Multi-channel spectral image prediction method and system based on hmsa-bp neural network and storage medium

Abstract

This invention discloses a multi-channel spectral image prediction method, system, and storage medium based on HMSSA-BP neural network. The method includes: First, constructing a BP neural network prediction model (BPNN) from the theoretical dot area ratios of C, M, Y, and K to multi-channel spectral maps using multi-source multispectral imaging technology. Second, optimizing the BPNN using the Hybrid Multi-Strategy Sparrow Search (HMSSA) algorithm, which introduces Tent mapping, a staged control step size strategy, and a chaotic cosine variation factor based on SSA. Third, introducing an adjustment factor Q and proposing a multi-channel spectral map fusion algorithm to obtain the predicted post-printing color image. This invention, based on multi-source multispectral imaging technology, constructs a high-precision multi-channel spectral prediction model to predict the post-printing image before actual printing, ensuring the color quality and consistency of the final printed product.

Description

Technical Field

[0001] This invention relates to the field of spectral prediction technology, specifically to a multi-channel spectral image prediction method, system, and storage medium based on an HMSSA-BP neural network. Background Technology

[0002] Color prediction models for printed materials aim to construct a mapping model between printing equipment control information and post-printing color information, allowing for a preview of the printed effect and ensuring the final printed product meets expectations. Currently, the mainstream color prediction models in printing include the Murray-Davies model and the Neugebauer model. However, due to nonlinear limitations, they do not fully consider actual coloring conditions or the complex interaction between ink and paper, leading to prediction bias. The YNSN model and the Clapper-Yule model also suffer from limitations in practical application and versatility due to neglecting optical interactions, parameter constraints, and the need for data adjustment. These mainstream models are built under ideal conditions, while the high nonlinearity of printing equipment and the complexity of printing conditions in reality require extensive modifications to the physical models and a large number of training samples, resulting in complex models and low computational efficiency. Therefore, using traditional prediction models for printed image prediction will inevitably result in errors and low prediction accuracy.

[0003] To maximize the accuracy of the printed matter prediction model and obtain a post-printing prediction image that most closely approximates the actual printed image, this invention proposes a new multi-channel spectral prediction model for printed matter based on a neural network and employing multi-source multispectral imaging technology. Simultaneously, an algorithm for synthesizing multi-channel spectral images is also proposed. Summary of the Invention

[0004] This invention addresses the shortcomings of existing spectral prediction models by proposing a multi-channel spectral image prediction method, system, and storage medium based on an HMSSA-BP neural network.

[0005] The present invention adopts the following technical solution:

[0006] A multi-channel spectral image prediction method based on HMSSA-BP neural network is disclosed. The method includes: using a multi-source multispectral imaging system, optimizing the BPNN-based prediction model from the theoretical dot area ratio of C, M, Y, and K to the multi-channel spectral map using the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA); the HMSSA algorithm combines Tent mapping, a staged control step size strategy, and a chaotic cosine variation factor on the basis of the SSA algorithm; and further introducing an adjustment factor Q on the basis of the optimization to fuse the predicted multi-channel spectral maps and obtain the predicted post-printing color image.

[0007] The above technical solution further includes the following steps:

[0008] The first step is to use a multi-source multispectral imaging system to form a multi-channel spectral prediction model based on BPNN, including setting the model input and output, model topology, loss function, and evaluation index.

[0009] The second step involves iterative optimization using the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA) to optimize the initial weights and thresholds of the BPNN model.

[0010] The third step is to fuse the multi-channel spectra after prediction. This includes first introducing an adjustment factor Q to convert the multi-channel spectral values ​​output by the prediction model to the XYZ color space; then converting them to the RGB space to obtain the fused color prediction image.

[0011] Furthermore, the BPNN-based prediction model for converting theoretical dot area ratios of C, M, Y, and K into multi-channel spectra is specifically as follows: the dot area ratios of C, M, Y, and K are used as input variables of the spectral prediction model, and the multi-channel spectra obtained by the multi-source multispectral imaging system are used as output variables of the model, wherein the dot area ratio is the proportion of dots per unit area in the ideal state where the ink has not diffused.

[0012] Furthermore, the topology of the BPNN-based prediction model is as follows:

[0013] The input layer has 4 nodes, and the output layer has N nodes, which is the number of channels in the multi-source multispectral imaging system. The number of hidden layer nodes is then determined by selecting each hidden layer node number within a preset range for training. The error component between the expected output and the actual output after training is calculated, and the hidden layer node number with the smallest L2 norm of the error component is taken as the optimal number of hidden layer nodes. The model uses the mean squared error as the loss function for model training, which is the average of the squares of the deviations of the predicted data values ​​from the measured data values.

[0014] Furthermore, the mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) between the standard output values ​​on the test set and the training output values ​​are used to measure the predictive performance of the prediction model, serving as a direct standard for evaluating the model's prediction accuracy. The NY model, Ham model, and LMS-MI model are used as evaluation indices for model metamerism to assess the degree of metamerism.

[0015] Let one spectral curve be R1 and the other spectral curve be R2. R1(λ) and R2(λ) are the spectral information of R1 and R2 at wavelength λ, respectively. S(λ) is the relative spectral energy corresponding to wavelength λ. x(λ), y(λ), and z(λ) are the matching functions of three standard observers.

[0016] The NY model is... Among them, M y M z These are the results after weighting the spectral differences according to the standard observer matching functions x(λ), y(λ), and z(λ), respectively.

[0017] The Ham model is Where S(λ) is the normalized spectral power distribution under a given illumination source D65, and n is the number of samples;

[0018] The LMS-MI model is Where x lms (λ), y lms (λ), z lms (λ) represent the three standard observer matching functions x(λ), y(λ), and z(λ), respectively. Thus, a metamerism index containing three-dimensional information is calculated.

[0019] Furthermore, the initial weights and thresholds of the BPNN model are iteratively optimized using the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA); the specific steps include:

[0020] 2.1) Based on the Sparrow Search Algorithm (SSA), a Tent mapping is introduced:

[0021]

[0022] Generate a chaotic sequence to optimize the initial position of the sparrow population in SSA, where x n The value of the current iteration belongs to the interval [0,1], x n+1 The value for the next iteration is given, and mu∈(0,2] is the chaos parameter, which is proportional to the chaos.

[0023] 2.2) In the discoverer position of the SSA algorithm, a phased step size control strategy is introduced, and a nonlinear decay factor μ is added, resulting in:

[0024]

[0025]

[0026] l=(a-1)·rand+1

[0027]

[0028] Where ST′∈[0.8,1] represents the safe value, rand is a random number in [0,1], and ω is a constant; This represents the j-th dimension value of the i-th sparrow in the t-th iteration; Iter represents the globally optimal position in the t-th iteration. max R2 represents the maximum number of iterations; R2 represents the alarm value, and the range of R2 is [0,1]; V represents a random number that follows a normal distribution; L represents a 1×d matrix with all elements being 1; d is the dimension of the variable to be optimized; and a is an intermediate variable.

[0029] 2.3) In the joiner position update of the SSA algorithm, a chaotic cosine variation factor and an inertia weight η that varies with the number of iterations are introduced; therefore:

[0030]

[0031] in: δ is a constant, k is a random number between 0 and 1 generated by the Circle mapping, and cos(2πk) is the introduced chaotic cosine variation factor. This represents the position of the global worst-case scenario in the t-th iteration. Represents the optimal position of the global discoverer in the (t+1)th iteration; A is a 1×d matrix where each element is randomly -1 or 1. + =A T (AA T ) -1 , where n represents the population size and α represents a uniform random number in the range (0,1).

[0032] Furthermore, the multi-channel spectra output by the prediction model are fused, specifically including the following:

[0033] Based on the principle of XYZ tristimulus values:

[0034]

[0035]

[0036]

[0037] Where: S(λ) is the relative spectral power distribution of the lighting source, R(λ) represents the spectral reflectance information, and λ represents the visible light wavelength in the range of a1~b1nm. This represents the human visual matching function, where X, Y, and Z are CIE tristimulus values, and k is an adjustment factor, calculated with Y taking a value of 100.

[0038] The multi-channel spectral values ​​output by the prediction model are light intensity, given by φ(λ) = R(λ)S(λ), where φ(λ) is the light intensity value. During the actual multi-channel fusion process, normalization is performed, and an adjustment factor Q is introduced. The processed light intensity, i.e., the multi-channel spectral value, is then:

[0039]

[0040] Then convert the multi-channel spectral values ​​to the XYZ color space:

[0041]

[0042]

[0043]

[0044] Here, the adjustment factor

[0045] λ i This corresponds to the predicted illumination wavelength of channel i;

[0046] Finally, convert from the XYZ color space to the RGB color space:

[0047]

[0048] The N-channel spectra predicted by the prediction model are then fused into a color image to obtain the predicted post-printing color image.

[0049] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-channel spectral image prediction method based on the HMSSA-BP neural network as described in any of the preceding claims.

[0050] An electronic device, the device comprising:

[0051] One or more processors;

[0052] Memory, used to store one or more programs;

[0053] When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-channel spectral image prediction method based on the HMSSA-BP neural network as described in any of the preceding claims.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] This invention proposes a multi-channel spectral image prediction method based on HMSSA-BP neural network. Based on multi-source multispectral technology, it establishes and optimizes a prediction model from C, M, Y, and K dot area ratios to multi-channel spectral images of printed matter. Compared with other prediction models, the multispectral prediction model proposed in this paper has higher accuracy and better prediction performance.

[0056] This invention proposes a Hybrid Multi-Strategy Sparrow Search (HMSSA) algorithm. Based on SSA, the algorithm introduces techniques such as Tent chaotic mapping, chaotic cosine variation factor and staged step size control to continuously iterate and optimize the initial weights and thresholds of the neural network.

[0057] This invention uses a multispectral prediction model to predict a multichannel spectrum, then introduces light intensity φ(λ) and adjustment factor Q to convert the multichannel spectrum to the XYZ tristimulus value color space, and then to the RGB color space, finally obtaining the predicted post-printing color image with the best fusion effect. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating a multi-channel spectral image prediction method based on an HMSSA-BP neural network.

[0059] Figure 2 A flowchart illustrating the HMSSA algorithm and its optimized BP neural network;

[0060] Figure 3 This represents the process of the loss function decreasing.

[0061] Figure 4 This is the pre-press original image for umbrella;

[0062] Figure 5 The 9-channel spectrum of the predicted umbrella plot;

[0063] Figure 6 The image is the actual printed color image predicted using the method of this invention. Detailed Implementation

[0064] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0065] like Figure 1 As shown, this invention first utilizes a multi-source multispectral imaging system to construct a BPNN-based prediction model from the theoretical dot area ratios of C, M, Y, and K to multi-channel spectral maps. Next, a Hybrid Multi-Strategy Sparrow Search (HMSSA) algorithm is proposed to further optimize the BPNN model. This algorithm combines Tent mapping, a staged control step size strategy, and a chaotic cosine variation factor on the basis of SSA. Finally, an adjustment factor Q is introduced to fuse the predicted multi-channel spectral maps, obtaining the predicted post-printing color image.

[0066] Specifically, the following steps are included:

[0067] The first step is to construct a multi-channel spectral prediction model for printed materials, including the design of model inputs and outputs, model topology, loss function, evaluation metrics, etc.

[0068] The second step involves using the Hybrid Multi-Strategy Sparrow Search (HMSSA) algorithm to iteratively optimize the initial weights and thresholds of the BPNN model.

[0069] The third step is to fuse the multi-channel spectra after prediction, which is roughly divided into two steps. First, an adjustment factor Q is proposed to convert the multi-channel spectral values ​​output by the prediction model to the XYZ color space, and then to the RGB space to obtain the fused color prediction image.

[0070] Furthermore, the specific steps for constructing the multi-channel spectral prediction model for printed matter described in the first step are as follows:

[0071] 1.1) The dot area ratios of C, M, Y, and K are used as input variables for the spectral prediction model. The dot area ratio used here refers to the theoretical dot area ratio, which is the proportion of dots per unit area in an ideal state where the ink has not diffused.

[0072] 1.2) Using the multi-channel spectrum obtained by the multi-source multispectral imaging system as the output variable of the model, according to a specific embodiment of the present invention, the multi-source multispectral imaging system used is the American Mega Vision system, which acquires the spectrum of nine wavelength light sources at 420nm, 450nm, 470nm, 505nm, 530nm, 560nm, 590nm, 630nm and 655nm, i.e., the nine-channel spectrum.

[0073] 1.3) The four colors of ink, C, M, Y and K, are arranged and combined according to different dot area ratios to obtain a large number of color block images. The multispectral imaging system acquires the spectral images of the nine channels of the printed color block images and divides the acquired data into training set and test set.

[0074] 1.4) The model has 4 input layer nodes and 9 output layer nodes. The number of hidden layer nodes is then determined using the empirical formula for hidden layer nodes. Where x represents the number of hidden layer nodes, m and N are the number of input and output layer nodes (i.e., the number of channels), respectively, and α is an adjustment constant between 1 and 10. The range of hidden layer node numbers for the training samples is determined to be [a, b]. Training is performed using each hidden layer node number within this range. The error component between the expected output and the actual output after training is calculated. The hidden layer node number with the smallest L2 norm of the error component is taken as the optimal number of hidden layer nodes. The L2 norm of the error is... Wherein e j The error of the j-th sample is the error between the predicted value and the actual value, and n is the number of samples.

[0075] 1.5) The model uses mean squared error as the loss function for model training, that is, the mean of the sum of squares of the deviations of the predicted data values ​​from the measured data values. Where N represents the number of channels for the spectral value, r i The measured spectral value, r i1 This represents the spectral values ​​predicted by the model;

[0076] 1.6) The mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) between the standard output value of the test set and the training output value provide different perspectives to measure the predictive performance of the model, serving as a direct standard for evaluating the model's prediction accuracy.

[0077] 1.7) The NY model, Ham model, and LMS-MI model are used as evaluation indices for model metamerism to assess the degree of metamerism. Metamerism refers to colors that match under specific lighting or observation conditions, but no longer match when the lighting or observation conditions change.

[0078] 1.8) Let one spectral curve be R1 and the other be R2. R1(λ) and R2(λ) are the spectral information of R1 and R2 at wavelength λ, respectively. S(λ) is the relative spectral energy corresponding to wavelength λ. x(λ), y(λ), and z(λ) are three standard observer matching functions. The NY model is... Wherein M x M y M z These are the results after weighting the spectral differences according to the standard observer matching functions x(λ), y(λ), and z(λ), respectively. The Ham model is... Where S(λ) is the normalized spectral power distribution under a given illumination source D65, and n is the number of samples. The LMS-MI model is... Where x lms (λ), y lms (λ), z lms (λ) represent the three standard observer matching functions x(λ), y(λ), and z(λ), respectively. Thus, a metamerism index containing three-dimensional information is calculated.

[0079] Furthermore, the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA) described in the second step optimizes the BPNN model, such as... Figure 2 As shown, the specific steps are as follows:

[0080] 2.1) Based on the Sparrow Search Algorithm (SSA), the Tent mapping algorithm is introduced. Generate a chaotic sequence to optimize the initial position of the sparrow population in SSA. Where x... n The value of the current iteration belongs to the interval [0,1], x n+1 The value for the next iteration is given, and mu∈(0,2] is the chaos parameter, which is proportional to the chaos.

[0081] 2.2) In the discoverer position of the SSA algorithm, a phased step size control strategy is introduced, and a nonlinear decay factor μ is added, resulting in:

[0082]

[0083]

[0084] l = (a-1)·rnd+1,

[0085]

[0086] Where ST′∈[0.8,1] represents a safe value, rand is a random number in [0,1], and ω is a constant. This represents the j-th dimension value of the i-th sparrow in the t-th iteration; Iter represents the globally optimal position in the t-th iteration. max R2 represents the maximum number of iterations; R2 represents the alarm value, and the range of R2 is [0,1]; V represents a random number that follows a normal distribution; L represents a 1×d matrix with all elements being 1, and d is the dimension of the variable to be optimized.

[0087] 2.3) In the joiner position update of the SSA algorithm, a chaotic cosine variation factor and an inertia weight η that varies with the number of iterations are introduced, resulting in:

[0088]

[0089] The above Where δ is a constant, and k is a random number between 0 and 1 generated by the Circle mapping. This represents the position of the global worst-case scenario in the t-th iteration. Let L represent the optimal position of the global discoverer in the (t+1)th iteration. Let L be a 1×d matrix with all elements being 1, d be the dimension of the variable to be optimized, and A be a 1×d matrix with each element randomly set to -1 or 1. + =A T (AA T ) -1 n represents the population size, α represents a uniformly random number in the range (0,1], and iter maxThis indicates the maximum number of iterations.

[0090] Furthermore, the specific steps of the multi-channel spectral fusion algorithm described in the third step are as follows:

[0091] 3.1) Convert the predicted multichannel spectrum to the XYZ color space. Based on the XYZ tristimulus principle, the... Wherein S(λ) represents the relative spectral power distribution of the illumination source, R(λ) represents the spectral reflectance information, and λ represents the visible light wavelength in the range of 380–780 nm, where a1 is 380 nm and b1 is 780 nm. This represents the human visual matching function, where X, Y, and Z are CIE tristimulus values, and k is an adjustment factor, calculated with Y taking a value of 100.

[0092] 3.2) The multi-channel spectral values ​​predicted in this invention are actually light intensities; therefore, φ(λ) = R(λ)S(λ) is proposed, where φ(λ) is the light intensity value. During the actual multi-channel fusion process, normalization was performed, and an adjustment factor Q was introduced to improve the fusion accuracy.

[0093] 3.3) Finally, the algorithm for converting multi-channel spectral values ​​into the XYZ color space is obtained. Where n is 9, and λ takes values ​​corresponding to the predicted illumination wavelengths of the 9 channels, with corresponding adjustment factors.

[0094] 3.4) Then convert to the RGB color space, the aforementioned

[0095] By combining the above knowledge of colorimetry and computer graphics, the 9-channel spectral images predicted by the model can be fused into a color image to obtain the predicted post-printing color image.

[0096] Example

[0097] In this embodiment, the multi-channel spectral image prediction method based on the HMSSA-BP neural network of the present invention is adopted, and the specific steps are as follows:

[0098] 1) To ensure the richness of the samples and the wide coverage of the color gamut, this embodiment uses multi-primary-color overprinting. The inks of the four colors C, M, Y, and K are arranged and combined according to the dot area ratios of 0%, 2%, 4%, 6%, 8%, 10%, 12%, 14%, 16%, 18%, 20%, 23%, 26%, 29%, 32%, 36%, 40%, 44%, 50%, 56%, 64%, 71%, and 91%, respectively, to form 279,841 color block images;

[0099] 2) In this embodiment, the Mega Vision multi-light source multispectral imaging system from the United States is used to capture the printed color block image in 1). This device is a multispectral imaging system based on LED lighting, which sequentially acquires the spectrum of nine wavelength light sources: 420nm, 450nm, 470nm, 505nm, 530nm, 560nm, 590nm, 630nm, and 655nm, i.e., a 9-channel spectrum.

[0100] 3) Samples were collected using a multi-source multispectral imaging system, resulting in 92*9 spectral images, corresponding to 279,841 data sets, each with 9 dimensions. The theoretical dot area ratios of the color patch images (C, M, Y, K) were used as model inputs, with 4*279,841 data sets as input data. The spectral values ​​of the corresponding 9 channels of the printed color patch images obtained in step 2) were used as outputs, with 9*279,841 data sets as output data. A 4-input, 9-output HMSSA-BP neural network model was established. 90% of the samples were randomly selected as training samples, and 10% as test samples.

[0101] 4) The number of hidden layer nodes is calculated according to the formula. The range of the number of nodes is determined to be [4, 13]. Training is performed using the number of nodes in each hidden layer within this range. The error component between the expected output and the actual output after training is calculated, and the L2 norm of the error component is taken. The minimum number of hidden layer nodes is the optimal number of hidden layer nodes, and the final number of hidden layer nodes is 11.

[0102] 5) The hidden layer activation function is the Tansig function, the output layer activation function is the Purelin function, the loss function is MSE, the number of training iterations is 2000, the learning rate is 0.01, the performance metric is RMSE, the target error is 0.0001, the gradient is 1.00e-06, and the training function is the trainlm function.

[0103] 6) In the HMSSA algorithm, the sparrow population popsize is 80, the dimension dim is 163, and the maximum number of iterations Max_iteration is 40;

[0104] 7) Based on the SSA algorithm, the Tent algorithm traverses and generates a chaotic sequence to uniformly cover the search space as the initial position of the sparrow population; a phased control step size strategy is adopted to improve the position update of the discoverer, and a chaotic cosine change factor and an inertial weight η that varies with the number of iterations are introduced into the position update of the joiner to reduce the risk of the algorithm getting trapped in local optima. This can improve the global and local search capabilities of the algorithm at different stages of iteration, and greatly improve the accuracy of the prediction model.

[0105] 8) The HMSSA algorithm iterates continuously to find the best value. The total number of iterations is 40. The changes in the fitness value of the sparrow during the iteration are shown in Table 1. The optimal position of the sparrow with the best fitness value (fitness = 0.019892) is assigned to the initial weights and threshold of the BP neural network to train the HMSSA-BP model.

[0106] Table 1. HMSSA Sparrow Iterative Fitness Values

[0107]

[0108] 9) Model training begins, using mean squared error (MSE) as the loss function. During training, the loss function continuously decreases, eventually reaching a minimum MSE of 0.00030132. Model training is then complete. Figure 3 As shown; and the fitting effect of the HMSSA-BP model is higher than 0.99, indicating that the model has good generalization ability and high prediction accuracy;

[0109] 10) Test the model using test set data. Validate the model's prediction accuracy using four metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) between the standard output values ​​and training output values ​​of the test samples. Table 2 compares the errors of the four metrics for the HMSSA-BP model, SSA-BP model, and BP model. Compared to the improved BP model using the standard SSA algorithm, the HMSSA algorithm shows a significant reduction in errors across all metrics, and also a substantial decrease compared to the BP model. Therefore, the hybrid multi-strategy sparrow search algorithm demonstrates significant improvement, and the HMSSA-BP model exhibits high prediction accuracy and good prediction performance.

[0110] Table 2. Errors of various indicators for multiple models

[0111]

[0112]

[0113] 11) The NY model, Ham model, and LMS-MI model are used as the metamerism evaluation index for the model. The metamerism index I for the HMSSA-BP model, SSA-BP model, and BP model is also used. NY The average values ​​were 7.938, 12.860, and 30.400, respectively. Ham The average values ​​were 9.334, 14.780, and 34.540, respectively. lms The mean values ​​are 0.025, 0.039, and 0.093, respectively, as shown in Table 3. Therefore, the HMSSA-BP model can predict spectral values ​​that are very close to the actual printed spectrum, and can largely avoid metamerism.

[0114] Table 3 Metamerism Evaluation Index

[0115]

[0116] 12) Finally, a high-precision multi-channel spectral prediction model based on HMSSA-BP is obtained.

[0117] Next, in this embodiment, an umbrella image with a resolution of 9072×12096 will be used as the pre-press master image, such as... Figure 4 As shown, the multi-channel spectral prediction model trained in step 12) is used to predict the umbrella image. The specific steps are as follows:

[0118] 1) Convert the original color digital image of the umbrella image to a halftone image, read the C, M, Y, K dot area ratios within its 8*8 pixel area, and move 4 pixels each time as the model input;

[0119] 2) The HMSSA-BP multi-channel spectral prediction model predicts the spectra of multiple channels after printing, such as... Figure 5 As shown;

[0120] 3) The spectral values ​​predicted in this invention are actually light intensity, i.e., φ(λ). The light intensity values ​​are normalized, and an adjustment factor Q of 150 is introduced. We obtain F(λ);

[0121] 4) According to Where n is 9, and λ takes values ​​corresponding to the predicted illumination wavelengths of the 9 channels, and is an adjustment factor. Convert the multi-channel spectrum to the XYZ color space;

[0122] 5) Then according to Converting to the RGB color space yields the predicted post-printed color image, such as... Figure 6 As shown.

[0123] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-channel spectral image prediction method based on HMSSA-BP neural network, characterized in that, The method includes: using a multi-source multispectral imaging system, and employing the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA) to optimize the BPNN-based prediction model from the theoretical dot area ratio of C, M, Y, and K to the multi-channel spectral map; the HMSSA algorithm is based on the SSA algorithm and combines Tent mapping, a staged control step size strategy, and a chaotic cosine variation factor; based on the optimization, an adjustment factor Q is introduced to fuse the predicted multi-channel spectral maps to obtain the predicted post-printing color image.

2. The multi-channel spectral image prediction method based on HMSSA-BP neural network according to claim 1, characterized in that, Specifically, the following steps are included: The first step is to use a multi-source multispectral imaging system to form a multi-channel spectral prediction model based on BPNN, including setting the model input and output, model topology, loss function, and evaluation index. The second step involves iterative optimization using the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA) to optimize the initial weights and thresholds of the BPNN model. The third step is to fuse the multi-channel spectra after prediction. This includes first introducing an adjustment factor Q to convert the multi-channel spectral values ​​output by the prediction model to the XYZ color space; then converting them to the RGB space to obtain the fused color prediction image.

3. The multi-channel spectral image prediction method based on HMSSA-BP neural network according to claim 1, characterized in that, The BPNN-based prediction model for converting theoretical dot area ratios of C, M, Y, and K into multi-channel spectra is as follows: the dot area ratios of C, M, Y, and K are used as input variables of the spectral prediction model, and the multi-channel spectra obtained by the multi-source multispectral imaging system are used as output variables of the model. The dot area ratio is the proportion of dots per unit area in the ideal state where the ink has not diffused.

4. The multi-channel spectral image prediction method based on HMSSA-BP neural network according to claim 3, characterized in that, The topology of the BPNN-based prediction model is as follows: The input layer has 4 nodes, and the output layer has N nodes, which is the number of channels in the multi-source multispectral imaging system. The number of hidden layer nodes is then determined by selecting each hidden layer node number within a preset range for training. The error component between the expected output and the actual output after training is calculated, and the hidden layer node number with the smallest L2 norm of the error component is taken as the optimal number of hidden layer nodes. The model uses the mean squared error as the loss function for model training, which is the average of the squares of the deviations of the predicted data values ​​from the measured data values.

5. The multi-channel spectral image prediction method based on HMSSA-BP neural network according to claim 3, characterized in that, The predictive performance of the prediction model is measured by the mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE) between the standard output values ​​on the test set and the training output values, serving as a direct standard for evaluating the model's prediction accuracy. The NY model, Ham model, and LMS-MI model are used as evaluation indices for model metamerism to assess the degree of metamerism. Let one spectral curve be R1 and the other spectral curve be R2. R1(λ) and R2(λ) are the spectral information of R1 and R2 at wavelength λ, respectively. S(λ) is the relative spectral energy corresponding to wavelength λ. x(λ), y(λ), and z(λ) are the matching functions of three standard observers. The NY model is... Among them, M y M z These are the results after weighting the spectral differences according to the standard observer matching functions x(λ), y(λ), and z(λ), respectively. The Ham model is Where S(λ) is the normalized spectral power distribution under a given illumination source D65, and n is the number of samples; The LMS-MI model is Where x lms (λ), y lms (λ), z lms (λ) represent the three standard observer matching functions x(λ), y(λ), and z(λ), respectively. Thus, a metamerism index containing three-dimensional information is calculated.

6. The multi-channel spectral image prediction method based on HMSSA-BP neural network according to claim 1, characterized in that, The initial weights and thresholds of the BPNN model are iteratively optimized using the Hybrid Multi-Strategy Sparrow Search Algorithm (HMSSA); the specific steps include: 2.1) Based on the Sparrow Search Algorithm (SSA), a Tent mapping is introduced: Generate a chaotic sequence to optimize the initial position of the sparrow population in SSA, where x n The value of the current iteration belongs to the interval [0,1], x n+1 The value for the next iteration is given, and mu∈(0,2] is the chaos parameter, which is proportional to the chaos. 2.2) In the discoverer position of the SSA algorithm, a phased step size control strategy is introduced, and a nonlinear decay factor μ is added, resulting in: l=(a-1)·rand+1 Where ST′∈[0.8,1] represents the safe value, rand is a random number in [0,1], and ω is a constant; This represents the j-th dimension value of the i-th sparrow in the t-th iteration; Iter represents the globally optimal position in the t-th iteration. max R2 represents the maximum number of iterations; R2 represents the alarm value, and the range of R2 is [0,1]; V represents a random number that follows a normal distribution; L represents a 1×d matrix with all elements being 1; d is the dimension of the variable to be optimized; and a is an intermediate variable. 2.3) In the joiner position update of the SSA algorithm, a chaotic cosine variation factor and an inertia weight η that varies with the number of iterations are introduced; therefore: in: δ is a constant, k is a random number between 0 and 1 generated by the Circle mapping, and cos(2πk) is the introduced chaotic cosine variation factor. This represents the position of the global worst-case scenario in the t-th iteration. Represents the optimal position of the global discoverer in the (t+1)th iteration; A is a 1×d matrix where each element is randomly -1 or 1. + =A T (AA T ) -1 , where n represents the population size and α represents a uniform random number in the range (0,1).

7. The multi-channel spectral image prediction method based on HMSSA-BP neural network according to claim 1, characterized in that, The multi-channel spectra output by the prediction model are fused, specifically including the following: based on the principle of XYZ tristimulus values: Where: S(λ) is the relative spectral power distribution of the lighting source, R(λ) represents the spectral reflectance information, and λ represents the visible light wavelength in the range of a1~b1nm. This represents the human visual matching function, where X, Y, and Z are CIE tristimulus values, and k is an adjustment factor, calculated with Y taking a value of 100. The multi-channel spectral values ​​output by the prediction model are light intensity, given by φ(λ) = R(λ)S(λ), where φ(λ) is the light intensity value. During the actual multi-channel fusion process, normalization is performed, and an adjustment factor Q is introduced. The processed light intensity, i.e., the multi-channel spectral value, is then: Then convert the multi-channel spectral values ​​to the XYZ color space: Here, the adjustment factor λ i This corresponds to the predicted illumination wavelength of channel i; Finally, convert from the XYZ color space to the RGB color space: The N-channel spectra predicted by the prediction model are then fused into a color image to obtain the predicted post-printing color image.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the multi-channel spectral image prediction method based on an HMSSA-BP neural network as described in any one of claims 1-7.

9. An electronic device, characterized in that, The device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the multi-channel spectral image prediction method based on the HMSSA-BP neural network as described in any one of claims 1-7.

Images