterized by: The following steps are involved: Step S1, obtaining a hyperspectral image polluted by haze and performing normalization processing as the input of the dehazing network; Step S2: construct an unsupervised dehazing network based on hybrid prior constraints, which includes four key components: low-rank generation module, spectral grouping module, spectral grouping feature fusion module, and residual module; The normalized hyperspectral image first undergoes low-rank matrix decomposition through a low-rank generation module. In the spectral grouping module, an adaptive grouping mechanism is constructed based on spectral correlation to divide the continuous band into multiple strongly complementary spectral subsets. Each spectral subset is treated as a spectral branch. After processing through the spectral grouping feature fusion module and the residual module, the output features of each spectral branch are finally spliced and fused across channels. The residuals are then superimposed with the hyperspectral image before spectral grouping through jump connections to complete end-to-end hyperspectral dehazing image reconstruction. Step S3: Iteratively restore the haze image using the dehazing network and continuously optimize the model parameters through the back propagation algorithm until the model converges; In step S4, the output image when the model converges is the optimal defogging image.

2. The unsupervised hyperspectral image dehazing method based on hybrid prior constraints according to claim 1, characterized in that: The hyperspectral image polluted by haze in step S1 is recorded as ,in H is the number of image rows, W is the number of image columns, B is the number of spectral bands. The pixel value normalization operation is performed on the image band by band, and the pixel value range is linearly mapped to the [0,1] interval. The specific normalization method is: in, represents the normalized image pixel value, Represents the original pixel value of the image, Indicates the minimum value of the band image pixel value. Indicates the maximum value of the pixel value of the band image.

3. The unsupervised hyperspectral image dehazing method based on hybrid prior constraints according to claim 1, characterized in that: In step S2, the low-rank generation module consists of two U-shaped generation networks, named two-dimensional spatial network and one-dimensional spectral network. The output of the two-dimensional spatial network is the left eigenvector of the matrix decomposition V , the output of the one-dimensional spectral network is the right eigenvector of the matrix decomposition U , the low-rank matrix decomposition process formula is expressed as: in, represents the normalized hyperspectral data, where H is the number of image rows, W is the number of image columns, B is the number of spectral bands; Represents the abundance matrix, i.e., the right eigenvector, where each row represents the linear combination coefficient of different end members in a certain pixel. Represents the endmember matrix, i.e., the left eigenvector, where each row represents the spectral eigenvector of an endmember. r Indicates the number of end members.

4. The unsupervised hyperspectral image dehazing method based on mixed prior constraints according to claim 1, characterized in that: In the spectral grouping module, the three-dimensional hyperspectral data is divided into multiple spectral groups according to the correlation between spectral bands. The correlation calculation formula is: in, Indicates the i Band and j The correlation coefficient between the bands, Indicates the i bands, represents the covariance, Represents variance.

5. The unsupervised hyperspectral image dehazing method based on hybrid prior constraints according to claim 1, characterized in that: In the spectral grouping feature fusion module, for two adjacent spectral groups, parallel convolution layers are first used to extract multi-scale features, and the feature maps of convolution outputs of different sizes are spliced and fused; then the multi-scale features of the two spectral groups are aggregated element by element, and global average pooling is used to compress the global spatial information into 1×1× C To enhance the feature response, the channel dimension is first compressed to C / r To capture key information, r is the dimensionality reduction ratio, and then restored to the original dimension C , and finally the multi-scale features are dynamically recalibrated through the channel attention mechanism.

6. The unsupervised hyperspectral image dehazing method based on mixed prior constraints according to claim 5, characterized in that: The parallel convolutional layer includes three convolutional layers, and the convolution kernel sizes are 3×3, 5×5, and 7×7 respectively.

7. The unsupervised hyperspectral image dehazing method based on hybrid prior constraints according to claim 1, characterized in that: In the residual module, the input feature map is split into three parallel paths after 1×1 convolution, and the inputs of the three paths are recorded as , where the input of the first path is Directly retain the original features and pass them to the output end, recorded as , the input of the second path After extracting local details through 3×3 convolution, it further bifurcates into two branches, one of which flows to the output To pass the mid-level semantic information, the other branch is connected with the input of the third path Perform feature fusion so that deep features can inherit the multi-granularity information of shallow layers. After the fused features are processed by a 3×3 convolutional layer, the output is By superimposing the cascade operation of small-sized convolution kernels, a large receptive field is constructed equivalently; finally, the multi-scale features of the three paths are After splicing, they are deeply fused through 1×1 convolution to achieve adaptive integration of cross-layer features and information complementarity.

8. The unsupervised hyperspectral image dehazing method based on mixed prior constraints according to claim 1, characterized in that: In step S3, the hyperspectral image acquired in step S1 is iteratively optimized using an unsupervised dehazing network based on hybrid prior constraints. The Adam optimization algorithm is used to drive the back propagation process, and the network weights are dynamically updated by minimizing the mean absolute error loss function, where the loss function is defined as the mean pixel-level absolute difference between the network output image and the input image. Its mathematical expression is: in, represents the original foggy input image, represents the dehazed image output by the network, N Indicates the number of spectral bands.

9. The unsupervised hyperspectral image dehazing method based on hybrid prior constraints according to claim 1, characterized in that: Each spectral branch is processed as follows: first, basic features are extracted through the convolution layer, and then the output features are divided into two branches in parallel. The main branch inputs the current branch features and the adjacent spectral branch features into the spectral grouping feature fusion module for cross-branch fusion interaction, and the auxiliary branch aggregates the features together with the fused features through the convolution layer; the aggregated features are subjected to deep feature extraction by the residual module; the output of the residual module is also divided into two branches, and the feature fusion interaction and aggregation operations are repeated to form a progressive feature optimization structure.

10. An unsupervised hyperspectral image dehazing system based on mixed prior constraints, characterized by: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for dehazing an unsupervised hyperspectral image based on mixed prior constraints as claimed in any one of claims 1 to 9 is implemented.

Images