Spectral reflectivity rapid reconstruction method for small-scale samples

Abstract

The invention discloses a spectral reflectivity rapid reconstruction method for small-scale samples, which comprises the following steps: collecting small-scale training samples, and generating an enhanced sample set through a data enhancement method; a Wiener estimation method is adopted to quickly solve a low-precision predicted value of system spectral characteristics, and a high-precision spectral reflectivity reconstruction result is obtained through the nonlinear mapping capability of the multi-layer perceptron neural network; a spectral image response matrix of a test sample is collected through a multispectral imaging device, and the spectral reflectivity of a to-be-tested sample is calculated and obtained through a two-stage spectral reflection reconstruction model of a Wiener estimation-multilayer perceptron neural network for prediction. According to the method, the dependence of a traditional spectral reflectivity reconstruction model on large-scale sample data is effectively overcome, the data acquisition and cost are reduced, the spectral reflectivity reconstruction efficiency and accuracy of the model under small-scale sample data are improved, and a technical support is provided for the practical application of a low-cost and miniaturized multi-spectral imager.

Description

Technical Field

[0001] This invention relates to the field of spectral imaging technology, and in particular to a method for rapid reconstruction of spectral reflectance for small-scale samples. Background Technology

[0002] Spectral reflectance is a crucial parameter describing the optical properties of an object's surface, providing vital information for computer vision and target attribute detection. Traditional spectrophotometers can measure fine spectral features within a small range, but they struggle to meet the real-time measurement requirements of spectral reflectance for large, complex structures. In recent years, spectral reflectance reconstruction technology has emerged as a method for calculating high-dimensional spectral reflectance information of a photographed object using low-dimensional response signals output from simple imaging devices. This method offers advantages such as low cost, high efficiency, and wide applicability, making it a research hotspot in the field of machine vision inspection.

[0003] Common methods for spectral reflectance reconstruction estimate spectral features by directly solving the inverse model of the imaging system, including pseudo-inverse methods, Wiener estimation, PCA principal component analysis, and sparse dictionary methods. However, these methods are susceptible to high reconstruction errors due to actual imaging noise. With the continuous development of machine learning and deep learning technologies, BP neural networks, RBF neural networks, and convolutional neural network models have been increasingly applied to spectral reflectance reconstruction, enabling the prediction of high-dimensional spectral information from multi-channel spectral data and effectively improving spectral reconstruction accuracy. However, neural networks typically exhibit strong learning performance only with a large number of training samples; when training samples are scarce, convolutional neural networks may experience overfitting. Summary of the Invention

[0004] To overcome the aforementioned problems in the existing technology, this invention proposes a method for rapid reconstruction of spectral reflectance for small-scale samples.

[0005] The technical solution adopted by this invention to solve its technical problem is: a method for rapid reconstruction of spectral reflectance for small-scale samples, comprising the following steps: Step 1: Collect multispectral response images of a small number of samples, obtain the spectral reflectance data of the samples, and form a training set; Step 2: Amplify the training set samples obtained in Step 1 to generate an enhanced training set, and process the enhanced training set to form a standardized training set; Step 3: Construct a linear reconstruction model based on Wiener estimation and calculate the low-precision spectral reflectance matrix of the training set samples; Step 4: Construct a spectral reflectance reconstruction compensation model based on a multilayer perceptron neural network, and train the model to obtain the reconstruction model; Step 5: Acquire the multispectral response image of the sample to be tested, obtain the standardized test set according to the method described in Step 2, input the standardized test set into the linear reconstruction model based on Wiener estimation obtained in Step 3 to obtain the initial prediction result, input the initial prediction result into the reconstruction model obtained in Step 4 to obtain the spectral reflectance reconstruction result.

[0006] The above-mentioned method for rapid reconstruction of spectral reflectance for small-scale samples, specifically the enhanced training set generation process in step 2, involves: using data augmentation methods to increase the sample size of the training set; using spectral image response interpolation methods, spectral imaging digital simulation methods, and image noise reduction methods, combined with spectral reflectance smoothing constraints, to generate physically reasonable and spatially spectrally consistent augmented samples; and then fusing the original samples to generate the enhanced training set.

[0007] The above-mentioned method for fast reconstruction of spectral reflectance for small-scale samples, in step 3, the linear reconstruction model takes the standardized training set sample response matrix obtained in step 2 as input and the spectral reflectance matrix of the training set sample as output to calculate the low-precision spectral reflectance matrix of the training set sample. The Wiener estimation rule is used to find an estimated multi-channel response matrix. G Reduce the original spectral reflectance R With the reconstructed spectral reflectance Minimize the mean square error between them.

[0008] The above-mentioned method for fast reconstruction of spectral reflectance for small-scale samples, in step 4, uses the low-precision spectral reflectance matrix of the training set samples obtained in step 3 as input and the spectral reflectance of the real samples as output. The input layer and the output layer are connected through one or more hidden layers. Each unit in the input layer is multiplied by the interconnection weight of the node, and the interconnection weight is continuously adjusted during the training iteration to minimize the error between the predicted reflectance and the true value.

[0009] The above-mentioned method for fast reconstruction of spectral reflectance for small-scale samples uses a particle swarm optimization algorithm to optimize the weights and bias parameters of each layer during the training process of the spectral reflectance reconstruction compensation model based on a multilayer perceptron neural network, and introduces a linear inertia factor reduction method in the particle swarm optimization algorithm.

[0010] The aforementioned method for fast reconstruction of spectral reflectance for small-scale samples involves a particle swarm optimization algorithm that uses weights and bias parameters as optimization variables. The algorithm generates position vectors for the particle swarm through encoding. For each particle, a solution is calculated, and a linear matrix reconstruction model and a nonlinear error compensation model are superimposed to obtain the reconstructed spectral reflectance. The actual spectral reflectance value is compared with the reconstruction result to calculate the fitness value of each particle, updating the global and individual extreme values. The inertia weights and learning factors are updated, along with the velocity and position of particles in the swarm. An iterative update is determined based on a termination condition; if the termination condition is met, the current globally optimal particle position vector is saved. The optimal particle position vector is decoded to generate the optimal structural parameters for the multilayer perceptron neural model, terminating the particle swarm optimization algorithm.

[0011] The beneficial effects of this invention are that it proposes a method for expanding the scale of small-scale samples based on data augmentation methods. By using spectral image response interpolation methods, spectral imaging digital simulation methods, image noise reduction processing methods, etc., and combined with spectral reflectance smoothing constraints, it generates physically reasonable and spatially spectrally consistent augmented samples, and fuses the original samples to generate new training samples. This mechanism helps to enhance the diversity of training samples and improve the generalization of spectral reflectance reconstruction.

[0012] This invention proposes a two-stage spectral reflectance reconstruction model based on Wiener estimation and a multilayer perceptron neural network. The Wiener estimation method can quickly solve for low-precision predictions of the system's spectral features, while the nonlinear mapping capability of the multilayer perceptron neural network can obtain high-precision spectral reflectance reconstruction results. By integrating the complementary advantages of linear regression model estimation and nonlinear fitting of the neural network, the efficiency and accuracy of spectral reflectance reconstruction are improved. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the process of this invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] This invention provides a fast method for reconstructing spectral reflectance for small-scale samples. By combining the complementary advantages of linear Wiener estimation and nonlinear multilayer perceptron neural network, it breaks through the dependence of traditional spectral reflectance reconstruction methods on large samples and achieves efficient and high-precision spectral reflectance reconstruction in small-scale sample data scenarios.

[0016] like Figure 1As shown, the specific steps are as follows: Step 1, use a multispectral imaging device to acquire multispectral response images of a small number of samples, and use a spectrophotometer to obtain the spectral reflectance data corresponding to the samples to form a training set.

[0017] In this embodiment, a 24-color standard color chart can be selected as the measurement sample. The spectral response range of the multispectral imaging device covers 400nm to 1000nm, and no fewer than 16 spectral channels are selected for imaging. In this embodiment, 16 spectral channels are selected for the multispectral imaging device. Image processing technology is used to identify and extract the multispectral response value of each color patch sample on the standard color chart, forming a 24×16 training set response matrix. The spectral reflectance data of each color patch sample is obtained using a spectrophotometer. In this embodiment, the measured spectral resolution is 10nm, forming a 24×61 training set spectral reflectance matrix.

[0018] Step 2: The sample size of the training set is increased by using data augmentation methods such as spectral image response interpolation, spectral imaging digital simulation, and image noise reduction, combined with spectral reflectance smoothing constraints, to generate physically reasonable and spatially spectral consistent augmented samples. The original samples are then fused to generate an augmented training set.

[0019] The spectral image linear interpolation method generates new sample image responses and spectral features by linearly interpolating the image responses and spectral curves of adjacent spectral channels of existing samples.

[0020] The spectral imaging digital simulation method is based on modeling and simulating the physical parameters of the spectral imaging system. By comparing the measured spectrum with the theoretical model, new samples are generated in terms of image response and spectral features.

[0021] The image noise-adding processing method involves adding Gaussian noise or salt-and-pepper noise to the response image of an existing sample using image processing techniques, thereby generating a new sample image response and spectral features that are affected by background noise transformation.

[0022] In this embodiment, 8 new samples are generated by linear interpolation of spectral images, 24 new samples are generated by image simulation based on spectral response curves, and 32 new samples are generated by image noise addition and spectral reflectance smoothing constraint methods. Physically reasonable and spatially spectrally consistent amplified samples are generated, and the original samples are fused to generate an enhanced training set containing a total of 80 samples.

[0023] Step 3: Perform data cleaning and normalization on the augmented training set to form a standardized training set.

[0024] The data cleaning process involves cleaning small-scale sample data and using data cleaning algorithms to identify and correct erroneous, missing, and duplicate values ​​in the data.

[0025] Normalization: Unifying the cleaned data with different features to the same numerical range.

[0026] Step 4: Construct a linear reconstruction model based on Wiener estimation, using the response matrix of the standardized training set samples as input and the spectral reflectance matrix of the training set samples as output, to calculate the low-precision spectral reflectance matrix of the training set samples.

[0027] The Wiener estimation rule is used to find an estimate of the multi-channel response matrix. G Reduce the original spectral reflectance R With the reconstructed spectral reflectance Minimizing the mean square error between them is expressed as: in, Q The camera's response matrix, and These represent the autocorrelation matrices of reflectivity and noise, respectively.

[0028] In this embodiment, the input training set has 80×16 channel response matrices, and the output is an 80×61 spectral reflectance matrix. The residual matrix is ​​obtained by subtracting the initial linear reconstruction result from the spectral reflectance matrix of the real sample.

[0029] Step 5: Construct a spectral reflectance reconstruction compensation model based on a multilayer perceptron neural network. This model uses the low-precision spectral reflectance matrix of the training set samples as the input layer and the spectral reflectance of the real samples as the output layer. The input layer and the output layer are connected through one or more hidden layers.

[0030] The input layer selects a hidden layer for connection. The number of hidden layer nodes is as follows: Where m is the number of input layer nodes, n is the number of output layer nodes, and α is an integer between 1 and 10. In this embodiment, m is 61 and n is 61, so the range of the number of hidden layer nodes is 12 to 21.

[0031] Each unit in the input layer is multiplied by the interconnection weights of the nodes; for the , i Hidden nodes x i Activate and the previous layer j The neuron in the previous layer i Weights of the interconnections between neurons w ij , bias is b j .

[0032] Hierarchical connections have weight attributes, and node functionality can perform both summation and activation functions. The summation function is... In the formula: n is the amount of input data, x i For input data i, b j w is the bias parameter. ij Interconnection weights; The output obtained by using the activation function in the hidden layer is Then the output of layer unit j y j The value can be represented as: During training, multilayer perceptron neural networks need to continuously adjust the weights during the iterative process to minimize the error between the predicted reflectance and the true value.

[0033] Step 6: During the training of the multilayer perceptron neural network, the particle swarm optimization algorithm is used to optimize the weights and bias parameters of each layer of the multilayer perceptron neural network model, overcoming the problem that the gradient descent method is prone to generating local optima when updating weights and biases.

[0034] The particle swarm optimization algorithm treats each individual particle in the swarm as a separate particle. During flight in the search space, each particle has a fitness value relative to its environment and its own position, velocity, and direction. In the iterative process, to find the optimal solution, particles continuously adjust their individual and global extrema, correcting their flight direction and velocity. Consider a... D In the dimensional search space, there are a total of N If _ examples form a group, then the _ ... i The position of each particle is x i And the speed is v i Then after k The position and velocity after flight can be expressed as: (5) (6) The linear inertia factor reduction method introduced in the particle swarm optimization algorithm addresses the "premature convergence" problem in the standard particle swarm optimization algorithm during the solution process. The formula for the linear inertia factor reduction is as follows: The fitness function g(x) of the particle swarm optimization algorithm uses the spectral reflectance of M training samples. and true reflectivity RThe root mean square error is calculated, and its mathematical formula can be described as follows: The particle swarm optimization algorithm uses the weights and bias parameters of the multilayer perceptron neural network as optimization variables, and generates the position vector of the particle swarm through encoding. For each particle, the solution is calculated, and a linear matrix reconstruction model and a nonlinear error compensation model are superimposed to obtain the reconstructed spectral reflectance. The actual spectral reflectance value is compared with the reconstruction result to calculate the fitness value of each particle, updating the global and individual extreme values. The inertia weights and learning factors are updated respectively, as well as the velocity and position of particles in the swarm. An iterative update is determined based on the termination condition; if the condition is met, the current globally optimal particle position vector is saved. The optimal particle position vector is decoded to generate the optimal structural parameters of the multilayer perceptron neural model, terminating the particle swarm optimization algorithm.

[0035] In this embodiment, the population size of the particle swarm optimization algorithm is 30, the learning factor of the particle swarm optimization algorithm is set to 2.0, the initial value of the inertia factor is set to 1.6, the inertia factor value after the optimal number of iterations is set to 0.6, and the maximum number of iterations is set to 200.

[0036] Step 7: Use a multispectral imaging device to acquire multispectral response images of the sample to be tested, and perform data cleaning and normalization on the test set according to the method in Step 3 to form a standardized test set.

[0037] In this embodiment, a multispectral response image of one sample is acquired using a multispectral imaging device. According to the method in step 3, the test set is cleaned and normalized to form a standardized test set. In this embodiment, the response matrix of the standardized test is 1×16.

[0038] Step 8: Input the standardized test set into the initial prediction result of the test set spectral reflectance calculated by the Wiener estimation model; then input the initial prediction result of the test set spectral reflectance into the multilayer perceptron neural network to calculate the predicted spectral reflectance of the sample to be tested.

[0039] The standardized test set response matrix is ​​input into the Wiener estimation model to calculate the 1×61 low-precision spectral reflectance matrix of the test set; then the low-precision spectral reflectance matrix of the test set is input into the multilayer perceptron neural network to calculate the 1×61 high-precision spectral reflectance matrix prediction result of the sample to be tested.

[0040] Compared with existing technologies, this embodiment discloses a method for rapid reconstruction of spectral reflectance for small-scale sample data. First, multispectral image data and spectral reflectance data of small-scale samples are collected as small-scale sample data. Data augmentation methods are used to expand the sample size of the training set. After data cleaning and normalization, a standardized training set is formed. Then, a Wiener estimation model is constructed to provide initial predicted values ​​of spectral reflectance, and a multilayer perceptron neural network model is trained to obtain corrected high-precision spectral reflectance predictions. Finally, the trained Wiener estimation model and multilayer perceptron neural network are used to predict spectral reflectance on test set samples.

[0041] This embodiment effectively overcomes the dependence of traditional deep learning spectral reflectance reconstruction models on large-scale sample data. It reduces data acquisition and processing costs through data augmentation methods and leverages the advantages of efficient computation using the Wiener estimation method and high-precision nonlinear fitting using multilayer perceptron neural networks. This improves the computational efficiency and accuracy of the spectral reflectance reconstruction model with small-scale sample data, providing a reliable and efficient solution for the application of multispectral imaging testing technology in fields with limited sample size.

[0042] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its scope and spirit, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A method for rapid reconstruction of spectral reflectance for small-scale samples, characterized in that, Includes the following steps: Step 1: Collect multispectral response images of a small number of samples, obtain the spectral reflectance data of the samples, and form a training set; Step 2: Amplify the training set samples obtained in Step 1 to generate an enhanced training set, and process the enhanced training set to form a standardized training set; Step 3: Construct a linear reconstruction model based on Wiener estimation and calculate the low-precision spectral reflectance matrix of the training set samples; Step 4: Construct a spectral reflectance reconstruction compensation model based on a multilayer perceptron neural network, and train the model to obtain the reconstruction model; Step 5: Acquire the multispectral response image of the sample to be tested, obtain the standardized test set according to the method described in Step 2, input the standardized test set into the linear reconstruction model based on Wiener estimation obtained in Step 3 to obtain the initial prediction result, input the initial prediction result into the reconstruction model obtained in Step 4 to obtain the spectral reflectance reconstruction result.

2. The method for rapid reconstruction of spectral reflectance for small-scale samples according to claim 1, characterized in that, The enhanced training set generation process in step 2 is as follows: the training set sample size is increased by using data augmentation methods. Through spectral image response interpolation, spectral imaging digital simulation, and image noise reduction processing, combined with spectral reflectance smoothing constraints, physically reasonable and spatially spectrally consistent augmented samples are generated. The original samples are then fused to generate the enhanced training set.

3. The method for rapid reconstruction of spectral reflectance for small-scale samples according to claim 1, characterized in that, In step 3, the linear reconstruction model takes the standardized training set sample response matrix obtained in step 2 as input and the spectral reflectance matrix of the training set sample as output to calculate the low-precision spectral reflectance matrix of the training set sample. The Wiener estimation rule is used to find an estimated multi-channel response matrix. G Reduce the original spectral reflectance R With the reconstructed spectral reflectance Minimize the mean square error between them.

4. The method for rapid reconstruction of spectral reflectance for small-scale samples according to claim 1, characterized in that, In step 4, the spectral reflectance reconstruction compensation model based on the multilayer perceptron neural network takes the low-precision spectral reflectance matrix of the training set samples obtained in step 3 as input and the spectral reflectance of the real samples as output. The input layer and the output layer are connected through one or more hidden layers. Each unit in the input layer is multiplied by the interconnection weight of the node, and the interconnection weight is continuously adjusted during the training iteration to minimize the error between the predicted reflectance and the real value.

5. The method for rapid reconstruction of spectral reflectance for small-scale samples according to claim 4, characterized in that, During the training process, the spectral reflectance reconstruction compensation model based on multilayer perceptron neural network uses particle swarm optimization to optimize the weights and bias parameters of each layer, and introduces a linear inertia factor reduction method into the particle swarm optimization algorithm.

6. The method for rapid reconstruction of spectral reflectance for small-scale samples according to claim 5, characterized in that, The particle swarm optimization algorithm uses weights and bias parameters as optimization variables, and generates the position vector of the particle swarm through encoding. The solution for each particle is calculated, and a linear matrix reconstruction model and a nonlinear error compensation model are superimposed to obtain the reconstructed spectral reflectance. The fitness value of each particle is calculated by comparing the actual value of the spectral reflectance with the reconstruction result, and the global extremum and individual extremum are updated. Update the inertia weights and learning factors respectively, and update the velocity and position of particles in the swarm; determine whether iterative updates are needed based on the termination condition. If the termination condition is met, save the current globally optimal particle position vector; decode the optimal particle position vector to generate the optimal structural parameters of the multilayer perceptron neural model, and end the particle swarm optimization algorithm.

Images