CN113393543A - Hyperspectral image compression method, device and equipment and readable storage medium - Google Patents

Abstract

The invention provides a hyperspectral image compression method, a hyperspectral image compression device, hyperspectral image compression equipment and a readable storage medium. The method comprises the following steps: training a convolutional neural network through a training set, wherein the convolutional neural network comprises a nonlinear transformation module, a quantization module and an entropy model; and verifying the compression performance of the trained convolutional neural network by using the test set, and compressing the hyperspectral image by using the trained convolutional neural network when the compression performance of the trained convolutional neural network reaches the standard. The method has better rate distortion performance for the compression of the hyperspectral image.

Description

Hyperspectral image compression method, apparatus, device and readable storage medium

technical field

The present invention relates to the technical field of image processing, and in particular, to a hyperspectral image compression method, apparatus, device and readable storage medium.

Background technique

Hyperspectral images have rich and unique spectral information, which brings great convenience to many applications based on hyperspectral images, such as crop classification, quality inspection, disaster prediction and other tasks. However, this advantage also restricts the further development of hyperspectral images under the limited transmission bandwidth and storage capacity. Therefore, how to effectively solve the various challenges brought by the large amount of hyperspectral image data is the premise and key to the wide application of hyperspectral images.

Among hyperspectral image compression algorithms, transform coding is widely used because of its small computational complexity and good adaptability. The image compression algorithm based on transform coding includes four parts: transform, quantization and entropy coding, inverse transform, separate encoding process and decorrelation process.

The transformation is to transform the image from the pixel domain to a more compact space in a certain way. The existing hyperspectral image compression methods based on transform coding generally assume that the hyperspectral image is a Gaussian source, and only a reversible linear transformation is required under this condition. That is, pixels can be mapped into independent latent representations, and latent variables can be compressed into code streams for storage and transmission through quantization and entropy coding. However, hyperspectral images of actual scenes have obvious non-Gaussian properties, which makes linear transformation no longer applicable. The exploration of nonlinear transformation provides new methods and ideas for this problem. In recent years, the development of nonlinear transformations using artificial neural networks, especially deep learning, has changed the situation of manual parameter setting in traditional image compression. The existing deep learning-based image compression technologies show great potential, and the performance has exceeded the H.266/VVC (Versatile Video Coding, VVC) standard in the industry. However, these methods are mostly used for processing three-band natural images, and the compression of hyperspectral images is relatively less.

The transformation process allows quantization and entropy encoding to be performed in a compact space. Compared with RGB natural images, the spectra of hyperspectral images have stronger correlation, which makes the potential representation of hyperspectral images after the same transformation process. Has different statistical properties than RGB images. After quantization, the latent variable becomes a discrete form, which is then encoded based on an entropy coding algorithm. The process of entropy coding depends on the probability distribution model of latent variables. According to the minimum entropy theory, the closer the designed entropy model is to the real latent variable distribution, the smaller the code rate, and the closer the solution obtained in the entropy rate optimization process is to the optimal solution. .

Combined with the above analysis, the current compression technology based on deep learning needs to further design a more flexible and accurate entropy model according to the characteristics of hyperspectral images, so as to reduce the mismatch between the entropy model and the distribution of real latent variables, so as to achieve the optimal rate. Distortion performance.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the present invention provides a hyperspectral image compression method, apparatus, device and readable storage medium.

In a first aspect, the present invention provides a hyperspectral image compression method, the hyperspectral image compression method comprising:

The convolutional neural network is trained through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model;

Use the test set to verify the compression performance of the trained convolutional neural network. When the compression performance of the trained convolutional neural network reaches the standard, the hyperspectral image is compressed by the trained convolutional neural network.

Optionally, before the step of training the convolutional neural network through the training set, the method further includes:

Divide the sample hyperspectral image into several fixed-size cubes in the spatial dimension;

The several fixed-size cubes are divided into a training set and a test set according to a preset ratio.

Optionally, the nonlinear transformation module performs forward nonlinear transformation on the spatial and spectral dimensions of the hyperspectral image to obtain latent variables; the quantization module quantifies the latent variables by adding uniform noise; the entropy model is used to obtain the latent variables. A probability distribution for determining the codeword assigned to each element in the latent variable based on the probability distribution during entropy encoding.

Optionally, the convolutional neural network training process is constrained based on the rate-distortion criterion, which is used to determine the nonlinear transformation module and parameter values in the entropy model.

其中，

表示输入的高光谱图像，

表示重构图像，H,W,B分别对应高光谱图像的行、列和波段数，

表示潜变量，h,w,N分别对应潜变量的行、列和滤波器的个数，

和

表示正变换网络参数，

和

表示反变换网络参数，g_a(.)表示非线性正向变换函数，g_s(.)表示非线性反向变换函数。Optionally, the nonlinear transformation includes forward transformation: Y=ga (W _a X+ _{b a} ), and reverse transformation:

in,

represents the input hyperspectral image,

represents the reconstructed image, H, W, B correspond to the number of rows, columns and bands of the hyperspectral image, respectively,

Represents latent variables, h, w, N correspond to the number of rows, columns and filters of latent variables, respectively,

and

represents the positive transformation network parameters,

and

represents the inverse transformation network parameters, ga (.) represents the nonlinear forward transformation function, and _{g s} (.) represents the nonlinear reverse transformation function.

Optionally, the function for quantizing latent variables by adding uniform noise is expressed as follows:

表示单位均匀噪声，round表示取整操作，

表示量化后的潜向量。Among them, training represents the training process, testing represents the testing process,

represents unit uniform noise, round represents the rounding operation,

represents the quantized latent vector.

Optionally, the statistical characteristics of latent variables are introduced into the design of the entropy model, and at the same time, additional variables are introduced to construct a conditional model to improve the accuracy of the entropy model.

In a second aspect, the present invention also provides a hyperspectral image compression device, the hyperspectral image compression device comprising:

a training module for training the convolutional neural network through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model;

The processing module is used to verify the compression performance of the trained convolutional neural network using the test set. When the compression performance of the trained convolutional neural network reaches the standard, the hyperspectral image is compressed by the trained convolutional neural network.

In a third aspect, the present invention also provides a hyperspectral image compression device, the hyperspectral image compression device includes a processor, a memory, and a hyperspectral image compression program stored on the memory and executable by the processor , wherein when the hyperspectral image compression program is executed by the processor, the steps of the hyperspectral image compression method described above are implemented.

In a fourth aspect, the present invention also provides a readable storage medium on which a hyperspectral image compression program is stored, wherein when the hyperspectral image compression program is executed by a processor, the Steps of a spectral image compression method.

In the present invention, the convolutional neural network is trained through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model; the compression performance of the trained convolutional neural network is verified by using the test set. When the compression performance of the completed convolutional neural network reaches the standard, the hyperspectral image is compressed through the trained convolutional neural network. The present invention has better rate-distortion performance for hyperspectral image compression.

Description of drawings

FIG. 1 is a schematic diagram of the hardware structure of a hyperspectral image compression device involved in an embodiment of the present invention;

2 is a schematic flowchart of an embodiment of a hyperspectral image compression method according to the present invention;

FIG. 3 is a schematic diagram of functional modules of an embodiment of a hyperspectral image compression apparatus according to the present invention.

The realization, functional characteristics and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

In a first aspect, an embodiment of the present invention provides a hyperspectral image compression device, where the hyperspectral image compression device may be a personal computer (personal computer, PC), a notebook computer, a server, and other devices with data processing functions.

Referring to FIG. 1 , FIG. 1 is a schematic diagram of a hardware structure of a hyperspectral image compression device involved in an embodiment of the present invention. In this embodiment of the present invention, the hyperspectral image compression device may include a processor 1001 (for example, a central processing unit, Central Processing Unit, CPU), a communication bus 1002 , a user interface 1003 , a network interface 1004 , and a memory 1005 . Wherein, the communication bus 1002 is used to realize the connection and communication between these components; the user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard); the network interface 1004 may optionally include a standard wired interface, a wireless interface (such as wireless fidelity WIreless-FIdelity, WI-FI interface); the memory 1005 can be a high-speed random access memory (random access memory, RAM), or can be a stable memory (non-volatile memory), such as disk memory, memory Optionally, 1005 may also be a storage device independent of the aforementioned processor 1001 . Those skilled in the art can understand that the hardware structure shown in FIG. 1 does not constitute a limitation of the present invention, and may include more or less components than those shown in the drawings, or combine some components, or arrange different components.

Continuing to refer to FIG. 1 , the memory 1005 as a computer storage medium in FIG. 1 may include an operating system, a network communication module, a user interface module and a hyperspectral image compression program. The processor 1001 may call the hyperspectral image compression program stored in the memory 1005, and execute the hyperspectral image compression method provided by the embodiment of the present invention.

In a second aspect, an embodiment of the present invention provides a hyperspectral image compression method.

In an embodiment, referring to FIG. 2 , FIG. 2 is a schematic flowchart of an embodiment of a hyperspectral image compression method according to the present invention. As shown in Figure 2, hyperspectral image compression methods include:

Step S10, train the convolutional neural network through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model;

In this embodiment, a training set is pre-built, and the convolutional neural network is trained through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module, and an entropy model.

Further, in an embodiment, before step S10, it further includes:

The sample hyperspectral image is divided into several fixed-size cubes in the spatial dimension; the several fixed-size cubes are divided into a training set and a test set according to a preset ratio.

In this embodiment, it is necessary to prepare a data set, including a training set and a test set, and set the hyperparameters of the convolutional neural network before training starts. For example, 27 KAIST datasets of size 2704 × 3376 × 31 (30) and 28 CAVE data of 512 × 512 × 31 (32) are randomly cropped into image patches of size 128 × 128 × 31. The experiment uses the tensorflow framework for model training, and sends the cropped 128×128×31 image block batches (batchesize=32) into the built network, and iterates 500,000 times. The loss function during training is as follows:

采用全分解的单位均匀密度函数，因此损失函数的第一项

表示失真项，在训练过程中采用带参数λ的均方误差MSE损失进行失真度量，λ的取值从0.00001到0.01之间，通过调整的取值可使得bppp(编码每个波段每个像素所占的比特数)的范围控制在0.1到2之间(λ越大，bppp越大)；

表示总的编码比特数。where the approximate posterior

A fully decomposed unit uniform density function is used, so the first term of the loss function

Represents the distortion term. In the training process, the mean square error MSE loss with the parameter λ is used to measure the distortion. The value of λ ranges from 0.00001 to 0.01. The range of the number of bits occupied) is controlled between 0.1 and 2 (the larger the λ, the larger the bppp);

Indicates the total number of encoded bits.

Further, in one embodiment, the nonlinear transformation module performs forward nonlinear transformation on the spatial and spectral dimensions of the hyperspectral image to obtain latent variables; the quantization module quantifies the latent variables by adding uniform noise; the entropy model is used for A probability distribution of latent variables is obtained for determining a codeword assigned to each element in the latent variable based on the probability distribution during entropy coding.

In this embodiment, the training set is input into the convolutional neural network, and the nonlinear transformation module in the convolutional neural network performs forward nonlinear transformation on the spatial and spectral dimensions of the hyperspectral image, so that the image is mapped from the pixel domain to a The compact latent layer space is used to obtain latent variables; then, the quantization module in the convolutional neural network quantizes the latent variables by adding uniform noise. The training process will make the gradient of back propagation to 0 due to quantization. In order to ensure smooth training In this embodiment, the method of adding uniform noise is used to replace the quantization process, so that the quantization can be guided, and the test process is directly rounded; and then, the probability distribution of the latent variables is obtained based on the entropy model, for entropy coding, based on the probability distribution Determine the codeword assigned to each element in the latent variable (ie, how many characters per element).

Further, in one embodiment, the training process of the convolutional neural network is constrained based on the rate-distortion criterion, which is used to determine the nonlinear transformation module and parameter values in the entropy model.

In this embodiment, the rate-distortion criterion is used to solve the parameter values in the nonlinear transformation module and the entropy model, and the concept of variational inference is combined with rate-distortion in the optimization process, and the process of rate-distortion optimization is explained from the perspective of probability:

表示重构图像，d(.)表示失真度量准则，在高光谱图像中一般采用峰值信噪比PSNR，结构相似度SSIM对像素进行失真度量，其值越大表示像素的重构效果越好，利用光谱角SAM度量光谱的重构精度，λ表示拉格朗日乘子。in,

Represents the reconstructed image, and d(.) represents the distortion measurement criterion. In hyperspectral images, the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) are generally used to measure the pixel distortion. The larger the value, the better the reconstruction effect of the pixel. The spectral reconstruction accuracy is measured by the spectral angle SAM, and λ represents the Lagrange multiplier.

In order to optimize the above loss function, the idea of variational inference is adopted, and an approximate posterior is designed to approximate the real posterior to solve the parameters, and the KL divergence is used to measure the two posteriors. The calculation formula is as follows:

表示近似后验，可用任何简单的分布表示，在压缩中一般表示为全分解的单位均匀分布，以便将该项从损失函数中去除，而其余除常数项const之外的三项，

对应失真

对应码率

对应额外的信息

in,

Represents the approximate posterior, which can be represented by any simple distribution, generally expressed as a fully decomposed unit uniform distribution in compression, so that this term is removed from the loss function, and the remaining three terms except the constant term const,

Corresponding distortion

Corresponding code rate

Corresponds to additional information

其中，

表示输入的高光谱图像，

表示重构图像，H,W,B分别对应高光谱图像的行、列和波段数，

表示潜变量，h,w,N分别对应潜变量的行、列和滤波器的个数，

和

表示正变换网络参数，

和

表示反变换网络参数，g_a(.)表示非线性正向变换函数，g_s(.)表示非线性反向变换函数。Further, in an embodiment, the nonlinear transformation includes a forward transformation: Y=ga (W _a X+b _a ), and _a reverse transformation:

in,

represents the input hyperspectral image,

represents the reconstructed image, H, W, B correspond to the number of rows, columns and bands of the hyperspectral image, respectively,

Represents latent variables, h, w, N correspond to the number of rows, columns and filters of latent variables, respectively,

and

represents the positive transformation network parameters,

and

represents the inverse transformation network parameters, ga (.) represents the nonlinear forward transformation function, and _{g s} (.) represents the nonlinear reverse transformation function.

Further, in an embodiment, the function for quantizing latent variables by adding uniform noise is expressed as follows:

表示单位均匀噪声，round表示取整操作，

表示量化后的潜向量。Among them, training represents the training process, testing represents the testing process,

represents unit uniform noise, round represents the rounding operation,

represents the quantized latent vector.

In this embodiment, the latent variable is quantized, and the unit uniform noise approximation is adopted during training, and the rounding method is adopted during testing.

Further, in one embodiment, the statistical characteristics of the latent variables are introduced into the design of the entropy model, and at the same time, an additional variable is introduced to construct a conditional model, so as to improve the accuracy of the entropy model.

In this embodiment, the statistical characteristics of latent variables are introduced into the design of the entropy model to reduce the difference between the distribution of the entropy model and the real latent variable. Certain prior knowledge is required to improve the accuracy of the entropy model. Here, a conditional model is constructed by introducing additional variables. The calculation formula is as follows:

表示量化后的潜层表示，

表示条件熵模型，

表示额外变量

的分布，作为熵模型的先验信息，

表示潜层表示的真实分布。in,

represents the latent layer representation after quantization,

represents the conditional entropy model,

represents an extra variable

The distribution of , as the prior information of the entropy model,

represents the true distribution represented by the latent layer.

In the design of the entropy model, the statistical prior represented by the latent layer is added, and the convolutional neural network is used to solve the parameters:

Among them, f represents a distribution that can describe the statistical characteristics of the latent layer. If the Gaussian characteristic is obvious, it can be expressed as a Gaussian distribution. If the non-Gaussian characteristic is obvious, T distribution, Laplace distribution, etc. can be selected; the selection of f is determined by It is determined by the statistical properties of the latent layer representation. The parameters of f are learned by the convolutional neural network, that is, on the premise of determining the type of f, the parameter information of the f distribution is learned by using variables.

Because the distribution of latent variables has obvious non-Gaussian characteristics after the hyperspectral image is nonlinearly transformed by the convolutional neural network, this prior information needs to be added in the design of the entropy model. During the experiment, it was found that the t distribution can be very good To capture this property, the t-distribution was chosen to model the latent variables of hyperspectral images.

In order to make the whole compression process differentiable, the quantization process adopts the additive unit uniform noise approximation; in order to make the entropy model fit the posterior distribution more closely, in the design of the entropy model, a unit uniform distribution is convolved, and the formula is as follows:

通过超先验网络求的。Among them, η _i represents the scale parameter of the t distribution (similar to variance but not equal to the variance), and v represents the degree of freedom, which can adjust the shape of the t distribution. c represents the analytical form of the probability distribution of the entropy model. Parametric variables of probability distributions take advantage of additional variables

obtained through a super-prior network.

Entropy coding adopts arithmetic coding. In the process of arithmetic coding, the entropy model provides probability distribution for the process of arithmetic coding and arithmetic decoding. According to the probability distribution of each element in the latent variable, the size of the code word allocated to it (that is, how many symbols each symbol occupies) is determined. bit), after entropy encoding, the latent variable becomes a binary code stream for storage or transmission.

Step S20, using the test set to verify the compression performance of the trained convolutional neural network, when the compression performance of the trained convolutional neural network reaches the standard, compress the hyperspectral image through the trained convolutional neural network.

In this embodiment, the uniform noise approximate quantization process is used during training, and the rounding method is directly used during testing. The entropy coding adopts the commonly used arithmetic coding, which minimizes the rate-distortion loss and trains the model until convergence. When testing, the entire large image is directly put in. On the CAVE data set, when the degree of freedom is 21, the bppp is 0.1219, the PSNR is 36.74dB, the SSIM is 0.9175, and the SAM is 0.2137. When the degree of freedom is 20, on the KAIST dataset, when the bppp is 0.0885, the PSNR is 39.99dB, the SSIM is 0.9524, and the SAM is 0.2331.

If the user needs to use the image information, the binary code stream can be restored to latent variables through arithmetic decoding, and then input into the inverse transformation network composed of two spatial and spectral modules. In the inverse transformation network, the spatial and spectral modules are connected by IGDN. Sampling is restored to the original image size. The compression framework is divided into four parts, including transform network, quantization, entropy coding, and inverse transform network.

In this embodiment, in view of the anisotropy of hyperspectral images, a spatial and spectral convolution module (SS module, including SS module_down (for constructing an encoding network) and SS module_up (for constructing a decoding network)) is proposed, and GDN is used in the middle. connect. In SS module_down, for an image tensor (B*H*W) whose spectral dimension is B, the first layer is down-sampled and then input a 5*5*B filter to generate N feature representations; after GDN, one more time Downsampling, input a 5*5*N convolutional layer, generate B feature representations, and then go through a 1*1*B convolutional layer to generate N feature representations. The process of SS module_up is similar to SSmodule_down, but the downsampling is replaced by upsampling. The spectral dimension of the hyperspectral image was introduced into the network during the design of the spectral module to realize the rearrangement of spectral information, thereby reducing the correlation between spectra. In view of the non-Gaussian characteristics of the latent layer representation of hyperspectral images, the assumption of traditional Gaussian distribution is no longer used in the design of the entropy model, and some non-Gaussian distributions are introduced as the statistical prior of latent layer representation to improve the entropy model and latent layer representation. The fit of the statistical distribution. According to the fitting of the latent variables of the hyperspectral datasets CAVE and KAIST datasets, it is found that the t distribution performs well, and the shape of the distribution can be flexibly changed by adjusting the degrees of freedom. When the degrees of freedom tend to be infinite, the t distribution and the Gaussian distribution can be equivalence. This property enables the t-distribution to capture both the non-Gaussian properties of the latent layer representation and the generality of the Gaussian distribution.

In this embodiment, the convolutional neural network is trained by the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model; the compression performance of the trained convolutional neural network is verified by using the test set. When the compression performance of the trained convolutional neural network reaches the standard, the hyperspectral image is compressed by the trained convolutional neural network. This embodiment has better rate-distortion performance for hyperspectral image compression.

In a third aspect, an embodiment of the present invention further provides a hyperspectral image compression device.

In an embodiment, referring to FIG. 3 , FIG. 3 is a schematic diagram of functional modules of an embodiment of a hyperspectral image compression apparatus according to the present invention. As shown in Figure 3, the hyperspectral image compression device includes:

The training module 10 is used for training the convolutional neural network through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model;

The processing module 20 is used to verify the compression performance of the trained convolutional neural network using the test set. When the compression performance of the trained convolutional neural network reaches the standard, the hyperspectral image is compressed by the trained convolutional neural network. .

Further, in one embodiment, the hyperspectral image compression device further includes a building block for:

Divide the sample hyperspectral image into several fixed-size cubes in the spatial dimension;

The several fixed-size cubes are divided into a training set and a test set according to a preset ratio.

Further, in one embodiment, the nonlinear transformation module performs forward nonlinear transformation on the spatial and spectral dimensions of the hyperspectral image to obtain latent variables; the quantization module quantifies the latent variables by adding uniform noise; the entropy model is used for A probability distribution of latent variables is obtained for determining a codeword assigned to each element in the latent variable based on the probability distribution during entropy coding.

Further, in one embodiment, the training process of the convolutional neural network is constrained based on the rate-distortion criterion, which is used to determine the nonlinear transformation module and parameter values in the entropy model.

其中，

表示输入的高光谱图像，

表示重构图像，H,W,B分别对应高光谱图像的行、列和波段数，

表示潜变量，h,w,N分别对应潜变量的行、列和滤波器的个数，

和

表示正变换网络参数，

和

表示反变换网络参数，g_a(.)表示非线性正向变换函数，g_s(.)表示非线性反向变换函数。Further, in an embodiment, the nonlinear transformation includes a forward transformation: Y=ga (W _a X+b _a ), and _a reverse transformation:

in,

represents the input hyperspectral image,

represents the reconstructed image, H, W, B correspond to the number of rows, columns and bands of the hyperspectral image, respectively,

Represents latent variables, h, w, N correspond to the number of rows, columns and filters of latent variables, respectively,

and

represents the positive transformation network parameters,

and

represents the inverse transformation network parameters, ga (.) represents the nonlinear forward transformation function, and _{g s} (.) represents the nonlinear reverse transformation function.

Further, in an embodiment, the function for quantizing latent variables by adding uniform noise is expressed as follows:

表示单位均匀噪声，round表示取整操作，

表示量化后的潜向量。Among them, training represents the training process, testing represents the testing process,

represents unit uniform noise, round represents the rounding operation,

represents the quantized latent vector.

Further, in one embodiment, the statistical characteristics of the latent variables are introduced into the design of the entropy model, and at the same time, an additional variable is introduced to construct a conditional model, so as to improve the accuracy of the entropy model.

The function implementation of each module in the above-mentioned hyperspectral image compression apparatus corresponds to each step in the above-mentioned embodiment of the hyperspectral image compression method, and the functions and implementation process thereof will not be repeated here.

In a fourth aspect, an embodiment of the present invention further provides a readable storage medium.

The readable storage medium of the present invention stores a hyperspectral image compression program, wherein when the hyperspectral image compression program is executed by the processor, the steps of the hyperspectral image compression method described above are implemented.

The method implemented when the hyperspectral image compression program is executed may refer to the various embodiments of the hyperspectral image compression method of the present invention, which will not be repeated here.

It should be noted that, herein, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or system comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or system. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, method, article or system that includes the element.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on such understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a storage medium (such as ROM/RAM) as described above. , magnetic disk, optical disk), including several instructions to make a terminal device execute the method described in each embodiment of the present invention.

The above are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields , are similarly included in the scope of patent protection of the present invention.

Claims (10)

1. A hyperspectral image compression method, wherein the hyperspectral image compression method comprises: The convolutional neural network is trained through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model; Use the test set to verify the compression performance of the trained convolutional neural network. When the compression performance of the trained convolutional neural network reaches the standard, the hyperspectral image is compressed by the trained convolutional neural network. 2. The hyperspectral image compression method according to claim 1, wherein, before the step of training the convolutional neural network through the training set, further comprising: Divide the sample hyperspectral image into several fixed-size cubes in the spatial dimension; The several fixed-size cubes are divided into a training set and a test set according to a preset ratio. 3. The hyperspectral image compression method according to claim 2, wherein the nonlinear transformation module performs forward nonlinear transformation on the space and spectral dimensions of the hyperspectral image to obtain latent variables; The latent variable is quantified by the method; the entropy model is used to obtain the probability distribution of the latent variable, so as to determine the codeword allocated to each element in the latent variable based on the probability distribution during entropy coding. 4 . The hyperspectral image compression method according to claim 3 , wherein the training process of the convolutional neural network is constrained based on the rate-distortion criterion, which is used to determine the parameter values in the nonlinear transformation module and the entropy model. 5 . 5. The hyperspectral image compression method according to claim 4, wherein the nonlinear transformation comprises a forward transformation: Y=g _a (W _a X+b _a ), and a reverse transformation:

in,

represents the input hyperspectral image,

represents the reconstructed image, H, W, B correspond to the number of rows, columns and bands of the hyperspectral image, respectively,

Represents latent variables, h, w, N correspond to the number of rows, columns and filters of latent variables, respectively,

and

represents the positive transformation network parameters,

and

represents the inverse transformation network parameters, ga (.) represents the nonlinear forward transformation function, and _{g s} (.) represents the nonlinear reverse transformation function. 6. The hyperspectral image compression method as claimed in claim 5, wherein the function for quantizing latent variables by adding uniform noise is expressed as follows:

Among them, training represents the training process, testing represents the testing process,

represents unit uniform noise, round represents the rounding operation,

represents the quantized latent vector. 7 . The hyperspectral image compression method according to claim 6 , wherein statistical characteristics of latent variables are introduced into the design of the entropy model, and additional variables are introduced to construct a conditional model to improve the accuracy of the entropy model. 8 . 8. A hyperspectral image compression device, wherein the hyperspectral image compression device comprises: a training module for training the convolutional neural network through the training set, wherein the convolutional neural network includes a nonlinear transformation module, a quantization module and an entropy model; The processing module is used to verify the compression performance of the trained convolutional neural network using the test set. When the compression performance of the trained convolutional neural network reaches the standard, the hyperspectral image is compressed by the trained convolutional neural network. 9. A hyperspectral image compression device, characterized in that the hyperspectral image compression device comprises a processor, a memory, and a hyperspectral image compression program stored on the memory and executable by the processor, wherein When the hyperspectral image compression program is executed by the processor, the steps of the hyperspectral image compression method according to any one of claims 1 to 7 are implemented. 10. A readable storage medium, characterized in that, a hyperspectral image compression program is stored on the readable storage medium, and when the hyperspectral image compression program is executed by a processor, the program as claimed in claims 1 to 7 is implemented. The steps of any one of the hyperspectral image compression methods.

Images