A spectral reflectance image acquisition method and system based on intrinsic decomposition

Abstract

The present invention relates to a spectral reflectance image acquisition method and system based on intrinsic decomposition, and belongs to the field of computer vision technology. First, spectral image detail gradient extraction is performed to generate detail gradient information of spectral intrinsic image. Then, the spectral image neighborhood autocorrelation matrix is ​​extracted. According to the spectral image detail gradient image and the neighborhood autocorrelation matrix, a spectral image intrinsic decomposition model is constructed, and iterative optimization is performed using a self-supervised learning method. The system includes a spectral image detail gradient image extraction subsystem, a spectral image neighborhood autocorrelation matrix extraction subsystem, a spectral image intrinsic decomposition self-supervised learning subsystem, a loss function calculation subsystem, and a generation result evaluation subsystem. The present invention realizes the generation of spectral reflectance images from pure noise based on unsupervised learning by introducing relevant detail gradient generation and spectral image autocorrelation constraints, and solves the problems of spectral images with the same object but different spectra, and the same spectrum but different objects.

Description

Technical Field

[0001] The invention relates to a method and system for acquiring a spectral reflectance image based on intrinsic decomposition, and belongs to the technical field of computer vision. Background Art

[0002] With the advancement of imaging technology, spectral imaging, as an advanced technology, has attracted widespread attention because it can provide rich spectral information for each pixel. This technology has been widely used in various fields due to its detailed spatial and spectral information. However, spectral image processing faces challenges such as mixed pixels, different spectra for the same object, and different objects for the same spectrum, which limits its further application. In response to these challenges, some researchers have proposed using the method of intrinsic decomposition to extract spectral reflectance images, aiming to overcome these complex phenomena in spectral images and remove the influence of atmospheric shadows and illumination on spectral images.

[0003] At present, the methods for obtaining spectral reflectance images through intrinsic decomposition mainly include methods based on model optimization and methods combined with deep learning optimization results. Although these methods have their own advantages, they all rely on the optimization quality of the model and the convergence of the optimization process, which directly affect the final decomposition effect.

[0004] In view of this, in order to improve the quality of spectral imaging, solve the problem of different spectra of the same object and different objects of the same spectrum, and optimize the quality of spectral reflectance images, it is urgent to develop a new spectral reflectance acquisition method. It should be able to automatically produce precise image details and still generate high-quality spectral reflectance images when processing complex actual spectral images, thereby improving the application potential and efficiency of spectral images. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings and defects of the prior art and creatively propose a method and system for acquiring spectral reflectance images based on intrinsic decomposition. The present invention does not need to use any model optimization method for spectral image intrinsic decomposition. By introducing relevant detail gradient generation and spectral image autocorrelation constraints, the generation of spectral reflectance images from pure noise is achieved based on unsupervised learning, solving the problems of different spectra for the same object and different objects for the same spectrum in spectral images.

[0006] The present invention is implemented by adopting the following technical solutions.

[0007] A method for acquiring a spectral reflectance image based on intrinsic decomposition comprises the following steps:

[0008] Step 1: Extract the detail gradient of the spectral image to generate the detail gradient information of the spectral intrinsic image.

[0009] Specifically, the non-differentiable problem of discrete variables is converted into a differentiable problem of continuous variables using tangible distribution. The tangible distribution variables of the gradient image are iteratively optimized using the similarity measurement method of spectral angles and discreteness constraints, and finally a spectral image detail gradient image is generated. .

[0010] Step 2: Extraction of spectral image neighborhood autocorrelation matrix.

[0011] Specifically, based on the characteristic that the intrinsic image of the spectral image can be linearly represented by the neighborhood, combined with the autocorrelation of the spectral image and the reflectance image, the spectral angle distance calculation method is used to calculate the spectral image neighborhood similarity matrix .

[0012] Step 3: Self-supervised learning of spectral image intrinsic decomposition.

[0013] According to the spectral image detail gradient image and neighborhood autocorrelation matrix, a spectral image intrinsic decomposition model is constructed, and self-supervised learning method is used for iterative optimization.

[0014] Specifically, step 3 includes the following steps:

[0015] Step 3.1: Generate a noise image of the same size as the spectral image of the reflectance image to be extracted.

[0016] Specifically, the spectral image is a three-dimensional array of size h*w*c, where h, w, and c are the length, width, and number of spectra of the spectral image, respectively.

[0017] From the normal distribution Randomly sample 2*h*w*c numbers from and combine them into two 3D arrays of size h*w*c as noise images.

[0018] Step 3.2: Construct a deep learning model for eigendecomposition of spectral images, generate corresponding reflectance images and illumination images, combine them with the spectral image gradient image and the neighborhood similarity matrix, calculate the loss function, and update the model parameters in the deep learning model for eigendecomposition of spectral images accordingly.

[0019] The specific implementation method is as follows:

[0020] Select the spectral image to be decomposed, generate a noise image that meets the requirements according to step 3.1, and input the two noise images into the spectral image intrinsic decomposition deep learning model.

[0021] Then, the deep learning model processes the two noise images and outputs the corresponding spectral reflectance image and spectral illumination image. The length, width and spectrum number of the spectral reflectance image and spectral illumination image output by the spectral intrinsic decomposition model are consistent with the input noise image.

[0022] Reflectance image generation, whose input is a noise image and output is a spectral image reflectance image The reflectance image generation adopts a fully convolutional network, which is based on translation invariance and generative autocorrelation. The convolutional layers, pooling layers and activation functions are all operated on the local input area and only depend on the relevant spatial coordinates.

[0023] Specifically, the fully convolutional network for reflectance image generation has parameters in each layer that are three-dimensional arrays of size h*w*d, where h and w are the spatial dimensions and d is the number of channels. The first layer is an image with an image size of h*w and d spectral channels.

[0024] The position in a specific layer where the intrinsic image is generated ( i, j ) is the vector X ij , after calculation, y ij :

[0025]

[0026] in, f ks It determines the layer type, including convolution, average pooling, maximum pooling, and activation function; Indicates the feature graph involved in the calculation In the location A vector of k is the convolution kernel size; s is the stride of the convolution operation; , For the value in All integers indicate that the vectors involved in the calculation are is the upper left vertex, and the side length is k All vectors within the convolution kernel.

[0027] Similarly, illumination image generation, whose input is a noise image, outputs a spectral illumination image .

[0028] The loss function used in the spectral image intrinsic decomposition model includes three parts: intrinsic image reconstruction constraint, reflectance image gradient constraint, and reflectance image neighborhood constraint.

[0029] ; Among them, the intrinsic image reconstruction constraint for:

[0030]

[0031] in, is the spectral reflectance image, is the spectral illumination image, is a spectral image. The intrinsic image reconstruction constraint enables the generated reflectance image and illumination image to satisfy the basic intrinsic decomposition assumption.

[0032] Reflectivity Image Gradient Constraint for:

[0033]

[0034] in, is the spectral reflectance image, To generate the spectral image detail gradient image, is the first-order gradient operator of the image, is the multiplication of the corresponding elements of the matrix. The reflectance image gradient constraint can save the corresponding gradient on the original spectral image in the corresponding reflectance image, and remove other image gradients to achieve local smoothing of the intrinsic image.

[0035] Reflectivity image neighborhood constraints for:

[0036]

[0037] in, is the generated neighborhood similarity matrix, is the spectral image reflectance image; the intrinsic image neighborhood constraint can maintain the autocorrelation of the spectral intrinsic image and satisfy the autocorrelation assumption of the spectral intrinsic image.

[0038] Step 3.3: Repeat step 3.2 until the set termination condition is met, and store the intrinsic image and shadow image output by the model.

[0039] On the other hand, based on the above method, the present invention proposes a spectral reflectance image acquisition system based on intrinsic decomposition, including a spectral image detail gradient image extraction subsystem, a spectral image neighborhood autocorrelation matrix extraction subsystem, a spectral image intrinsic decomposition self-supervised learning subsystem, a loss function calculation subsystem, and a generated result evaluation subsystem.

[0040] The spectral image detail gradient image extraction subsystem uses tangible distribution to convert the discrete variable non-differentiable problem into a continuous variable differentiable problem, uses the spectral angle similarity measurement method and discreteness constraints to iteratively optimize the tangible distribution variables of the gradient image, and finally generates the spectral image detail gradient image. , and constrain the subsequent model generation results.

[0041] The spectral image neighborhood autocorrelation matrix extraction subsystem calculates the spectral image neighborhood similarity matrix using the spectral angle based on the characteristics that the spectral image reflectance image can be linearly represented by the neighborhood, combined with the autocorrelation of the spectral image and the reflectance image. , to realize the constraint on the autocorrelation of spectral image.

[0042] The spectral image intrinsic decomposition self-supervised learning subsystem includes a noise generation module, a spectral reflectance image generation module and a spectral illumination image generation module. The noise generation module generates a noise image of corresponding size according to the spectral image; the input of the spectral reflectance image generation module is the output noise of the noise generation module, and a high-quality spectral reflectance image is generated according to the network parameters; the input of the spectral illumination image generation module is the output noise of the noise generation module, and a high-quality spectral illumination image is generated according to the network parameters.

[0043] The loss function calculation subsystem includes an intrinsic image reconstruction constraint module, a reflectance image gradient constraint module and a reflectance image neighborhood constraint module. Among them, the intrinsic image reconstruction constraint module calculates the loss of the generated reflectance image and illumination image with the original spectral image; the reflectance image gradient constraint module calculates the loss of the generated intrinsic image with the output result of the gradient image extraction subsystem; the reflectance image neighborhood constraint module calculates the loss of the generated intrinsic image with the output result of the neighborhood correlation matrix generation subsystem. All losses will be back-propagated to update the parameters of each neural network layer in the spectral intrinsic decomposition subsystem and correct the spectral reflectance image and spectral light and shadow image finally generated by the system.

[0044] The generation result evaluation subsystem is used to evaluate the generation effect of the spectral reflectance image. It can automatically calculate the objective evaluation indicators of the spectral intrinsic image such as the spectral intrinsic image object classification accuracy and Kappa coefficient according to the spectral real object classification. As a supplement to the user's subjective feeling, the user can quantitatively understand the quality of the reflectance image generated by the system.

[0045] The connection relationship between the above components is:

[0046] The spectral image detail gradient image extraction subsystem is connected to the loss function calculation subsystem, and provides the spectral gradient image to the loss function calculation subsystem;

[0047] The spectral image neighborhood autocorrelation matrix extraction subsystem is connected to the loss function calculation subsystem, and provides the neighborhood similarity matrix to the loss function calculation subsystem;

[0048] The spectral image intrinsic decomposition self-supervised learning subsystem is connected to the loss function calculation subsystem and the generated result evaluation subsystem, provides the generated spectral intrinsic image and spectral shadow image to the loss function calculation subsystem, and receives the loss value returned by the loss function calculation subsystem for training itself;

[0049] The loss function calculation subsystem obtains the generated intrinsic image and shadow image from the spectral image intrinsic decomposition self-supervised learning subsystem, calculates the loss value according to the loss constraint designed by the loss function calculation subsystem, and then transmits the loss value back to the spectral image intrinsic decomposition self-supervised learning subsystem for training;

[0050] The generated result evaluation subsystem is connected to the spectral image intrinsic decomposition self-supervised learning subsystem. It obtains the generated reflectance image from the spectral image intrinsic decomposition self-supervised learning subsystem and calculates the classification accuracy to help users quantitatively understand the quality of the reflectance image converted and generated by the system. Beneficial Effects

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. The present invention replaces the traditional method of using the spectral image intrinsic decomposition model with a deep learning method, uses the spectral intrinsic image and spectral shadow image to generate a network, and implements a deep learning method for spectral image intrinsic decomposition, thereby improving the scope of application and stability of the spectral image intrinsic decomposition algorithm.

[0053] 2. The present invention extracts important gradient information from the spectral image by using a tangible learning gradient image extraction module, and performs constraints on the spectral intrinsic image to achieve key gradient maintenance on the intrinsic image, thereby avoiding the loss of excessive image details during the intrinsic decomposition process.

[0054] 3. The present invention uses an autocorrelation constraint based on the neighborhood correlation of the spectral image, extracts the neighborhood parameter matrix through the spectral angle, and performs an autocorrelation constraint on the intrinsic image to maintain the autocorrelation on the spectral image and its intrinsic image. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of the method of the present invention.

[0056] Figure 2 It is a schematic diagram of training the core algorithm model described in the method of the present invention.

[0057] Figure 3 It is a schematic diagram of the composition of the system of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0059] Example

[0060] like Figure 1 As shown, a method for acquiring a spectral reflectance image based on intrinsic decomposition comprises the following steps:

[0061] Step 1: Extract spectral image detail gradient based on tangible distribution and generate spectral detail gradient image .

[0062] Specifically, the mathematical representation of the detail gradient image is as follows:

[0063]

[0064] in, represents the detail gradient image, In pixels Is there a detail gradient in the position image? If Description in pixels There is a detail gradient in the position image, otherwise there is no detail gradient. Represents the set of all pixels.

[0065] To solve the differentiable problem of Equation 1, use the tangible distribution , the discrete variable Convert to continuous variable ; The mathematical expression of tangible distribution is as follows:

[0066]

[0067] in, represents a distribution variable; represents uniform distribution; is the transformed continuous variable, The larger the pixel The probability of The smaller the pixel The greater the probability; , is an intermediate variable; is the Sigmoid activation function; is a parameter that controls the shape of the visible branch. The original distribution degenerates into Bernoulli distribution and cannot be differentiated. When , the distribution is a continuous distribution similar to the Bernoulli distribution; in this example ; represents the lower bound of the distribution; Represents the upper bound of the distribution.

[0068] About Optimization The loss function for:

[0069]

[0070] in, Represents the weight of the loss function; To avoid retaining too much gradient information in the gradient image, it is a sparse constraint. It is a key gradient constraint that forces the model gradient image to retain pixels with larger gradients. is the spectral angle gradient parameter, and the calculation method is as follows:

[0071]

[0072] in, Pixel The neighborhood of , in this example, the neighborhood size is 3x3; Pixel The neighborhood size; arccos is the inverse cosine function; represents the spectral angle calculation formula; T Indicates matrix transformation; Representing images In pixels The pixel value of Represents an image Represents an image Represents an image In pixels The pixel value of .

[0073] Afterwards, As the objective function, the target variable is iteratively optimized to generate the final spectral image gradient image and store it.

[0074] Step 2: Extraction of the neighborhood autocorrelation matrix of the spectral image. Based on the characteristic that the intrinsic image of the spectral image can be linearly represented by the neighborhood, combined with the autocorrelation of the spectral image and the reflectance image, the spectral angle distance calculation method is used to calculate the neighborhood similarity matrix of the spectral image. .

[0075] Specifically, the neighborhood correlation matrix The calculation method is:

[0076] in, For the matrix exist OK The value of the column; and Represents the spectral image and pixels, and satisfy , i.e. pixel In pixels In the neighborhood of; represents the similarity coefficient; is an exponential function; Represents the inverse cosine function.

[0077] Calculate the neighborhood correlation matrix according to Formula 5 , and store.

[0078] Step 3: Self-supervised learning of spectral image intrinsic decomposition.

[0079] like Figure 2 As shown, according to the spectral image detail gradient image and the neighborhood autocorrelation matrix, a spectral image intrinsic decomposition model is constructed, and a self-supervised learning method is used for iterative optimization.

[0080] Specifically, step 3 includes the following steps:

[0081] Step 3.1: Generate a noise image of the same size as the spectral image of the reflectance image to be extracted.

[0082] Specifically, the spectral image is a three-dimensional array of size h*w*c, where h, w, and c are the length, width, and number of spectra of the spectral image, respectively.

[0083] Then, from the normal distribution Randomly sample 2*h*w*c numbers from and combine them into two 3D arrays of size h*w*c as noise images.

[0084] Step 3.2: Construct a deep learning model for eigendecomposition of spectral images to generate corresponding reflectance images and illumination images; then, combine the spectral image detail gradient image and the neighborhood similarity matrix, calculate the loss function, and update the model parameters in the deep learning model for eigendecomposition of spectral images accordingly.

[0085] The specific implementation method is as follows:

[0086] Select the spectral image to be decomposed, generate a noise image that meets the requirements according to step 3.1, and input the two noise images into the spectral image intrinsic decomposition deep learning model.

[0087] Then, the deep learning model processes the two noise images and outputs the corresponding spectral reflectance image and spectral illumination image. The length, width and spectrum number of the spectral reflectance image and spectral illumination image output by the spectral intrinsic decomposition model are consistent with the input noise image.

[0088] Reflectance image generation, whose input is a noise image and output is a spectral image reflectance image The reflectance image generation adopts a fully convolutional network, which is based on translation invariance and generative autocorrelation. The convolutional layers, pooling layers and activation functions are all operated on the local input area and only depend on the relevant spatial coordinates.

[0089] Specifically, the fully convolutional network for reflectance image generation has parameters in each layer that are three-dimensional arrays of size h*w*d, where h and w are the spatial dimensions and d is the number of channels. The first layer is an image with an image size of h*w and d spectral channels.

[0090] The position in a specific layer where the intrinsic image is generated ( i, j ) is the vector X ij , after calculation, :

[0091]

[0092] in, f ks It determines the layer type, including convolution, average pooling, maximum pooling, and activation function; Indicates the feature graph involved in the calculation In the location A vector of k is the convolution kernel size; s is the stride of the convolution operation; , For the value in All integers indicate that the vectors involved in the calculation are is the upper left vertex, and the side length is k All vectors within the convolution kernel.

[0093] Similarly, illumination image generation, whose input is a noise image, outputs a spectral illumination image .

[0094] The loss function used in the spectral image intrinsic decomposition model includes three parts: intrinsic image reconstruction constraint, reflectance image gradient constraint, and reflectance image neighborhood constraint.

[0095] Among them, the intrinsic image reconstruction constraint for:

[0096]

[0097] in, is the spectral reflectance image, is the spectral illumination image, is a spectral image. The intrinsic image reconstruction constraint enables the generated reflectance image and illumination image to satisfy the basic intrinsic decomposition assumption.

[0098] Reflectivity Image Gradient Constraint for:

[0099]

[0100] in, is the spectral reflectance image, To generate the spectral image detail gradient image, is the first-order gradient operator of the image, is the multiplication of the corresponding elements of the matrix. The reflectance image gradient constraint can save the corresponding gradient on the original spectral image in the corresponding reflectance image, and remove other image gradients to achieve local smoothing of the intrinsic image.

[0101] Reflectivity image neighborhood constraints for:

[0102]

[0103] in, is the generated neighborhood similarity matrix, is the spectral image reflectance image; the intrinsic image neighborhood constraint can maintain the autocorrelation of the spectral intrinsic image and satisfy the autocorrelation assumption of the spectral intrinsic image.

[0104] Step 3.3: Repeat step 3.2 until the set termination condition is met, and store the intrinsic image and shadow image output by the model.

[0105] Furthermore, if Figure 3 As shown, a spectral reflectance image acquisition system based on intrinsic decomposition according to an embodiment of the present invention includes a spectral image detail gradient image extraction subsystem, a neighborhood correlation matrix generation subsystem, a spectral image intrinsic decomposition subsystem, a loss function calculation subsystem, and a generation result evaluation subsystem.

[0106] Among them, the spectral image detail gradient image extraction subsystem generates the spectral image detail gradient image , and constrain the subsequent model generation results.

[0107] The neighborhood correlation matrix generation subsystem combines the autocorrelations of the spectral image and the reflectance image to calculate the neighborhood similarity matrix of the spectral image. , to realize the constraint on the autocorrelation of spectral image.

[0108] The spectral image intrinsic decomposition subsystem includes a noise generation module, a spectral intrinsic image generation module and a spectral shadow image generation module. The noise generation module generates a noise image of corresponding size according to the spectral image; the input of the spectral intrinsic image generation module is the output noise of the noise generation module, and a high-quality spectral intrinsic image is generated according to the network parameters; the input of the spectral shadow image generation module is the output noise of the noise generation module, and a high-quality spectral shadow image is generated according to the network parameters.

[0109] The loss function calculation subsystem includes an intrinsic image reconstruction constraint module, an intrinsic image gradient constraint module and an intrinsic image neighborhood constraint module. Among them, the intrinsic image reconstruction constraint module calculates the loss of the generated intrinsic image and shadow image with the original spectral image; the intrinsic image gradient constraint module calculates the loss of the generated intrinsic image with the output result of the gradient image extraction subsystem; the intrinsic image neighborhood constraint subsystem calculates the loss of the generated intrinsic image with the output result of the neighborhood correlation matrix generation subsystem. All losses will be back-propagated to update the parameters of each neural network layer in the spectral intrinsic decomposition subsystem and correct the spectral eigenimage and spectral shadow image finally generated by the system.

[0110] The generation result evaluation subsystem is used to evaluate the generation effect of the spectral intrinsic image. It can classify the real objects according to the spectrum, and as a supplement to the user's subjective feelings, it helps the user to quantitatively understand the quality of the intrinsic image generated by the system.

Claims

1. A spectral reflectance image acquisition method based on intrinsic decomposition, characterized in that: The generation of relevant detail gradients and spectral image autocorrelation constraints are introduced to realize the generation of spectral reflectance images from pure noise based on unsupervised learning, including the following steps: Step 1: Extract the detail gradient of the spectral image to generate the detail gradient information of the spectral intrinsic image; Use tangible distribution to convert the non-differentiable problem of discrete variables into a differentiable problem of continuous variables; Use the similarity measurement method of spectral angles and discreteness constraints to iteratively optimize the tangible distribution variables of the gradient image, and finally generate the spectral image detail gradient image ; Step 2: Extraction of spectral image neighborhood autocorrelation matrix; Based on the characteristic that the spectral image intrinsic image can be linearly represented by the neighborhood, combined with the autocorrelation of the spectral image and the reflectance image, the spectral angular distance calculation method is used to calculate the spectral image neighborhood similarity matrix ; Step 3: Self-supervised learning of spectral image intrinsic decomposition; According to the spectral image detail gradient image and the neighborhood autocorrelation matrix, the spectral image intrinsic decomposition model is constructed, and the self-supervised learning method is used for iterative optimization; Step 3.1: According to the size of the spectral image of the reflectance image to be extracted, a noise image of the same size is generated; Step 3.2: Construct a spectral image intrinsic decomposition deep learning model, generate the corresponding reflectance image and illumination image, combine them with the spectral image gradient image and the neighborhood similarity matrix, calculate the loss function, and update the model parameters in the spectral image intrinsic decomposition deep learning model accordingly; Step 3.3: Repeat step 3.2 until the set termination condition is met, and store the intrinsic image and shadow image output by the model.

2. The method for acquiring spectral reflectance images based on intrinsic decomposition according to claim 1, characterized in that: In step 3.1, let the spectral image be a three-dimensional array of size h*w*c, where h, w, and c are the length, width, and number of spectra of the spectral image, respectively; from the normal distribution Randomly sample 2*h*w*c numbers from and combine them into two 3D arrays of size h*w*c as noise images.

3. The method for acquiring spectral reflectance images based on intrinsic decomposition according to claim 1, characterized in that: In step 3.2, select the spectral image to be decomposed, generate a noise image that meets the requirements according to step 3.1, and input the two noise images into the spectral image intrinsic decomposition deep learning model; then, the deep learning model processes the two noise images and outputs the corresponding spectral reflectance image and spectral illumination image; wherein, the length, width and spectral number of the spectral reflectance image and spectral illumination image output by the spectral intrinsic decomposition model are consistent with the input noise image; the reflectance image is generated, and its input is the noise image, and the output is the spectral image reflectance image ; The reflectivity image is generated using a fully convolutional network, which is based on translation invariance and generative autocorrelation. The convolutional layers, pooling layers, and activation functions are all operated on the local input area and only depend on the relevant spatial coordinates.

4. The method for acquiring spectral reflectance images based on intrinsic decomposition according to claim 3, characterized in that: The fully convolutional network for reflectance image generation has parameters in each layer that are three-dimensional arrays of size h*w*d, where h and w are the spatial dimensions and d is the number of channels; the first layer is an image with size h*w and d spectral channels; let the position in a specific layer for intrinsic image generation ( i,j ) is the vector X ij , after calculation, y ij : in, f ks It determines the layer type, including convolution, average pooling, maximum pooling, and activation function; Indicates the feature graph involved in the calculation In the location A vector of k is the convolution kernel size; s is the stride of the convolution operation; , For the value in All integers indicate that the vectors involved in the calculation are is the upper left vertex, and the side length is k All vectors in the convolution kernel of ; illumination image generation, its input is a noise image, and the output is a spectral illumination image .

5. The method for acquiring spectral reflectance images based on intrinsic decomposition according to claim 3, characterized in that: The loss function used in the spectral image intrinsic decomposition model consists of three parts: intrinsic image reconstruction constraint, reflectance image gradient constraint, and reflectance image neighborhood constraint. for: in, is the spectral reflectance image, is the spectral illumination image, is a spectral image; the intrinsic image reconstruction constraint enables the generated reflectance image and illumination image to satisfy the basic intrinsic decomposition assumption; the reflectance image gradient constraint for: in, is the spectral reflectance image, To generate the spectral image detail gradient image, is the first-order gradient operator of the image, is the multiplication of the corresponding elements of the matrix; the reflectance image gradient constraint can save the corresponding gradient on the original spectral image in the corresponding reflectance image, and remove other image gradients to achieve local smoothing of the intrinsic image; the reflectance image neighborhood constraint for: in, is the generated neighborhood similarity matrix, is the spectral image reflectance image; the intrinsic image neighborhood constraint can maintain the autocorrelation of the spectral intrinsic image and satisfy the autocorrelation assumption of the spectral intrinsic image.

6. The method for acquiring spectral reflectance images based on intrinsic decomposition according to claim 1, characterized in that: In step 1, the mathematical representation of the detail gradient image is as follows: in, represents the detail gradient image, In pixels Is there a detail gradient in the position image? If Description in pixels There is a detail gradient in the position image, otherwise there is no detail gradient; Represents the set of all pixels; uses a tangible distribution , the discrete variable Convert to continuous variable ; The mathematical expression of tangible distribution is as follows: in, represents a distribution variable; represents uniform distribution; is the transformed continuous variable, The larger the pixel The probability of The smaller the pixel The greater the probability; , is an intermediate variable; is the Sigmoid activation function; is a parameter that controls the shape of the visible branch. The original distribution degenerates into Bernoulli distribution and cannot be differentiated. When , the distribution is a continuous distribution similar to the Bernoulli distribution; represents the lower bound of the distribution; represents the upper bound of the distribution; optimization The loss function for: in, Represents the weight of the loss function; To avoid retaining too much gradient information in the gradient image, it is a sparse constraint. It is a key gradient constraint that forces the model gradient image to retain pixels with larger gradients. is the spectral angle gradient parameter, and the calculation method is as follows: in, Pixel Neighborhood of; Pixel The neighborhood size; arccos is the inverse cosine function; represents the spectral angle calculation formula; T Indicates matrix transformation; Representing images In pixels The pixel value of Represents an image Represents an image Represents an image In pixels ; then, As the objective function, the target variable is iteratively optimized to generate the final spectral image gradient image and store it.

7. The method for acquiring spectral reflectance images based on intrinsic decomposition according to claim 1, characterized in that: In step 2, the neighborhood correlation matrix The calculation method is: in, For the matrix exist OK The value of the column; and Represents the spectral image and pixels, and satisfy , i.e. pixel In pixels In the neighborhood of; represents the similarity coefficient; is an exponential function; Represents the inverse cosine function.

8. A spectral reflectance image acquisition system based on intrinsic decomposition for implementing the spectral reflectance image acquisition method based on intrinsic decomposition as claimed in claim 1, characterized in that: It includes a spectral image detail gradient image extraction subsystem, a spectral image neighborhood autocorrelation matrix extraction subsystem, a spectral image intrinsic decomposition self-supervised learning subsystem, a loss function calculation subsystem, and a generation result evaluation subsystem; among them, the spectral image detail gradient image extraction subsystem uses tangible distribution to convert the discrete variable non-differentiable problem into a continuous variable differentiable problem, uses the spectral angle similarity measurement method and discreteness constraints to iteratively optimize the tangible distribution variables of the gradient image, and finally generates a spectral image detail gradient image. , and constrain the subsequent model generation results; the spectral image neighborhood autocorrelation matrix extraction subsystem, based on the characteristics that the spectral image reflectance image can be linearly represented by the neighborhood, combines the autocorrelation of the spectral image and the reflectance image, and uses the spectral angle to calculate the spectral image neighborhood similarity matrix. , realizing the constraint on the autocorrelation of spectral images; the spectral image intrinsic decomposition self-supervised learning subsystem, including a noise generation module, a spectral reflectance image generation module and a spectral illumination image generation module; wherein the noise generation module generates a noise image of corresponding size according to the spectral image; the input of the spectral reflectance image generation module is the output noise of the noise generation module, and a high-quality spectral reflectance image is generated according to the network parameters; the input of the spectral illumination image generation module is the output noise of the noise generation module, and a high-quality spectral illumination image is generated according to the network parameters; the loss function calculation subsystem, including the intrinsic image reconstruction constraint module, the reflectance image gradient approximation module, The beam module and the reflectivity image neighborhood constraint module; the intrinsic image reconstruction constraint module calculates the loss of the generated reflectivity image and illumination image with the original spectral image; the reflectivity image gradient constraint module calculates the loss of the generated intrinsic image and the gradient image extraction subsystem output result; the reflectivity image neighborhood constraint module calculates the loss of the generated intrinsic image and the neighborhood correlation matrix generation subsystem output result, all losses will be back-propagated to update the parameters of each neural network layer in the spectral intrinsic decomposition subsystem, and correct the spectral reflectivity image and spectral light and shadow image finally generated by the system; the generation result evaluation subsystem is used to evaluate the spectral reflectivity image. The generation effect can automatically calculate the spectral intrinsic image object classification accuracy and Kappa coefficient spectral intrinsic image objective evaluation index according to the spectral real object classification; among them: the spectral image detail gradient image extraction subsystem is connected to the loss function calculation subsystem, and provides the spectral gradient image to the loss function calculation subsystem; the spectral image neighborhood autocorrelation matrix extraction subsystem is connected to the loss function calculation subsystem, and provides the neighborhood similarity matrix to the loss function calculation subsystem; the spectral image intrinsic decomposition self-supervised learning subsystem is connected to the loss function calculation subsystem and the generation result evaluation subsystem, and provides the generated spectral intrinsic image and spectrum to the loss function calculation subsystem. The loss function calculation subsystem obtains the generated intrinsic image and shadow image from the spectral image intrinsic decomposition self-supervised learning subsystem, and calculates the loss value according to the loss constraint designed by the loss function calculation subsystem. The loss value is then returned to the spectral image intrinsic decomposition self-supervised learning subsystem for training. The generated result evaluation subsystem is connected to the spectral image intrinsic decomposition self-supervised learning subsystem, which obtains the generated reflectance image from the spectral image intrinsic decomposition self-supervised learning subsystem, and calculates the classification accuracy to help users quantitatively understand the quality of the reflectance image converted and generated by the system.

Images