An infrared wide-spectrum classification method and system for isomorphism objects of different materials

Abstract

The application provides an infrared wide-spectrum classification method and system for isomorphism objects with different materials, and belongs to the technical field of infrared photoelectric imaging and target classification and identification. The method comprises the following steps: normalizing a one-dimensional discrete spectrum signal to a 0-1 range; decomposing the normalized one-dimensional discrete spectrum signal into an intrinsic mode signal by EMD; obtaining an i*n-sized spectrum signal interpolation matrix by cubic spline interpolation of the intrinsic mode signal; performing point multiplication of the spectrum signal interpolation matrix and a full-spectrum weight matrix to obtain a spectrum feature matrix; and inputting the spectrum feature matrix into a spectrum classification model to obtain a spectrum category. The application can effectively extract real information in the spectrum while suppressing noise in the spectrum, and can guarantee the accuracy of the classification result by using a neural network under the condition of few samples.

Description

Technical Field

[0001] This invention belongs to the field of infrared photoelectric imaging and target classification and recognition technology, and more specifically, relates to an infrared broadband classification method and system for targets of the same shape but different materials. Background Technology

[0002] Any target on the Earth's surface and in the atmosphere will generate spectral characteristics due to the interaction between its own properties and the environment during the process of reflecting, scattering, and transmitting solar radiation and the target's own radiation. Comprehensive and effective utilization of the characteristics in the target's spectral information can effectively improve the performance of target detection and identification.

[0003] Currently, commonly used one-dimensional spectral sequence classification methods include nearest neighbor, logistic regression, Naive Bayes, decision trees, support vector machines, and neural networks. These methods can theoretically achieve the classification of one-dimensional spectral sequences. However, infrared signals have weak energy and are easily interfered with, making spectral preprocessing crucial, directly impacting the quality and application range of the calibration model. Conventional denoising and baseline correction methods, such as wavelet transform, derivative methods, interpolation, and polynomial fitting, require timely adjustments to the algorithm's basis functions, decomposition levels, and signal reconstruction methods based on the characteristics of different spectra. Parameter settings lack prior knowledge, making optimal preprocessing difficult for spectra with sharp peaks and drastic waveform changes, thus affecting the accuracy of classification results and the model's applicability, leading to reduced generalization ability. Furthermore, current spectral acquisition equipment operates only in the 0.4µm–2.5µm band, which is narrow. At room temperature, targets (below 100℃) lack near-infrared and short-wave radiation capabilities, making it difficult to acquire the spectral features determined by the target's inherent characteristics. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an infrared broadband classification method and system for targets of the same shape but different materials. It aims to solve the problem that existing methods are difficult to achieve optimal preprocessing for spectra with sharp peaks and drastic waveform changes, which affects the accuracy of classification results and the applicability of the model, resulting in a reduction in generalization ability.

[0005] To achieve the above objectives, on the one hand, the present invention provides an infrared broadband classification method for targets of the same shape but different materials, comprising the following steps:

[0006] S1: Normalize the one-dimensional discrete spectral signal to the range of 0-1;

[0007] S2: Perform EMD decomposition on the normalized one-dimensional discrete spectral signal to decompose the normalized one-dimensional discrete spectral signal into l intrinsic mode signals (IMFs). l (λ);

[0008] S3: Interpolate the intrinsic mode signals of l spectra using cubic splines to obtain i total intrinsic mode signals for each wavelength, and obtain a spectral signal interpolation matrix of size n×i; where n is the number of one-dimensional discrete spectral signals;

[0009] S4: Calculate the full-spectrum weight matrix based on the spectral signal interpolation matrix and the normalized one-dimensional discrete spectral signal. Multiply the n×i spectral signal interpolation matrix and the full-spectrum weight matrix element by element to obtain the spectral feature matrix.

[0010] S5: Input the spectral feature matrix into the spectral classification model to obtain the spectral category; wherein, the spectral classification model is a neural network model; wherein, the spectral classification model is obtained by inputting the spectral feature matrices corresponding to several homomorphic targets of different materials into the neural network model for spectral classification training.

[0011] More preferably, the band range of the one-dimensional discrete spectral signal is from near-infrared to long-wave infrared.

[0012] More preferably, S2 specifically includes the following steps:

[0013] S2.1: Let f(λ) = s'(λ), find all local maxima and local minima in the signal f(λ); where s'(λ) is the normalized one-dimensional discrete spectral signal s'(λ), λ∈{λ1,λ2,λ3...λ n};

[0014] S2.2: Construct the upper envelope f(λ) of the signal using cubic spline interpolation by utilizing the local maxima and minima of f(λ). n (λ) and lower envelope f u (λ), and calculate the mean m(λ) of the upper and lower envelopes of f(λ) = (f u (λ)+f n (λ)) / 2;

[0015] S2.3: Calculate the first intermediate signal h1(λ) based on the mean of the upper and lower envelopes of f(λ), count the number of extreme points and zero-crossing points of the first intermediate signal, and determine whether the difference between the number of extreme points and the number of zero-crossing points satisfies the condition of being less than or equal to 1; at the same time, construct the upper envelope formed by the local maxima of the first intermediate signal and the lower envelope formed by the local minima of the first intermediate signal through cubic spline interpolation, and determine whether the average value of the upper and lower envelopes of the first intermediate signal satisfies the condition of being less than a preset threshold; proceed to S2.4;

[0016] S2.4: If the number of extreme points and zero-crossing points of the first intermediate signal is less than or equal to 1, and the average value of the upper and lower envelopes of the first intermediate signal is less than a preset threshold, then the first intermediate signal is taken as the first intrinsic mode signal, and proceed to S2.5; otherwise, let f(λ) = h1(λ), ​​find all local maxima and local minima in the signal f(λ), and proceed to S2.2 for signal sieving until the k-th iteration is repeated to obtain h k (λ)=h k (λ)-m k (λ), and satisfy h k (λ) The number of extreme points and the number of zero-crossing points are less than or equal to 1, and h k The first intrinsic mode signal IMF1(λ) = h is obtained when the average value of the upper and lower envelopes of (λ) is less than a preset threshold. k (λ); k is greater than or equal to 2;

[0017] S2.5: Obtain the residual signal r1(λ) = f(λ) - IMF1(λ) by subtracting the signal f(λ) from the intrinsic mode signal IMF1(λ). Let f(λ) = r1(λ), ​​find all local maxima and local minima in the signal f(λ), and return to S2.2 for signal sieving to obtain the second IMF2(λ); j (λ)=f(λ)-IMF j (λ), and so on, until the amplitude of the last residual signal is less than the preset value β or becomes a monotonic signal, at which point the decomposition process terminates, resulting in a total of l intrinsic mode signals (IMF). l (λ); where j is greater than or equal to 2.

[0018] More preferably, S3 specifically includes the following steps:

[0019] S3.1: Based on a certain λ in the wavelength λ a In l intrinsic mode signals IMF l (λ a The value in ) is x al The intermediate interpolation signal g is obtained by using cubic spline interpolation. a (l);

[0020] S3.2: Based on the interpolated intermediate signal g a (l), obtain the interpolated IMFc j (λ a ) = g a (l), so that λ a This results in i total eigenmodes; where 1≤j≤i;

[0021] S3.3: Repeat S3.1 to S3.2 until all values ​​of wavelength λ are traversed. Interpolate to obtain i total eigenmodes for each wavelength and obtain an n×i spectral signal interpolation matrix.

[0022] More preferably, S4 specifically includes the following steps:

[0023] S4.1: The normalized one-dimensional discrete spectral signal is divided into discrete spectral signals in the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window; and the n×i spectral signal interpolation matrix is ​​represented by the spectral signal interpolation matrix of the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window.

[0024] S4.2: Calculate the intermediate weight matrix of the short-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; calculate the intermediate weight matrix of the mid-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; calculate the intermediate weight matrix of the long-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; calculate the intermediate weight matrix of the spectral band outside the atmospheric window using discrete spectral signals and spectral signal interpolation matrices.

[0025] S4.3: The intermediate weight matrices of the shortwave, midwave, longwave, and atmospheric window outside spectral bands are used to expand the matrix to obtain the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outside spectral bands respectively; wherein the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outside spectral bands have the same dimension as the interpolation matrices of the spectral signals of the shortwave, midwave, longwave, and atmospheric window outside spectral bands respectively.

[0026] S4.4: Combine the final weight matrices of the shortwave, midwave, longwave, and atmospheric window bands to obtain the full-spectrum weight matrix;

[0027] S4.5: Multiply the spectral signal interpolation matrix element by element with the full spectral weight matrix to obtain the spectral feature matrix.

[0028] More preferably, the neural network model sequentially includes: a 1*1 convolutional layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a ReLU activation layer, a fully connected layer, and a ReLU activation layer.

[0029] On the other hand, the present invention provides an infrared broadband classification system for targets of the same shape but different materials, including:

[0030] The normalization module is used to normalize one-dimensional discrete spectral signals to the range of 0-1;

[0031] The EMD decomposition module is used to perform EMD decomposition on the normalized one-dimensional discrete spectral signal, decomposing the normalized one-dimensional discrete spectral signal into l intrinsic mode signals (IMFs). l (λ);

[0032] The cubic spline interpolation module is used to interpolate the intrinsic mode signals of l spectra using cubic splines, so that each wavelength corresponds to i total intrinsic mode signals, and obtain a spectral signal interpolation matrix of size n×i; where n is the number of one-dimensional discrete spectral signals;

[0033] The spectral feature matrix acquisition module is used to calculate the full-spectrum weight matrix based on the spectral signal interpolation matrix and the normalized one-dimensional discrete spectral signal. The spectral signal interpolation matrix of size n×i is compared with the full-spectrum weight matrix element by element to obtain the spectral feature matrix.

[0034] The spectral classification module is used to input the spectral feature matrix into the spectral classification model to obtain the spectral category; the spectral classification model is obtained by inputting the spectral feature matrices corresponding to several objects of the same shape but different materials into the neural network model for spectral classification training.

[0035] More preferably, the band range of the one-dimensional discrete spectral signal is from near-infrared to long-wave infrared.

[0036] More preferably, the EMD decomposition module includes:

[0037] The f(λ) signal analysis unit is used to set f(λ) = s'(λ) or f(λ) = h. k (λ) or f(λ)=h1(λ) or f(λ)=r1(λ) or f(λ)=r j Find all local maxima and local minima in the signal f(λ); where s'(λ) is the normalized one-dimensional discrete spectral signal s'(λ), λ∈{λ1,λ2,λ3...λ} n};h k (λ) is the k-th intermediate signal; h1(λ) is the first intermediate signal; r j (λ) is the j-th residual signal, where j is greater than or equal to 2;

[0038] The f(λ) upper and lower envelope mean calculation unit is used to construct the upper envelope f(λ) of the signal by using cubic spline interpolation based on the local maxima and local minima of f(λ). n (λ) and lower envelope f u (λ), and calculate the mean m(λ) of the upper and lower envelopes of f(λ) = (fu (λ)+f n (λ)) / 2;

[0039] The condition determination unit is used to calculate the first intermediate signal h1(λ) based on the mean of the upper and lower envelopes of f(λ), count the number of extreme points and zero crossings of the first intermediate signal, and determine whether the difference between the number of extreme points and the number of zero crossings satisfies the condition of being less than or equal to 1; at the same time, it constructs the upper envelope formed by the local maxima of the first intermediate signal and the lower envelope formed by the local minima of the first intermediate signal through cubic spline interpolation, and determines whether the average value of the upper and lower envelopes of the first intermediate signal satisfies the condition of being less than a preset threshold.

[0040] The first intrinsic mode signal acquisition unit is configured to, if the number of extreme points and the number of zero-crossing points of the first intermediate signal are less than or equal to 1, and the average value of the upper and lower envelopes of the first intermediate signal is less than a preset threshold, then use the first intermediate signal as the first intrinsic mode signal; otherwise, it drives the f(λ) signal analysis unit to set f(λ) = h1(λ) or f(λ) = h k (λ), find all local maxima and local minima in the signal f(λ), drive the upper and lower envelope mean calculation unit and the condition judgment unit of f(λ) to perform signal screening and condition judgment, and obtain the first intrinsic mode signal; where k is greater than or equal to 2;

[0041] Other intrinsic mode signal acquisition units are used to obtain the residual signal r1(λ) = f(λ) - IMF1(λ) by subtracting the signal f(λ) from the intrinsic mode signal IMF1(λ), ​​driving the f(λ) signal analysis unit to set f(λ) = r1(λ), ​​finding all local maxima and local minima in the signal f(λ), and driving the f(λ) signal analysis unit to obtain the second IMF2(λ); r j (λ)=f(λ)-IMF j (λ), and so on, until the amplitude of the last residual signal is less than the preset value β or becomes a monotonic signal, at which point the decomposition process terminates, resulting in a total of l intrinsic mode signals (IMF). l (λ); where j is greater than or equal to 2.

[0042] More preferably, the specific execution process of the cubic spline interpolation processing module is as follows:

[0043] Based on a certain λ in the wavelength λ a In l intrinsic mode signals IMF l (λ a The value in ) is x al The intermediate interpolation signal g is obtained by using cubic spline interpolation. a (l), obtain the interpolated IMFc j(λ a ) = g a (l) Iterate through all the values ​​in wavelength λ, and obtain i total eigenmodes for each wavelength by interpolation, and obtain an i×n spectral signal interpolation matrix.

[0044] More preferably, the spectral feature matrix acquisition module includes:

[0045] The matrix splitting unit is used to divide the normalized one-dimensional discrete spectral signal into discrete spectral signals in the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window; and to represent the n×i spectral signal interpolation matrix using the spectral signal interpolation matrix of the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window.

[0046] The intermediate weight matrix calculation unit is used to calculate the intermediate weight matrix of the short-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; to calculate the intermediate weight matrix of the mid-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; to calculate the intermediate weight matrix of the long-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; and to calculate the intermediate weight matrix of the spectral band outside the atmospheric window using discrete spectral signals and spectral signal interpolation matrices.

[0047] The final weight matrix calculation unit is used to expand the matrix using the intermediate weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands respectively, to obtain the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands; wherein the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands have the same dimensions as the spectral signal interpolation matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands respectively;

[0048] The matrix combination unit is used to combine the final weight matrices of the short-wavelength, mid-wavelength, long-wavelength, and atmospheric window-outside spectral bands to obtain the full-spectrum weight matrix.

[0049] The matrix multiplication unit is used to multiply the spectral signal interpolation matrix and the full spectral weight matrix element by element to obtain the spectral feature matrix.

[0050] More preferably, the neural network model sequentially includes: a 1*1 convolutional layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a ReLU activation layer, a fully connected layer, and a ReLU activation layer.

[0051] In summary, compared with the prior art, the above-described technical solutions conceived by this invention have the following advantages:

[0052] Beneficial effects:

[0053] This invention provides an infrared broadband classification method for targets of the same shape but different materials. This method involves interpolation and local weighted suppression of the intrinsic mode signals derived from spectral decomposition (more specifically, based on a certain λ within the wavelength λ). a In l intrinsic mode signals IMF l (λ a The value in ) is x al The intermediate interpolation signal g is obtained by using cubic spline interpolation. a (l); based on the interpolated intermediate signal g a (l), obtain the interpolated IMFc j (λ a ) = g a (l), so that λ a The corresponding i total intrinsic modes are obtained; the full spectrum weight matrix is ​​calculated based on the spectral signal interpolation matrix and the normalized one-dimensional discrete spectral signal. The spectral signal interpolation matrix of size n×i is multiplied element by element with the full spectrum weight matrix to obtain the spectral feature matrix. This can effectively extract the real information in the spectrum while suppressing the noise in the spectrum. This allows the neural network to use only a small number of samples, i.e., 10% of the total samples, for training and learning, so that the classification results can be highly accurate and the model can be applied, i.e., it has good generalization ability. Attached Figure Description

[0054] Figure 1 This is a flowchart of the infrared broadband classification method for targets of the same shape but different materials provided in the embodiments of the present invention;

[0055] Figure 2 This is a schematic diagram of fitting a certain wavelength provided in an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the interpolation results provided in an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of a one-dimensional spectral matrix extension provided in an embodiment of the present invention;

[0058] Figure 5 This is a network structure diagram provided in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0060] Metallic targets are composed of different materials and temperatures (from the inside out), and their thermal infrared images will show texture. Polymer targets, composed of internal air and thin-film polymer materials, are homogeneous and at the same temperature (except for local heat sources), and their thermal infrared images lack texture information. Metallic targets are considered true targets; polymer targets are considered false targets. The positions of adjacent extreme spectral points corresponding to reflection and absorption differ between true and false targets. Metallic targets typically have distributed heat sources internally, relatively large specific heat capacity, and large mass. Internal thermal energy is transferred to the outer shell through the internal structure and radiated outwards. Therefore, the total spectrum radiated by a true target includes not only the spectral energy reflected and absorbed externally but also its own radiation spectrum. Polymer targets, on the other hand, have no distributed heat sources internally, and their radiated spectrum only includes the spectral energy reflected and absorbed. Therefore, metallic and polymeric targets differ due to their different compositions. False targets are filled with air, have different specific heat capacities, and exchange matter and energy with their surrounding environment at different rates. Metallic objects have high mass (density) and complex structure, especially their internal metal structure, which is complex and has a large heat capacity. They also have complex and varied thermal and temperature fields inside, and their energy exchange with the outside world is complex and unique.

[0061] Polymer targets have low mass and density, and simple structure, consisting of a thin chemical film on the surface and internal air. The internal air, in particular, has a low heat capacity, a uniform internal thermal field and corresponding temperature field, and simple energy exchange with the external environment. This energy exchange is reflected in the spectral energy incident on the target from the external environment, which consists of two components: reflection (scattering) and absorption. Since the incident spectral energy on both metal and polymer targets is the same, metal targets are likely to absorb spectral energy more strongly and reflect it less strongly than polymer targets. Therefore, the spectral curves of metal targets are significantly different from those of polymer targets.

[0062] Based on the differences in the spectral curves of the aforementioned metallic and polymeric material targets, such as Figure 1 As shown, the spectral data acquisition device in this invention patent uses a self-developed spectral correlation device, capable of acquiring wavelengths from near-infrared to long-wave infrared (1.7µm to 14µm). This invention embodiment provides a method for infrared broadband classification of targets of the same shape but different materials, including the following steps:

[0063] S1: Normalize the spectral signal to the range of 0-1; if the one-dimensional discrete spectral signal is s(λ), λ∈{λ1,λ2,λ3...λ} n The normalization steps are as follows:

[0064] S1.1: Find the maximum value s in the spectral signal. max and minimum value s min ;

[0065] S1.2: Normalization yields

[0066] S2: Perform EMD decomposition on the normalized spectral signal s'(λ). Through EMD decomposition, the normalized one-dimensional discrete spectral signal s'(λ), λ∈{λ1,λ2,λ3...λ} is obtained. n The intrinsic mode signals (IMFs) of l spectra were decomposed. l (λ); Let f(λ) = s'(λ), then the decomposition steps are as follows:

[0067] S2.1: Find all local maxima and local minima of signal f(λ);

[0068] S2.2: Construct the upper envelope f of the signal using cubic spline interpolation at the local maxima and local minima of f(λ). n (λ) and lower envelope f u (λ), and calculate the mean m(λ) of the upper and lower envelopes = (f u (λ)+f n (λ)) / 2;

[0069] S2.3: Calculate the first intermediate signal h(λ) = f(λ) - m(λ), count the number of extreme points and zero-crossing points of the first intermediate signal h(λ), and determine whether the difference between the number of extreme points and the number of zero-crossing points satisfies the condition of being less than or equal to 1; construct the upper envelope h formed by the local maxima of the first intermediate signal h(λ) using cubic spline interpolation. n (λ) and the lower envelope h formed by the local minimum point of the first intermediate signal. u (λ), the condition for determining that the average value of the upper and lower envelopes is less than a preset threshold α, that is, determining that for any λ, α>(h) u (λ)+h n (λ)) / 2;

[0070] S2.4: If the first intermediate signal h(λ) satisfies the two conditions in S2.3, then the first intrinsic mode signal IMF1(λ) = h(λ) is obtained, and proceed to S2.5; otherwise, let f(λ) = h(λ), and repeat S2.1 to S2.3 for signal sieving until the k-th repetition yields h. k (λ)=h k (λ)-m k If (λ) satisfies the two conditions in S2.3, then the first intrinsic mode signal IMF1(λ) = h is obtained. k (λ);

[0071] S2.5: Obtain the residual signal r1(λ) = f(λ) - IMF1(λ) by subtracting the signal f(λ) from the intrinsic mode signal IMF1(λ), ​​and let f(λ) = r l (λ), repeat S2.1~S2.4 to obtain the second IMF2(λ); repeat this process until the amplitude of the last residual signal is less than the preset value β or becomes a monotonic signal, then terminate the decomposition process, and obtain a total of l intrinsic mode signals IMF. l (λ);

[0072] S3: Due to the differences between each spectral signal, the number of intrinsic mode signals decomposed from different spectral signals is variable. Therefore, it is necessary to obtain a fixed number of intrinsic mode signals from the decomposed intrinsic mode signals through cubic spline interpolation; more specifically:

[0073] S2 decomposes the spectral signal s'(λ) into l intrinsic mode signals (IMFs). l (λ), in the intrinsic mode signal IMF l i interpolated intrinsic mode signals IMFc are obtained from (λ) through cubic spline interpolation. i (λ), specifically including the following steps:

[0074] S3.1: For a certain λ in the wavelength λ a In l intrinsic mode signals IMF l (λ a The value in ) is x al ,like Figure 2 As shown, the value x is fitted using cubic spline interpolation. al Obtain the interpolated intermediate signal g a (l), g a (l) that is Figure 2 The dashed line segment in the middle;

[0075] S3.2: Based on the interpolated intermediate signal g a (l), obtain the interpolated IMFc j (λ a ) = g a (l);

[0076] S3.3: Repeat S3.1 to S3.2 until all values ​​of wavelength λ have been traversed, and the IMFc can be obtained. i (λ), such as Figure 3 The IMFc obtained by interpolation is shown. i (λ);

[0077] S4: Through the above two steps, the spectral signal is decomposed into multiple feature signals, that is, the spectral signal matrix of size n×1 is expanded to a matrix S of size n×i, such as... Figure 4As shown; then the spectral signal interpolation matrix S is multiplied element-wise by the full-spectrum weight matrix A to finally obtain the spectral feature matrix T = S * A; in practical applications, the spectral signal matrix of size n×1 is fixedly expanded to a matrix S of size n×10. The calculation of its weight matrix A specifically includes the following steps:

[0078] S4.1: Since the spectral data acquisition range is from 1.7µm to 14µm, its spectral range includes the shortwavelength λ band. s , medium wave spectrum band λ m Longwavelength spectral band λ l and the spectral band outside the atmospheric window λ o That is, for the normalized one-dimensional discrete spectral signal s(λ), λ∈{λ s ,λ m ,λ l ,λ o Therefore, the spectral signal interpolation matrix S also includes the spectral signal interpolation matrix S for the shortwave band. s The spectral signal interpolation matrix S in the mid-wave band m The spectral signal interpolation matrix S in the long-wavelength band l and the spectral signal interpolation matrix S of the spectral band outside the atmospheric window o That is, S∈{S s ,S m ,S l ,S o ,};Since the discrete spectral signal s(λ)∈{s(λ s ),s(λ m ),s(λ l ),s(λ o If s(λ) is one-dimensional, then s(λ) can be... s ), s(λ m ), s(λ l ) and s(λ o ) can be viewed as column vectors, and each is represented by a one-dimensional matrix G. s G m G l and G o express;

[0079] S4.2: Calculate the intermediate weight matrix M for the shortwave band. s :

[0080]

[0081] Where “.*” means multiplying the elements at the same position in the matrix one by one, “. / ” means dividing the elements at the same position in the matrix one by one, and “sqrt()” means taking the square root of each element in the matrix one by one;

[0082] S4.3: Let matrix A's =0.5*M s +0.5, and expand matrix A' s The final weight matrix of the obtained shortwave band Make the final weight matrix A of the shortwave band s Dimension and matrix S s Consistent;

[0083] S4.4: Similarly, the final weighting matrix A for the mid-wave spectrum can be calculated. m The final weighting matrix A for the long-wavelength spectral band l And the final weight matrix A of the spectral band outside the atmospheric window. o Combining these four weight matrices yields the full-spectrum weight matrix A∈{A s A m A l A o ,};

[0084] S5: The spectral feature matrix sequence obtained through the above steps is randomly divided into a training set and a validation set, and input into a system such as... Figure 5 The neural network shown is trained and classified, and its output is the type of target material. If the spectral dataset consists of targets of the same type as metal and targets of the same type as chemical materials, then the neural network output is the same as the metal material and the chemical material. The structure of the neural network model is as follows: a 1*1 convolutional layer with 10 input channels and 1 output channel; a fully connected layer with an input size of n and an output size of 256; a ReLU activation layer; a fully connected layer with an input size of 256 and an output size of 64; a ReLU activation layer; a fully connected layer with an input size of 64 and an output size of 16; a ReLU activation layer; a fully connected layer with an input size of 16 and an output size equal to the number of target material types collected; and a ReLU activation layer.

[0085] Thus, by completing all the above steps, the high-resolution infrared broadband classification of targets of the same shape but different materials has been achieved.

[0086] On the other hand, the present invention provides an infrared broadband classification system for targets of the same shape but different materials, including:

[0087] The normalization module is used to normalize one-dimensional discrete spectral signals to the range of 0-1;

[0088] The EMD decomposition module is used to perform EMD decomposition on the normalized one-dimensional discrete spectral signal, decomposing the normalized one-dimensional discrete spectral signal into l intrinsic mode signals (IMFs). l (λ);

[0089] The cubic spline interpolation module is used to interpolate the intrinsic mode signals of l spectra using cubic splines, so that each wavelength corresponds to i total intrinsic mode signals, and obtain a spectral signal interpolation matrix of size n×i; where n is the number of one-dimensional discrete spectral signals;

[0090] The spectral feature matrix acquisition module is used to calculate the full-spectrum weight matrix based on the spectral signal interpolation matrix and the normalized one-dimensional discrete spectral signal. The spectral signal interpolation matrix of size n×i is compared with the full-spectrum weight matrix element by element to obtain the spectral feature matrix.

[0091] The spectral classification module is used to input the spectral feature matrix into the spectral classification model to obtain the spectral category; the spectral classification model is obtained by inputting the spectral feature matrices corresponding to several objects of the same shape but different materials into the neural network model for spectral classification training.

[0092] More preferably, the band range of the one-dimensional discrete spectral signal is from near-infrared to long-wave infrared.

[0093] More preferably, the EMD decomposition module includes:

[0094] The f(λ) signal analysis unit is used to set f(λ) = s'(λ) or f(λ) = h. k (λ) or f(λ)=h1(λ) or f(λ)=r1(λ) or f(λ)=r j Find all local maxima and local minima in the signal f(λ); where s'(λ) is the normalized one-dimensional discrete spectral signal s'(λ), λ∈{λ1,λ2,λ3...λ} n};h k (λ) is the k-th intermediate signal; h1(λ) is the first intermediate signal; r j (λ) is the j-th residual signal, where j is greater than or equal to 2;

[0095] The f(λ) upper and lower envelope mean calculation unit is used to construct the upper envelope f(λ) of the signal by using cubic spline interpolation based on the local maxima and local minima of f(λ). n (λ) and lower envelope f u (λ), and calculate the mean m(λ) of the upper and lower envelopes of f(λ) = (f u (λ)+f n (λ)) / 2;

[0096] The condition determination unit is used to calculate the first intermediate signal h1(λ) based on the mean of the upper and lower envelopes of f(λ), count the number of extreme points and zero crossings of the first intermediate signal, and determine whether the difference between the number of extreme points and the number of zero crossings satisfies the condition of being less than or equal to 1; at the same time, it constructs the upper envelope formed by the local maxima of the first intermediate signal and the lower envelope formed by the local minima of the first intermediate signal through cubic spline interpolation, and determines whether the average value of the upper and lower envelopes of the first intermediate signal satisfies the condition of being less than a preset threshold.

[0097] The first intrinsic mode signal acquisition unit is configured to, if the number of extreme points and the number of zero-crossing points of the first intermediate signal are less than or equal to 1, and the average value of the upper and lower envelopes of the first intermediate signal is less than a preset threshold, then use the first intermediate signal as the first intrinsic mode signal; otherwise, it drives the f(λ) signal analysis unit to set f(λ) = h1(λ) or f(λ) = h k (λ), find all local maxima and local minima in the signal f(λ), drive the upper and lower envelope mean calculation unit and the condition judgment unit of f(λ) to perform signal screening and condition judgment, and obtain the first intrinsic mode signal; where k is greater than or equal to 2;

[0098] Other intrinsic mode signal acquisition units are used to obtain the residual signal r1(λ) = f(λ) - IMF1(λ) by subtracting the signal f(λ) from the intrinsic mode signal IMF1(λ), ​​driving the f(λ) signal analysis unit to set f(λ) = r1(λ), ​​finding all local maxima and local minima in the signal f(λ), and driving the f(λ) signal analysis unit to obtain the second IMF2(λ); r j (λ)=f(λ)-IMF j (λ), and so on, until the amplitude of the last residual signal is less than the preset value β or becomes a monotonic signal, at which point the decomposition process terminates, resulting in a total of l intrinsic mode signals (IMF). l (λ); where j is greater than or equal to 2.

[0099] More preferably, the specific execution process of the cubic spline interpolation processing module is as follows:

[0100] Based on a certain λ in the wavelength λ a In l intrinsic mode signals IMF l (λ a The value in ) is x al The intermediate interpolation signal g is obtained by using cubic spline interpolation. a (l), obtain the interpolated IMFc j (λ a ) = g a(l) Iterate through all the values ​​in wavelength λ, and obtain i total eigenmodes for each wavelength by interpolation, and obtain an i×n spectral signal interpolation matrix.

[0101] More preferably, the spectral feature matrix acquisition module includes:

[0102] The matrix splitting unit is used to divide the normalized one-dimensional discrete spectral signal into discrete spectral signals in the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window; and to represent the n×i spectral signal interpolation matrix using the spectral signal interpolation matrix of the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window.

[0103] The intermediate weight matrix calculation unit is used to calculate the intermediate weight matrix of the short-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; to calculate the intermediate weight matrix of the mid-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; to calculate the intermediate weight matrix of the long-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; and to calculate the intermediate weight matrix of the spectral band outside the atmospheric window using discrete spectral signals and spectral signal interpolation matrices.

[0104] The final weight matrix calculation unit is used to expand the matrix using the intermediate weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands respectively, to obtain the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands; wherein the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands have the same dimensions as the spectral signal interpolation matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands respectively;

[0105] The matrix combination unit is used to combine the final weight matrices of the short-wavelength, mid-wavelength, long-wavelength, and atmospheric window-outside spectral bands to obtain the full-spectrum weight matrix.

[0106] The matrix multiplication unit is used to multiply the spectral signal interpolation matrix and the full spectral weight matrix element by element to obtain the spectral feature matrix.

[0107] More preferably, the neural network model sequentially comprises: a 1*1 convolutional layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a ReLU activation layer, a fully connected layer, a ReLU activation layer, a fully connected layer, and a ReLU activation layer. In summary, compared with the prior art, the present invention has the following advantages:

[0108] This invention provides an infrared broadband classification method for targets of the same shape but different materials. This method involves interpolation and local weighted suppression of the intrinsic mode signals derived from spectral decomposition (more specifically, based on a certain λ within the wavelength λ). a In l intrinsic mode signals IMF l (λ a The value in ) is x al The intermediate interpolation signal g is obtained by using cubic spline interpolation. a (l); based on the interpolated intermediate signal g a (l), obtain the interpolated IMFc j (λ a ) = g a (l), so that λ a The corresponding i total intrinsic modes are obtained; the weight matrix is ​​set according to the strength of the noise signal contained in each wavelength spectrum, and the spectral signal matrix of size n×i is multiplied by the weight matrix to obtain the spectral feature matrix. This can effectively extract the real information in the spectrum while suppressing the noise in the spectrum, so that only a small number of samples, i.e. 10% of the total samples, are used for training in the neural network to ensure that the classification results have high accuracy, while ensuring the applicability of the model, i.e., good generalization ability.

[0109] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for infrared broadband classification of targets of the same shape but different materials, characterized in that, Includes the following steps: S1: Normalize the one-dimensional discrete spectral signal to the range of 0-1; S2: Perform EMD decomposition on the normalized one-dimensional discrete spectral signal to decompose the normalized one-dimensional discrete spectral signal into... l The intrinsic mode signals of each spectrum ; S3: Will l The intrinsic mode signals of each spectrum are obtained by cubic spline interpolation, so that each wavelength corresponds to... i Identify the total intrinsic mode signals and obtain them. A spectral signal interpolation matrix of size; where, n The number of one-dimensional discrete spectral signals; S4: Calculate the full-spectrum weight matrix based on the spectral signal interpolation matrix and the normalized one-dimensional discrete spectral signal. The spectral feature matrix is ​​obtained by element-wise multiplying the spectral signal interpolation matrix of the magnitude of the signal with the full spectral weight matrix. S5: Input the spectral feature matrix into the spectral classification model to obtain the spectral category; wherein, the spectral classification model is obtained by inputting the spectral feature matrices corresponding to several homomorphic targets of different materials into the neural network model for spectral classification training.

2. The infrared broadband classification method according to claim 1, characterized in that, The band range of one-dimensional discrete spectral signals extends from near-infrared to long-wave infrared.

3. The infrared broadband classification method according to claim 1 or 2, characterized in that, S2 specifically includes the following steps: S2.1: Order Find the signal All local maxima and local minima in the ; among them, The normalized one-dimensional discrete spectral signal ; S2.2: Utilizing The local maxima and local minima are constructed using cubic spline interpolation. upper envelope of the signal and lower envelope and calculate Mean of upper and lower envelopes ; S2.3: According to and Calculate the first intermediate signal by the mean of the upper and lower envelopes. Count the number of extreme points and zero-crossing points of the first intermediate signal, and determine whether the difference between the number of extreme points and the number of zero-crossing points satisfies the condition of being less than or equal to 1; at the same time, construct the upper envelope formed by the local maxima of the first intermediate signal and the lower envelope formed by the local minima of the first intermediate signal through cubic spline interpolation, and determine whether the average value of the upper and lower envelopes of the first intermediate signal satisfies the condition of being less than a preset threshold; proceed to S2.4; S2.4: If the number of extreme points and the number of zero-crossing points of the first intermediate signal are less than or equal to 1, and the average value of the upper and lower envelopes of the first intermediate signal is less than a preset threshold, then the first intermediate signal is taken as the first intrinsic mode signal, and proceed to S2.5; otherwise, let Find the signal All local maxima and local minima are identified, and the process proceeds to S2.2 for signal sieving, repeating this process until the next step is completed. k This time received And satisfy The number of extreme points and the number of zero-crossing points are less than or equal to 1, and The first intrinsic mode signal is obtained when the average value of the upper and lower envelopes is less than a preset threshold. ; k Greater than or equal to 2; S2.5: via signal With intrinsic mode signals The residual signal is obtained by subtraction. ,make Find the signal All local maxima and local minima are identified, and the signal is then processed in step S2.2 to obtain the second... ; This process continues until the amplitude of the last residual signal is less than the preset value β or becomes a monotonic signal, at which point the decomposition process terminates, resulting in a total of [number missing]. Individual eigenmode signals ;in, j Greater than or equal to 2.

4. The infrared broadband classification method according to claim 3, characterized in that, S3 specifically includes the following steps: S3.1: Based on wavelength One of them exist Individual eigenmode signals The value in The intermediate interpolation signal is obtained by using cubic spline interpolation. ; S3.2: Based on the interpolated intermediate signal Obtain interpolation ,make Corresponding to i There are 1 total intrinsic modes; among them, 1 j i ; S3.3: Repeat S3.1 to S3.2 until the wavelengths have been traversed. All values ​​in the range are interpolated to obtain the corresponding wavelength. i Each total intrinsic mode is obtained. A spectral signal interpolation matrix of a certain size.

5. The infrared broadband classification method according to claim 1, characterized in that, S4 specifically includes the following steps: S4.1: Divide the normalized one-dimensional discrete spectral signal into discrete spectral signals in the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer atmospheric window band; and... The magnitude of the spectral signal interpolation matrix is ​​represented by the spectral signal interpolation matrix of the short-wavelength band, the spectral signal interpolation matrix of the mid-wavelength band, the spectral signal interpolation matrix of the long-wavelength band, and the spectral signal interpolation matrix of the band outside the atmospheric window; S4.2: Calculate the intermediate weight matrix of the short-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; calculate the intermediate weight matrix of the mid-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; calculate the intermediate weight matrix of the long-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; calculate the intermediate weight matrix of the spectral band outside the atmospheric window using discrete spectral signals and spectral signal interpolation matrices. S4.3: The intermediate weight matrices of the shortwave, midwave, longwave, and atmospheric window outside spectral bands are used to expand the matrix to obtain the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outside spectral bands respectively; wherein the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outside spectral bands have the same dimension as the interpolation matrices of the spectral signals of the shortwave, midwave, longwave, and atmospheric window outside spectral bands respectively. S4.4: Combine the final weight matrices of the shortwave, midwave, longwave, and atmospheric window bands to obtain the full-spectrum weight matrix; S4.5: Multiply the spectral signal interpolation matrix element by element with the full spectral weight matrix to obtain the spectral feature matrix.

6. An infrared broadband classification system for targets of the same shape but different materials, characterized in that, include: The normalization module is used to normalize one-dimensional discrete spectral signals to the range of 0-1; The EMD decomposition module is used to perform EMD decomposition on the normalized one-dimensional discrete spectral signal, decomposing the normalized one-dimensional discrete spectral signal into EMD components. l The intrinsic mode signals of each spectrum ; The cubic spline interpolation processing module is used to process... l The intrinsic mode signals of each spectrum are obtained by cubic spline interpolation, so that each wavelength corresponds to... i Identify the total intrinsic mode signals and obtain them. A spectral signal interpolation matrix of size; where, n The number of one-dimensional discrete spectral signals; The spectral feature matrix acquisition module is used to calculate the full-spectrum weight matrix based on the spectral signal interpolation matrix and the normalized one-dimensional discrete spectral signal. The spectral feature matrix is ​​obtained by element-wise multiplying the spectral signal interpolation matrix of the magnitude of the signal with the full spectral weight matrix. The spectral classification module is used to input the spectral feature matrix into the spectral classification model to obtain the spectral category; the spectral classification model is obtained by inputting the spectral feature matrices corresponding to several objects of the same shape but different materials into the neural network model for spectral classification training.

7. The infrared broadband classification system according to claim 6, characterized in that, The band range of one-dimensional discrete spectral signals extends from near-infrared to long-wave infrared.

8. The infrared broadband classification system according to claim 6 or 7, characterized in that, The EMD decomposition module includes: Signal analysis unit, used to enable or or or or Find the signal All local maxima and local minima in the ; among them, The normalized one-dimensional discrete spectral signal ; This is the kth intermediate signal; This is the first intermediate signal; For the first j sub-residual signal, j Greater than or equal to 2; The upper and lower envelope mean calculation unit is used to utilize... The local maxima and local minima are constructed using cubic spline interpolation. upper envelope of the signal and lower envelope and calculate Mean of upper and lower envelopes ; Condition determination unit, used to determine based on and Calculate the first intermediate signal by the mean of the upper and lower envelopes. The number of extreme points and zero-crossing points of the first intermediate signal are counted, and it is determined whether the difference between the number of extreme points and the number of zero-crossing points satisfies the condition of being less than or equal to 1. At the same time, the upper envelope formed by the local maxima of the first intermediate signal and the lower envelope formed by the local minima of the first intermediate signal are constructed by cubic spline interpolation, and it is determined whether the average value of the upper and lower envelopes of the first intermediate signal satisfies the condition of being less than a preset threshold. The first intrinsic mode signal acquisition unit is configured to, if the number of extreme points and the number of zero-crossing points of the first intermediate signal are less than or equal to 1, and the average value of the upper and lower envelopes of the first intermediate signal is less than a preset threshold, then use the first intermediate signal as the first intrinsic mode signal; otherwise, drive... Signal analysis unit command or Find the signal All local maxima and local minima in the driving force After signal screening by the upper and lower envelope mean calculation unit and the condition determination unit, the first intrinsic mode signal is obtained; whereby... k Greater than or equal to 2; Other intrinsic mode signal acquisition units are used to acquire signals With intrinsic mode signals The residual signal is obtained by subtraction. ,drive Signal analysis unit command Find the signal All local maxima and local minima in the driving force The signal analysis unit obtains the second... ; This process continues until the amplitude of the last residual signal is less than the preset value β or becomes a monotonic signal, at which point the decomposition process terminates, resulting in a total of [number missing]. Individual eigenmode signals ;in, j Greater than or equal to 2.

9. The infrared broadband classification system according to claim 8, characterized in that, The specific execution process of the cubic spline interpolation module is as follows: Based on wavelength One of them exist Individual eigenmode signals The value in The intermediate interpolation signal is obtained by using cubic spline interpolation. Obtain interpolation ; Traversing wavelengths All values ​​in the range are interpolated to obtain the corresponding wavelength. i Each total intrinsic mode is obtained. A spectral signal interpolation matrix of a certain size.

10. The infrared broadband classification system according to claim 6, characterized in that, The spectral feature matrix acquisition module includes: The matrix splitting unit is used to divide the normalized one-dimensional discrete spectral signal into discrete spectral signals in the short-wavelength band, the mid-wavelength band, the long-wavelength band, and the outer-spectral band outside the atmospheric window; and... The magnitude of the spectral signal interpolation matrix is ​​represented by the spectral signal interpolation matrix of the short-wavelength band, the spectral signal interpolation matrix of the mid-wavelength band, the spectral signal interpolation matrix of the long-wavelength band, and the spectral signal interpolation matrix of the band outside the atmospheric window; The intermediate weight matrix calculation unit is used to calculate the intermediate weight matrix of the short-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; to calculate the intermediate weight matrix of the mid-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; to calculate the intermediate weight matrix of the long-wavelength spectral band using discrete spectral signals and spectral signal interpolation matrices; and to calculate the intermediate weight matrix of the spectral band outside the atmospheric window using discrete spectral signals and spectral signal interpolation matrices. The final weight matrix calculation unit is used to expand the matrix using the intermediate weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands respectively, to obtain the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands; wherein the final weight matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands have the same dimensions as the spectral signal interpolation matrices of the shortwave, midwave, longwave, and atmospheric window outer spectral bands respectively; The matrix combination unit is used to combine the final weight matrices of the short-wavelength, mid-wavelength, long-wavelength, and atmospheric window-outside spectral bands to obtain the full-spectrum weight matrix. The matrix multiplication unit is used to multiply the spectral signal interpolation matrix and the full spectral weight matrix element by element to obtain the spectral feature matrix.

Images