Hyperspectral remote sensing image compression method, equipment and product based on hybrid hyper-prior

Abstract

This invention discloses a method, device, and product for hyperspectral remote sensing image compression based on a hybrid hyper-prior. This method utilizes a CNN- and MLP-based backbone analysis network and synthesis network to transform the original image into a compact latent representation, obtaining superior latent representation features. Furthermore, a spatial-pass-through attention module is used to further enhance the representational capabilities of the CNN-MLP hybrid network module, enabling in-depth exploration of redundant information within hyperspectral images. When constructing the entropy model, a CNN-based hyper-prior is used to mine local redundant information, while a Transformer-based hyper-prior is used to extract non-local redundant information. The two are then combined to predict the parameters of the entropy model, improving the accuracy of the entropy model construction and, consequently, the final image compression performance.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing and relates to a hyperspectral remote sensing image compression method, device and product, and in particular to a hyperspectral remote sensing image compression method, device and product based on Transformer-CNN hybrid hyper-prior and spatial-channel attention. Background Art

[0002] Hyperspectral images contain hundreds or even thousands of spectral bands, which contain a wealth of spectral information and can accurately characterize the classification characteristics of ground objects. With the recent advancement of imaging technology, the spatial and spectral resolution of hyperspectral images has become increasingly higher, providing richer information for Earth observation. However, the increasing size of hyperspectral image data in these and other dimensions places a significant burden on image storage and transmission. Furthermore, various remote sensing satellite-based Earth-based platforms continuously generate vast amounts of data, further increasing the storage and transmission burden of remote sensing systems.

[0003] In order to save storage space and transmission bandwidth, it is imperative to compress hyperspectral remote sensing image data. Currently, there are many compression methods for hyperspectral remote sensing images, including some models based on traditional algorithms and some models based on deep learning algorithms. However, these algorithms fail to fully consider the widespread spatial redundancy and spectral redundancy in remote sensing images, and do not fully exploit the non-local spatial redundancy in hyperspectral images, resulting in limited compression performance.

[0004] Therefore, there is an urgent need for a lossy image compression algorithm for hyperspectral remote sensing images to fully exploit the local and non-local spatial redundancy and spectral redundancy in hyperspectral remote sensing images, so as to improve compression performance and save image storage space and transmission bandwidth. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a hyperspectral remote sensing image compression method, device and product based on Transformer-CNN hybrid super prior and spatial-channel attention.

[0006] In a first aspect, the present invention provides a method for compressing hyperspectral remote sensing images based on a hybrid hyper-prior, characterized in that it includes the following steps:

[0007] Step 1: Fill the space size of the original hyperspectral remote sensing image to an integer multiple of N, where the value of N is related to the number of downsampling times. When the number of downsampling times is k, N = 2 k ;

[0008] Step 2: Use a network based on spatial-channel attention combined with CNN and MLP to convert the original image into a latent subspace representation, obtain the latent representation coefficients, and quantize the latent representation coefficients;

[0009] Step 3: Use two networks based on Transformer and CNN to extract deep potential representation information from the potential representation respectively, and extract hyper-prior information from the deep potential representation information, and mix the hyper-prior information to obtain hyperprior;

[0010] Step 4: Use the Factorized entropy model parameters to construct the entropy model for the two deep potential representations, and then perform entropy encoding and decoding on the deep potential representations;

[0011] Step 5: Use the hybrid hyper-prior to predict the entropy model parameters for the latent representation;

[0012] Step 6: Using the constructed entropy model, entropy code each pixel of the potential representation one by one and convert it into a binary code stream;

[0013] Step 7: Using a hybrid backbone network based on CNN and MLP, the entropy model and binary code stream are used to reconstruct the latent representation using the entropy decoding algorithm, and the decoded latent representation data is used to generate the reconstructed hyperspectral image.

[0014] Preferably, in step 2, the latent subspace coefficients are quantized for the original hyperspectral remote sensing image X after filling, and are quantized as follows: Y=G a (X); where G a represents the backbone analysis network, which is used to convert the image X into potential representation coefficients; Y, They represent the potential representation information before and after quantization, and U|Q represents the quantization process.

[0015] Preferably, the backbone analysis network and the backbone synthesis network each include four CNN-MLP hybrid modules, and the structure of CNN-MLP is as follows: Figure 2As shown in the figure. In this module, the input features are fed into the CNN and MLP branches respectively, which mine local and non-local features in the image, respectively. The two features are then concatenated and fed into a spatial channel attention module. A convolution / deconvolution layer with a stride of 2 is then used for downsampling or upsampling. In the backbone analysis network, a convolution with a stride of 2 is used for downsampling, while in the backbone synthesis network, a deconvolution layer with a stride of 2 is used for upsampling. The CNN module in the CNN-MLP module first uses a convolution with a kernel of 1 for channel compression, followed by multiple convolutions with kernels of 1, 3, 5, and 7 to extract multi-scale information. These are then concatenated through two convolutional layers to obtain the final features, and a residual link is performed on the output features. In the MLP module, the spatial mapping part uses multi-scale convolutions similar to the CNN module, while the channel mapping part uses two fully connected layers and a LeakyReLU layer. In the spatial channel attention module SC att, two branch networks learn spatial and channel features respectively, then perform point multiplication weighting with the input features and then concatenate them as output.

[0016] As a preference, in step 3, first obtain the data of deep potential representation based on Transformer and CNN network before and after quantization and Among them, H a_t 、H a_c They represent the Transformer-based super-prior extraction network and the CNN-based super-prior extraction network respectively; U|Q represents the quantization process;

[0017] Then based on the Transformer synthesis network H s_t And based on CNN super prior synthesis network and H s_c , extracting super prior information from latent representation information and

[0018] The final mixed hyper-prior information hyperprior=[hyperprior _t ,hyperprior _c ].

[0019] As a preference, the Transformer-based super-prior extraction network and the Transformer-based super-prior synthesis network H s_t, which consists of a sequentially connected transformer module, downsampling layer, transformer module, downsampling layer, U|Q quantization layer, upsampling layer, transformer module, upsampling layer and transformer module;

[0020] The CNN super prior extraction network and the CNN super prior synthesis network and H s_c , which consists of a sequentially connected conv5 network layer, downsampling layer, conv 5 network layer, downsampling layer, U|Q quantization layer, upsampling layer, conv 5 network layer block, upsampling layer and conv 5 network layer.

[0021] Preferably, in step 4, the entropy modeling is to construct a Gaussian mixture model for each pixel of the two quantized depth potential representations. The entropy model of each pixel has k*3 parameters that need to be predicted, where k represents the number of Gaussian models in the Gaussian mixture model; then the probability value p of each point is calculated, and the probability value is used to count the entropy value of the depth potential representation; where, The parameter w of the Gaussian mixture model i , μ i , δ i Through the parameter prediction network Parameter to make predictions, [w i ,μ i ,δ i ]=Parameter(Hyperprior).

[0022] Preferably, the entropy models are all trained models; during the training process of the model, the rate distortion is used as the loss, and the back propagation algorithm is used to optimize the network parameters;

[0023]

[0024] Among them, H_t, H_c, and H_y all represent entropy values, representing the average bits occupied by the bitstream of the two deep potential representation storage and the bitstream occupied by the potential representation; X, is the distortion between the original hyperspectral image and the original hyperspectral image; λ is a balance parameter. Using different λ values, reconstructed images with different compression rates are obtained.

[0025] In the second aspect, the present invention provides a hyperspectral remote sensing image compression device based on hybrid hyper-prior, including a memory, a processor and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, a hyperspectral remote sensing image compression method based on hybrid hyper-prior is implemented.

[0026] In a third aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the hybrid super-prior-based hyperspectral remote sensing image compression method.

[0027] In a fourth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the hybrid super-prior-based hyperspectral remote sensing image compression method.

[0028] Compared with the existing hyperspectral image compression technology, the beneficial effects of the present invention are:

[0029] (1) The present invention uses a spatial channel attention network in the backbone network to fully exploit the spatial-spectral information in hyperspectral remote sensing images and obtain a more compact deep potential representation.

[0030] (2) In the backbone network, the present invention integrates the CNN-based network and the MLP-based network, which improves the representation performance of the backbone network and the image compression effect.

[0031] (3) When modeling the latent representation, the present invention utilizes the Transformer-based hyper-prior to mine the non-local redundant information in the latent representation, utilizes the CNN-based prior to mine the local redundant information in the latent representation, and fuses the local redundancy with the non-local redundancy, thereby constructing a more accurate entropy model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The technical solution of the present invention is further illustrated below using embodiments and specific implementation methods. In addition, some drawings are also used in the process of illustrating the technical solution. For those skilled in the art, other drawings and the intention of the present invention can be obtained based on these drawings without making any creative efforts.

[0033] Figure 1 Schematic diagram of the method in an embodiment of the present invention;

[0034] Figure 2 This is a diagram of the backbone analysis network structure in an embodiment of the present invention;

[0035] Figure 3 This is a network structure diagram of a hybrid super-prior extraction module in an embodiment of the present invention;

[0036] Figure 4 This is a flow chart of encoding and decoding in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0038] In recent years, MLP has regained good performance in multiple visual tasks with its global attention, while CNN has very powerful capabilities in local modeling. Therefore, this embodiment uses a network combining MLP and CNN in the backbone network, and provides an efficient hyperspectral image lossy compression method, device and product in which the backbone network uses spatial-channel attention and CNN-MLP hybrid modules to optimize potential representation features, and super-prior information based on the fusion of Transformer, CNN and local context to fully exploit the local and non-local redundancy and channel redundancy of hyperspectral images.

[0039] In the network proposed in this invention, the information that needs to be stored includes the latent representation and two deep latent representations. Both coefficients require entropy modeling. During entropy encoding and decoding, each quantized pixel has an entropy model associated with it. During encoding and decoding, the probability distribution that this pixel satisfies is known. Therefore, arithmetic coding or other entropy coding methods such as ANS can be used based on this to convert the latent representation into a binary bitstream sequence.

[0040] Please see Figure 1 , this embodiment provides a hyperspectral remote sensing image compression method based on hybrid hyper-prior, comprising the following steps:

[0041] Step 1: Fill the space size of the original hyperspectral remote sensing image to an integer multiple of N, where the value of N is related to the number of downsampling times. When the number of downsampling times is k, N = 2 k ;

[0042] In one embodiment, during the entire compression process, the hyperspectral image needs to undergo a process of multiple downsampling four times. In order to ensure that the reconstructed image is consistent with the original image in size, it is necessary to ensure that the spatial size of the original image is an integer multiple of 32. The backbone network (a network based on spatial-channel attention combined with CNN and MLP) includes four downsampling and upsampling, while the hybrid super-prior extraction module (a super-prior extraction network based on Transformer and CNN) includes one downsampling. The size of the original image is reduced by 32 times). At this time, when the spatial size of the hyperspectral image is not an integer multiple of 32, the edges of the original image are padded with zeros to expand it to a multiple of 32. At this time, assuming that the original image size is [c, w, h] and the expanded image size is [c, w_1, h_1], then w_1 and h_1 should meet the following conditions:

[0043] w_1%32=0,w_1-w<32;

[0044] h_1%32=0,h_1-h<32.

[0045] Step 2: Use a network based on spatial-channel attention combined with CNN and MLP to convert the original image into a latent subspace representation, obtain the latent representation coefficients, and quantize the latent representation coefficients;

[0046] In one embodiment, the original image is partially passed through the backbone analysis network (a network based on spatial-channel attention combined with CNN and MLP) to obtain the potential representation coefficients. This process can be expressed as:

[0047] Y=G a (X);

[0048]

[0049] Where X represents the expanded hyperspectral image data, Y and Represents the potential representation data before and after quantization. G a Represents the backbone analysis network. U|Q is the quantizer. During training, the quantizer adds uniform noise, which causes gradient truncation and makes the network untrainable. During testing or actual compression, the quantizer can use integer quantization.

[0050] The process of extracting deep latent representation from the latent representation and quantizing it can be expressed as:

[0051]

[0052] Here, two hyper-prior branches are used to obtain the Transformer-based and CNN-based deep latent representations, respectively. These values ​​need to be stored. To do this, a factorized entropy model is constructed for each pixel. This entropy model can be used to entropy encode the two deep latent representations.

[0053] In one embodiment, see Figure 2 The backbone analysis network and the backbone synthesis network both include four CNN-MLP hybrid modules. The structure of CNN-MLP is as follows: Figure 2As shown in the figure. In this module, the input features are fed into the CNN and MLP branches respectively, which mine local and non-local features in the image, respectively. The two features are then concatenated and fed into a spatial channel attention module. A convolution / deconvolution layer with a stride of 2 is then used for downsampling or upsampling. In the backbone analysis network, a convolution with a stride of 2 is used for downsampling, while in the backbone synthesis network, a deconvolution layer with a stride of 2 is used for upsampling. The CNN module in the CNN-MLP module first uses a convolution with a kernel of 1 for channel compression, followed by multiple convolutions with kernels of 1, 3, 5, and 7 to extract multi-scale information. These are then concatenated through two convolutional layers to obtain the final features, and a residual link is performed on the output features. In the MLP module, the spatial mapping part uses multi-scale convolutions similar to the CNN module, while the channel mapping part uses two fully connected layers and a LeakyReLU layer. In the spatial channel attention module SC att, two branch networks learn spatial and channel features respectively, then perform point multiplication weighting with the input features and then concatenate them as output.

[0054] Step 3: Use the Transformer-based hyper-prior extraction network and the CNN-based hyper-prior extraction network to extract deep potential representation information from the potential representation, and extract hyper-prior information from the deep potential representation information, and mix the hyper-prior information to obtain the hyperprior;

[0055] In one embodiment, see Figure 3 , the Transformer-based super-prior extraction network and the Transformer-based super-prior synthesis network H s_t , which consists of a sequentially connected transformer module, downsampling layer, transformer module, downsampling layer, U|Q quantization layer, upsampling layer, transformer module, upsampling layer and transformer module;

[0056] The CNN super prior extraction network and the CNN super prior synthesis network and H s_c , which consists of a sequentially connected conv5 network layer, downsampling layer, conv 5 network layer, downsampling layer, U|Q quantization layer, upsampling layer, conv 5 network layer block, upsampling layer and conv 5 network layer.

[0057] In one embodiment, this process can be expressed as:

[0058]

[0059] Among them, Z_t, Z_c and Represents the data of deep potential representation based on Transformer and CNN network before and after quantization, respectively. a_t 、H s_t 、H a_c and H s_c The two represent the Transformer-based and CNN-based super-prior analysis and synthesis networks, respectively. Due to the Transformer's strong non-local representation capabilities, it can fully exploit the non-local redundancy present in the latent representation of hyperspectral images. CNNs, on the other hand, have a strong ability to represent local information, so they can fully capture the local redundant information present in the latent representation. The combination of these two redundant information forms the final hybrid super-prior information, namely:

[0060] hyperprior=[hyperprior _t ,hyperprior _c ]

[0061] During compression, in order to ensure that the super-prior information obtained by the encoding segment and the decoding end is consistent, it is necessary to store the deep potential representation to obtain the super-prior information to construct an entropy model for the potential representation. During training, this step only requires predicting the entropy value. During testing or actual compression, entropy coding is required to encode it first, and then decode it to obtain the decoded coefficient value. It is worth mentioning that the entropy coding algorithm is a lossless data compression algorithm, so this step does not lose any information. Here, it is assumed that the entropy values ​​obtained are H_c and H_t respectively (during testing, actual compression is used to obtain two code streams, bit_1 and bit_2).

[0062] Step 4: Use the Factorized entropy model parameters to construct the entropy model for the two deep potential representations, and then perform entropy encoding and decoding on the deep potential representations;

[0063] In one embodiment, each pixel in the depth latent representation is assumed to follow a single Gaussian distribution whose parameters are trained on the entire dataset. The entropy model used is a factorized model. Once the factorized entropy model is obtained, it can be used during compression to estimate the entropy values ​​H_t and Hc of the depth latent representation or, in combination with arithmetic coding, directly obtain the binary bitstreams bit-t and bit-c.

[0064] Step 5: Use the mixed hyper-prior to predict the entropy model parameters for the potential representation. The entropy model can use a Gaussian mixture model.

[0065] In one embodiment, when constructing the entropy model of the potential representation, the hyperprior value obtained from the deep potential representation is used as the prior for probability modeling, which can improve the accuracy of the entropy modeling, thereby making the predicted entropy value smaller. Considering that when the number of mixtures of the Gaussian mixture model is large enough, the Gaussian mixture model can approximate the actual predicted probability model with arbitrary precision. In this embodiment, the mixed hyperprior is used to predict the entropy model of the potential representation. For each pixel point of the potential representation, a Gaussian mixture model is constructed. That is, each pixel point of the potential representation obeys a Gaussian distribution: (The compression process is mainly to store the potential representation. In order to construct the entropy model of the potential representation, some prior knowledge is required. Therefore, the deep potential representation is extracted using the Transformer and CNN dual-branch network and stored as the prior for the upper model prediction.)

[0066]

[0067] Here, the entropy model of each pixel has k*3 parameters to be predicted, where k represents the number of Gaussian models in the Gaussian mixture model. The entropy model parameter prediction network is used to predict it, which can be expressed as:

[0068] [w i ,μ i ,δ i ]=Parameter(Hyperprior)

[0069] Above, the input to the parameter prediction network is the mixed hyper-prior information.

[0070] By obtaining the entropy model of each pixel of the potential representation through the parameter prediction network, the image can be entropy predicted or encoded and decoded based on the entropy model of the potential representation.

[0071] When training the network, the entropy value is used to predict the average bit rate of the final encoding. During testing, based on these probabilities, encoding algorithms such as arithmetic coding can be used to obtain a bit rate close to the entropy value. Once the probability value p is known, the entropy value is calculated as follows:

[0072] H=-p*logp

[0073] During the network training phase, the entropy values ​​of two deep potential representations can be obtained, namely H_t and H_c.

[0074] During training, after building the entropy model, the entropy value H_y of all potential representations can be calculated according to the information entropy formula (during testing, encoding can be performed based on the probability value, and the encoding bit rate is bit_y).

[0075] Step 5: Using the constructed entropy model, entropy code each pixel of the potential representation one by one and convert it into a binary code stream;

[0076] In one embodiment, see Figure 4 After constructing an entropy model for each pixel in the latent representation, entropy coding algorithms such as arithmetic coding can be used to compress the latent representation. During compression, a probability model is constructed for each pixel, allowing the probability value corresponding to each current pixel to be calculated. During encoding, the probability distribution and the latent representation pixel value to be encoded are known, resulting in the entropy prediction result H_y or the compressed bitstream bit_y. Entropy coding algorithms are lossless, so the coefficient values ​​of the latent representation remain unchanged after decoding.

[0077] Step 6: Using a hybrid backbone network based on CNN and MLP, the entropy model and binary code stream are used to reconstruct the latent representation using the entropy decoding algorithm, and the decoded latent representation data is used to generate the reconstructed hyperspectral image.

[0078] In one embodiment, the potential representation Through the backbone synthesis network, a reconstructed hyperspectral image is generated. This process can be expressed as:

[0079]

[0080] The reconstructed hyperspectral image is equal in size to the original (expanded) image. Once the reconstructed image is obtained, losses can be calculated relative to the original image, such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and multi-scale structural similarity. During network training, the MSE loss is calculated and combined with the three predicted entropy values ​​for model training.

[0081] In one embodiment, the entropy model is a trained model; during the model training process, the rate distortion is used as the loss, and the back propagation algorithm is used to optimize the network parameters;

[0082]

[0083] Among them, H_t, H_c, and H_y all represent entropy values, representing the average bits of the three bitstreams (i.e., the bitstream occupied by the two deep potential representation storage and the bitstream occupied by the potential representation); X, is the distortion between the original hyperspectral image and the original hyperspectral image; λ is a balancing parameter. Using different λ values, we can obtain reconstructed images with different compression rates. It is worth mentioning that for different λ values, the proposed algorithm needs to use different model parameters to compress the image.

[0084] This paper proposes a method for lossy compression of hyperspectral remote sensing images. It uses a spatial-channel attention module combined with a hybrid CNN and MLP backbone network to obtain a deep latent representation. Within the deep latent representation, a hybrid super-prior based on a Transformer and CNN is used to extract local and non-local redundancy in the latent representation.

[0085] In addition, in hyperspectral remote sensing images, there is a lot of spectral redundancy and local-non-local spatial redundancy. In order to further improve the performance of the backbone network, the present invention further uses the spatial-channel attention network in the hyperspectral network to further improve the nonlinear representation ability of the backbone network and obtain better potential representation features. In the potential representation, there is still a large amount of spatial local and non-local redundancy. Fully exploiting these spatial local and non-local redundancies can improve the accuracy of entropy model construction. The present invention uses an entropy model prediction framework based on super-prior. Considering the extremely strong representation ability of CNN and Transformer, we use a super-prior extraction module based on a mixture of CNN and Transformer to remove local and non-local redundant information respectively, so as to improve the final entropy model prediction accuracy.

[0086] At the same time, the algorithm proposed in this invention is a lossy compression algorithm for hyperspectral images. Therefore, the entire process needs to consider not only the length of the code stream sequence (bit rate), but also the image reconstruction effect. Therefore, the final training loss compares the rate-distortion performance, that is, under the same bit rate conditions, the better the quality of the reconstructed image, the better the performance of the image compression algorithm.

[0087] It should be understood that the embodiments described above are only some of the embodiments of the present invention, rather than all of the embodiments. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention may be arbitrarily combined with each other to form a feasible technical solution. Such combination is not restricted by the order of steps and / or structural composition mode, but must be based on the ability of ordinary technicians in this field to implement it. When the combination of technical solutions is mutually inconsistent or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0088] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A hyperspectral remote sensing image compression method based on hybrid hyper-prior, characterized in that: The following steps are involved: Step 1: Fill the space size of the original hyperspectral remote sensing image to an integer multiple of N, where the value of N is related to the number of downsampling times. When the number of downsampling times is k, N = 2 k ; Step 2: Use a backbone analysis network based on spatial-channel attention combined with CNN and MLP to convert the original image into a latent subspace representation, obtain the latent representation coefficients, and quantize the latent representation coefficients; The backbone analysis network includes four CNN-MLP hybrid modules. The input features are input into the CNN and MLP branches respectively, and the local and non-local features in the image are mined respectively. Then the two features are concatenated and input into a spatial channel attention module, and then a convolution / deconvolution layer with a stride of 2 is used for downsampling or upsampling. In the backbone analysis network part, a convolution with a stride of 2 is used for downsampling, and in the backbone synthesis network, a deconvolution layer with a stride of 2 is used for upsampling. The CNN module in the CNN-MLP hybrid module first uses a convolution with a convolution kernel of 1 for channel compression, and then uses multiple convolutions with convolution kernels of 1, 3, 5, and 7 to extract multi-scale information. They are then concatenated through two convolution layers to obtain the final features, and a residual link is performed on the output features. In the MLP module, the spatial mapping part uses multi-scale convolution similar to the CNN module, and the channel mapping part uses two fully connected layers and a Leakyrelu layer. The spatial channel attention module SC In att, two branch networks learn spatial and channel features respectively, then perform point multiplication weighting with the input features and concatenate them as output; Step 3: Use the Transformer-based hyper-prior extraction network and the CNN-based hyper-prior extraction network to extract deep potential representation information from the potential representation, and extract hyper-prior information from the deep potential representation information, and mix the hyper-prior information to obtain the hyperprior; Step 4: Use the Factorized entropy model parameters to construct the entropy model for the two deep potential representations, and then perform entropy encoding and decoding on the deep potential representations; Step 5: Use the hybrid hyper-prior to predict the entropy model parameters for the latent representation; Step 6: Using the constructed entropy model, entropy code each pixel of the potential representation one by one and convert it into a binary code stream; Step 7: Using the CNN and MLP-based backbone synthesis network, the entropy model and binary code stream are used to reconstruct the latent representation using the entropy decoding algorithm, and the decoded latent representation data is used to generate the reconstructed hyperspectral image.

2. The method for hyperspectral remote sensing image compression based on hybrid super prior according to claim 1, characterized in that: In step 2, the potential representation coefficient is quantized for the original hyperspectral remote sensing image X after filling, and is quantized as: Y=G a (X); where G a represents the backbone analysis network, which is used to convert the image X into potential representation coefficients; Y, They represent the potential representation information before and after quantization, and U|Q represents the quantization process.

3. The method for hyperspectral remote sensing image compression based on hybrid super prior according to claim 1, characterized in that: The backbone synthesis network includes four CNN-MLP hybrid modules. The input features are respectively input into the CNN and MLP branches to mine the local and non-local features in the image respectively. Then the two features are concatenated and input into a spatial channel attention module. Then a convolution / deconvolution layer with a stride of 2 is used for downsampling or upsampling. In the backbone analysis network part, a convolution with a stride of 2 is used for downsampling, and in the backbone synthesis network, a deconvolution layer with a stride of 2 is used for upsampling. The CNN module in the CNN-MLP hybrid module first uses a convolution with a convolution kernel of 1 for channel compression, and then uses convolution kernels of 1, 3, 5, and 7 to extract multi-scale information. They are then concatenated through two convolution layers to obtain the final features, and a residual link is performed on the output features. In the MLP module, the spatial mapping part uses multi-scale convolution similar to the CNN module, and the channel mapping part uses two fully connected layers and a Leakyrelu layer. The spatial channel attention module SC In ATT, two branch networks learn spatial and channel features respectively, which are then weighted by point multiplication with the input features and then concatenated as output.

4. The method for hyperspectral remote sensing image compression based on hybrid super prior according to claim 1, characterized in that: In step 3, we first obtain the data of the deep potential representation based on the Transformer and CNN networks before and after quantization. and Among them, H a_t 、H a_c They represent the Transformer-based super-prior extraction network and the CNN-based super-prior extraction network respectively; U|Q represents the quantization process; Then based on the Transformer synthesis network H s_t And based on CNN super prior synthesis network and H s_c , extracting super prior information from latent representation information and The final mixed hyper-prior information hyperprior=[hyperprior _t ,hyperprior _c ].

5. The method for hyperspectral remote sensing image compression based on hybrid super prior according to claim 1, characterized in that: The Transformer-based super-prior extraction network and the Transformer-based super-prior synthesis network H s_t , which consists of a sequentially connected transformer module, downsampling layer, transformer module, downsampling layer, U|Q quantization layer, upsampling layer, transformer module, upsampling layer and transformer module; The CNN super prior extraction network and the CNN super prior synthesis network and H s_c , which consists of a sequentially connected conv 5 network layer, downsampling layer, conv 5 network layer, downsampling layer, U|Q quantization layer, upsampling layer, conv 5 network layer block, upsampling layer and conv 5 network layer.

6. The method for hyperspectral remote sensing image compression based on hybrid super prior according to claim 1, characterized in that: In step 4, the entropy model is to construct a Gaussian mixture model for each pixel of the two quantized depth potential representations. The entropy model of each pixel has k*3 parameters that need to be predicted, where k represents the number of Gaussian models in the mixed Gaussian model; then the probability value p of each point is calculated, and the probability value is used to count the entropy value of the depth potential representation; where The parameter w of the Gaussian mixture model i , μ i , δ i Prediction is performed through the parameter prediction network Parameter, [w i ,m i ,d i ] = Parameter (Hyperprior).

7. The method for hyperspectral remote sensing image compression based on hybrid super prior according to any one of claims 1 to 6, characterized in that: The entropy models are all trained models. During the training process of the model, the rate distortion is used as the loss, and the back propagation algorithm is used to optimize the network parameters. Among them, H_t, H_c, and H_y all represent entropy values, representing the average bits occupied by the bitstream of the two deep potential representation storage and the bitstream occupied by the potential representation; X, is the distortion between the original hyperspectral image and the original hyperspectral image; λ is a balance parameter. Using different λ values, reconstructed images with different compression rates are obtained.

8. A hyperspectral remote sensing image compression device based on hybrid hyper-prior, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the hyperspectral remote sensing image compression method based on hybrid super prior according to any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for compressing hyperspectral remote sensing images based on hybrid super-prior according to any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for compressing hyperspectral remote sensing images based on hybrid super-prior according to any one of claims 1 to 7 is implemented.

Images