Spatial conversion projection and depth migration fused energy spectrum feature extraction method
Through the method of fusing space conversion projection and deep migration, radionuclide energy spectrum data is converted from Euclidean space to Banach and RGB color space, combined with multi-level depth visual descriptor migration, the problem of high computing resource consumption in traditional methods is solved, and the efficient extraction of radionuclide energy spectrum features is achieved.
Patent Information
- Application Number
- CN202510527285.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-25
- Publication Date
- 2025-08-08
AI Technical Summary
In the prior art, in radionuclide recognition, traditional methods rely on expert knowledge and consume large computing resources, and the neural network training time is long, making it difficult to effectively extract energy spectrum features in complex backgrounds.
Using the method of fusion of spatial conversion projection and deep migration, the radionuclide energy spectrum data is converted from Euclidean space to Banach space and RGB color space by pre-training the deep convolutional neural network model, and combined with multi-level depth visual descriptor migration, global and local features are extracted.
Effectively extracting the global and local characteristics of the radio performance spectrum, improving the utilization rate of energy spectrum information, reducing the computing resource requirements and training time, and is suitable for radionuclide recognition in complex contexts.
Smart Images
Figure CN120451577A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of nuclear radiation detection, and more specifically to the technical field of an energy spectrum feature extraction method that integrates space conversion projection and depth migration. Background Art
[0002] The experimental gamma spectrum obtained by measuring a monoenergetic gamma-ray-emitting radionuclide using a NaI(Tl) scintillator detector does not simply correspond to the pulse distribution resulting from the total absorption of gamma-ray energies. In addition to the full-energy peak, a continuous distribution exists, with most energies below the full-energy peak. The instrumental spectral data obtained during detection (the counts of each channel of a multichannel analyzer) may comprise contributions from dozens of interacting processes. Furthermore, interference from bremsstrahlung and characteristic X-ray peaks can also contribute to the gamma spectrum. In other words, even for a monoenergetic gamma-ray source, the gamma spectrum obtained using a NaI(Tl) detector is not a single, simple spectral line. The gamma spectrum of radionuclides emitting multiple energies, or of samples containing mixed nuclides, is even more complex. Furthermore, due to the limited energy resolution of the NaI(Tl) detector, the spectral data lack sharp selectivity, and a single data point often represents the combined contribution of multiple factors. Although full-energy peak data are an important basis for quantitative gamma-ray spectrum analysis, relying solely on full-energy peak data often presents numerous challenges.
[0003] To address this problem, traditional feature extraction methods for radionuclide identification algorithms can be summarized as searching for the full-energy peak position and peak height from the gamma-ray energy and matching them with peaks in the radionuclide library. This is generally based on physical rules and does not require training. These methods rely heavily on expert knowledge and only utilize a portion of the energy spectrum. Their limitations are magnified when characteristic peaks are overwhelmed by background noise or when several interfering peaks are extracted. To improve the completeness of information used, some researchers have used the raw counts of the full spectrum as the identification parameter and a neural network as the training model, achieving impressive recognition results.
[0004] However, since the raw spectral data is directly used as the identification feature, the data volume is large, the neural network structure is large, and the training time is long. The essence of feature extraction is effective dimensionality reduction. More researchers are beginning to experiment with energy spectrum feature extraction using transform analysis methods. Radioactive gamma ray spectrum feature extraction based on these methods has achieved good results. However, some machine learning methods with good classification results have high computational requirements. The time, effort, and computing power required to train a new model from scratch are factors that cannot be ignored in classification models. Summary of the Invention
[0005] The purpose of the present invention is: in order to solve the above technical problems, the present invention provides an energy spectrum feature extraction method that integrates space conversion projection and depth migration.
[0006] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:
[0007] The present invention provides a method for extracting energy spectrum features by integrating spatial transformation projection and depth migration, comprising the following steps:
[0008] S1. Based on the task requirements and data characteristics, taking into account factors such as model performance, complexity, and computing resources, pre-training with public datasets is used to learn common feature representations. The resulting deep convolutional neural network model is used as the source domain model, and the radionuclide energy spectrum data is used as the target domain.
[0009] S2. Through vector-matrix-image cascade representation transformation, the characteristic vector of the gamma energy spectrum from the Euclidean space is represented as a characteristic matrix in the Banach space, and further mapped into a characteristic image in the RGB color space;
[0010] S3. The radionuclide energy spectrum feature image in the RGB color space is used as the source domain model input, and the global and local features of the radionuclide energy spectrum are obtained through multi-level deep visual descriptor migration.
[0011] In one embodiment, in step S1, a small amount of labeled measured radionuclide gamma spectrum data obtained by measuring a radioactive source using a detector is used as the target domain. An open experimental area that meets safety standards is selected as a fixed experimental operation point. To accurately measure the relative distance between the spectrometer and various radioactive sources, a ruler is placed on the ground in the experimental area to ensure that the radiation source is accurately positioned at the origin of the ruler. Parameters such as measurement distance, measurement duration, and measurement angle are set to obtain measured spectrum samples.
[0012] Specifically, we selected public datasets, such as ImageNet, which is a large and rich visual database with more than 14 million annotated images covering more than 20,000 categories. The categories cover a wide range and have similarities to certain features in radionuclide energy spectrum feature images to a certain extent.
[0013] In one embodiment, the specific process of step S2 is as follows:
[0014] S21, through the vector-matrix-image cascade representation transformation, the characteristic vector of the gamma energy spectrum from the Euclidean space is represented as a characteristic matrix in the Banach space;
[0015] S22. Select a color map with high discrimination and good continuity to convert the feature matrix into a feature image in the RGB color space that is intuitive and easy to visually perceive.
[0016] In one embodiment, the specific process of step S21 is as follows:
[0017] Let the radionuclide γ spectrum s be a wide-sense stationary random vector in Euclidean space, expressed as s = {s k}, where s k is the number of k-th energy events distributed in the energy domain. In an ideal situation, s k It is positively correlated with the duration of the measurement, that is:
[0018] s k =α k ·t
[0019] Where t is the measurement duration, α k is the parameter of the kth count value affected by the environmental background noise, measurement angle and distance, and radiation source intensity.
[0020] The matrix of radionuclide energy spectrum in Banach space can be expressed as:
[0021]
[0022] Where f is the mapping of the radionuclide gamma spectrum s from the vector form of Euclidean space to the matrix form of Banach space; p ij Represents the element in the i-th row and j-th column of the matrix; i = 1…m, j = 1…n; specifically, Where k is the index of the energy spectrum vector value, and n is the number of columns in the matrix.
[0023] In one embodiment, the specific process of step S22 is as follows:
[0024] Each entry in the color map is an RGB triplet that defines a color, with each component intensity ranging from [0, 1]; Figure 3 The figure shows the mapping process from normalized values to actual RGB colors. It accepts a normalized value, scales it to the number of colormap entries, and returns the corresponding entry. This process enhances data visualization and helps introduce image processing and analysis techniques to mine more potential feature information.
[0025] The mapping of matrix data to RGB color space can be expressed as:
[0026]
[0027] Where Q is the image representation of the radionuclide γ spectrum in RGB color space, q ij is the element value of the image representation, m and n are the number of rows and columns of the image representation. g represents the function of color mapping using the color map. This function represents the mapping process from the normalized value to the actual RGB color, accepts a value normalized and scaled to the number of color map entries, and returns the corresponding entry, Nc is the number of entries in the colormap. p ij is the element in the i-th row and j-th column of the energy spectrum matrix.
[0028] In one embodiment, the specific process of step S3 is as follows:
[0029] S31. Fine-tune the source domain model based on the target domain data;
[0030] S32. Use the output of the convolutional layer and the fully connected layer of the source domain model as the image descriptor to perform activation vector migration. Source domain model construction and activation vector migration;
[0031] S33, obtaining the activation map corresponding to the feature map, which contains key information of the input data at different levels and dimensions;
[0032] S33. The fusion operation involves the combination, transformation or aggregation of activation maps, aiming to generate a new vector that can integrate the advantages of the original activation map, reduce redundancy, and generate a one-dimensional feature vector.
[0033] In one embodiment, in step S31, the process of fine-tuning the source domain model based on the target domain data is as follows: some layers are fixed and the parameter gradients are not updated; the learning rate is 5e -5 , optimizer SGD stochastic gradient descent, loss function cross entropy, epochs is 20.
[0034] In one embodiment, in step S32, the outputs of the convolutional layer and the fully connected layer of the source domain model are used as image descriptors to perform activation vector migration source domain model construction and activation vector migration; the specific process is as follows:
[0035] In public datasets The source domain model obtained through pre-training is:
[0036]
[0037] Where, For public datasets The sample space of Its corresponding category label space, f PM To use public datasets A pre-trained source domain model that enables reasoning from samples to class labels;
[0038] Activation vectors extracted from one or more layers of the source domain model:
[0039] A=φ(fPM,Q,l);
[0040] Where A represents the extracted activation vector, f PM To use public datasets The pre-trained source domain model, Q is the input image, l represents the index (or identity) of the selected single layer or multiple layers, and φ is a function that extracts the activation vector from one or more layers (convolutional layers or fully connected layers) of the source domain model.
[0041] In one embodiment, in step S33, an activation map corresponding to the feature map is obtained, which contains key information of the input data at different levels and dimensions. The specific process is as follows:
[0042] like A=C l ={c i |i∈{1,2,…,M}};
[0043]
[0044] Where, is a collection of convolutional layers, is a collection of fully connected layers.
[0045] If l represents one or more layers in the convolutional layer, then the set of activation vectors C passed from the convolutional layer is defined as l , A represents the extracted activation vector, M is the number of activation vectors passed from the convolutional layer, c i is the activation vector from the i-th convolutional layer.
[0046] C i is the activation vector of the convolutional layer, which consists of multiple feature maps, c (i) is the feature map generated by each convolution kernel in the i-th convolution layer after performing the convolution operation on the input data, d (i) is the number of features extracted in the i-th convolutional layer.
[0047] like
[0048] Where, if l represents one or more layers in the fully connected layer, then the set of activation vectors transferred from the fully connected layer is defined as A represents the extracted activation vector, N is the number of activation vectors passed from the fully connected layer, and f j is the activation vector from the jth convolutional layer.
[0049] In one embodiment, in step S34, the fusion operation involves combining, transforming, or aggregating the activation maps, aiming to generate a new vector that can integrate the advantages of the original activation maps, reduce redundancy, and generate a one-dimensional feature vector. The specific process is as follows:
[0050] Traverse all elements of the feature map and perform global average pooling:
[0051]
[0052] Where, Representation feature map The value of a specific element at position (h,w); h o and w o are the height and width of the feature map, respectively, and are calculated using the following formula.
[0053]
[0054] Where h i and w i Represents the height and width of the input image, h k and w k They represent the height and width of the convolution kernel respectively, pad and str represent the padding and stride respectively.
[0055] One-dimensional feature vector generation:
[0056] like
[0057] Where, is the set of convolutional layers; c i is the final convolutional layer feature; i represents c i is the activation vector from the i-th convolutional layer. c (i) is the feature map generated by each convolution kernel in the i-th convolution layer after performing the convolution operation on the input data, d (i) is the number of features extracted in the i-th convolutional layer. Represents the dth extracted from the i-th convolutional layer (i) The final vector of the features after global average pooling.
[0058] like
[0059] Where, is the set of fully connected layers; f j is the fully connected layer feature finally obtained; j represents f j is the activation vector from the jth fully connected layer. Each activation vector f j (j=1,2,…,N) is a one-dimensional vector of length K, where N is the number of activation vectors and K is the number of neurons in the jth fully connected layer.
[0060] The beneficial effects of the present invention are as follows:
[0061] The global and local features of the radioactive energy spectrum with strong semantic information were effectively extracted, which improved the utilization rate of the energy spectrum information and enhanced the credibility of the features. At the same time, it avoided the huge amount of data, strong hardware computing power and long training time usually required for deep learning, and opened up a new research perspective for the study of feature extraction methods of radionuclide energy spectra in complex background scenes. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.
[0063] Figure 1 It is a flow chart of an energy spectrum feature extraction method of the present invention that integrates space conversion projection and depth migration.
[0064] Figure 2 It is a flow chart of the vector-matrix-image cascade representation transformation process of the present invention.
[0065] Figure 3 Schematic diagram of color mapping of the present invention.
[0066] Figure 4 It is a flowchart of the multi-level depth visual descriptor migration process of the present invention. DETAILED DESCRIPTION
[0067] To make the technical problems, technical solutions, and technical effects of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.
[0068] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.
[0069] Example 1
[0070] like Figures 1 to 4 As shown, this embodiment provides an energy spectrum feature extraction method that combines spatial transformation projection with depth migration, including the following steps:
[0071] S1. Based on the task requirements and data characteristics, taking into account factors such as model performance, complexity, and computing resources, pre-training with public datasets is used to learn common feature representations. The resulting deep convolutional neural network model is used as the source domain model, and the radionuclide energy spectrum data is used as the target domain.
[0072] The target domain is a small amount of labeled radionuclide gamma energy spectrum data obtained by measuring radioactive sources using detectors. An open experimental area that meets safety standards is selected as a fixed operating point for the experiment. To accurately measure the relative distance between the spectrometer and various radioactive sources, a ruler is placed on the ground in the experimental area to ensure that the radiation source is accurately positioned at the ruler's origin. Parameters such as measurement distance, measurement duration, and measurement angle are set to obtain measured energy spectrum samples.
[0073] In addition, public datasets such as ImageNet are selected. They are large and rich visual databases with more than 14 million annotated images covering more than 20,000 categories. The categories cover a wide range and have similarities with certain features in radionuclide energy spectrum feature images to a certain extent.
[0074] S2. Through the vector-matrix-image cascade representation transformation, the characteristic vector of the gamma spectrum from the Euclidean space is represented as a characteristic matrix in the Banach space, and further mapped into a characteristic image in the RGB color space. The specific process is as follows:
[0075] S21. The characteristic vector of the gamma energy spectrum from the Euclidean space is represented as a characteristic matrix in the Banach space through a vector-matrix-image cascade representation transformation. The specific process is as follows:
[0076] Let the radionuclide γ spectrum s be a wide-sense stationary random vector in Euclidean space, expressed as s = {s k}, where s k is the number of k-th energy events distributed in the energy domain. In an ideal situation, s k It is positively correlated with the duration of the measurement, that is:
[0077] s k =α k ·t
[0078] Where t is the measurement duration, α k is the parameter of the kth count value affected by the environmental background noise, measurement angle and distance, and radiation source intensity.
[0079] The matrix of radionuclide energy spectrum in Banach space can be expressed as:
[0080]
[0081] Where f is the mapping of the radionuclide gamma spectrum s from the vector form of Euclidean space to the matrix form of Banach space; p ij Represents the element in the i-th row and j-th column of the matrix; i = 1…m, j = 1…n; specifically, Where k is the index of the energy spectrum vector value, and n is the number of columns in the matrix.
[0082] S22. Select a color map with high discrimination and good continuity to convert the feature matrix into a feature image in the RGB color space that is intuitive and easy to visually perceive. The specific process is as follows:
[0083] Each entry in the color map is an RGB triplet that defines a color, with each component intensity ranging from [0, 1]; Figure 3 The figure shows the mapping process from normalized values to actual RGB colors. It accepts a normalized value, scales it to the number of colormap entries, and returns the corresponding entry. This process enhances data visualization and helps introduce image processing and analysis techniques to mine more potential feature information.
[0084] The mapping of matrix data to RGB color space can be expressed as:
[0085]
[0086] Where Q is the image representation of the radionuclide γ spectrum in RGB color space, q ij is the element value of the image representation, m and n are the number of rows and columns of the image representation. g represents the function of color mapping using the color map. This function represents the mapping process from the normalized value to the actual RGB color, accepts a value normalized and scaled to the number of color map entries, and returns the corresponding entry, N c is the number of entries in the colormap. p ij is the element in the i-th row and j-th column of the energy spectrum matrix.
[0087] S3. The radionuclide energy spectrum feature image in the RGB color space is used as the source domain model input, and the global and local features of the radionuclide energy spectrum are obtained through multi-level deep visual descriptor migration.
[0088] S31. Fine-tune the source domain model based on the target domain data: The process of fine-tuning the source domain model based on the target domain data is as follows: fix some layers and do not update the parameter gradient; the learning rate is 5e -5 , optimizer SGD stochastic gradient descent, loss function cross entropy, epochs is 20.
[0089] S32. Use the output of the convolutional layer and the fully connected layer of the source domain model as image descriptors to perform activation vector migration. The source domain model is constructed and the activation vector is migrated. The specific process is as follows:
[0090] In public datasets The source domain model obtained through pre-training is:
[0091]
[0092] Where, For public datasets The sample space of Its corresponding category label space, f PM To use public datasets A pre-trained source domain model that enables reasoning from samples to class labels;
[0093] Activation vectors extracted from one or more layers of the source domain model:
[0094] A=φ(fPM,Q,l);
[0095] Where A represents the extracted activation vector, f PM To use public datasets The pre-trained source domain model, Q is the input image, l represents the index (or identity) of the selected single layer or multiple layers, and φ is a function that extracts the activation vector from one or more layers (convolutional layers or fully connected layers) of the source domain model.
[0096] S33. Obtain the activation map corresponding to the feature map, which contains key information of the input data at different levels and dimensions. The specific process is as follows:
[0097] like A=C l ={c i |i∈{1,2,…,M}};
[0098]
[0099] Where, is a collection of convolutional layers, is a collection of fully connected layers.
[0100] If l represents one or more layers in the convolutional layer, then the set of activation vectors C passed from the convolutional layer is defined as l , A represents the extracted activation vector, M is the number of activation vectors passed from the convolutional layer, c i is the activation vector from the i-th convolutional layer.
[0101] c iis the activation vector of the convolutional layer, which consists of multiple feature maps, c (i) is the feature map generated by each convolution kernel in the i-th convolution layer after performing the convolution operation on the input data, d (i) is the number of features extracted in the i-th convolutional layer.
[0102] like
[0103] Where, if l represents one or more layers in the fully connected layer, then the set of activation vectors transferred from the fully connected layer is defined as A represents the extracted activation vector, M is the number of activation vectors passed from the fully connected layer, and f j is the activation vector from the jth convolutional layer.
[0104] In S34, the fusion operation involves combining, transforming, or aggregating activation maps. The goal is to generate a new vector that can combine the advantages of the original activation maps, reduce redundancy, and generate a one-dimensional feature vector. The specific process is as follows:
[0105] Traverse all elements of the feature map and perform global average pooling:
[0106]
[0107] Where, Representation feature map The value of a specific element at position (h,w); h o and w o are the height and width of the feature map, respectively, and are calculated using the following formula.
[0108]
[0109] Where h i and w i Represents the height and width of the input image, h k and w k They represent the height and width of the convolution kernel respectively, pad and str represent the padding and stride respectively.
[0110] One-dimensional feature vector generation:
[0111] like
[0112] Where, is the set of convolutional layers; c i is the final convolutional layer feature; i represents c i is the activation vector from the i-th convolutional layer. c (i) is the feature map generated by each convolution kernel in the i-th convolution layer after performing the convolution operation on the input data, d (i)is the number of features extracted in the i-th convolutional layer. Represents the dth extracted from the i-th convolutional layer (i) The final vector of the features after global average pooling.
[0113] like
[0114] Where, is the set of fully connected layers; f j is the fully connected layer feature finally obtained; j represents f j is the activation vector from the jth fully connected layer. Each activation vector f j (j=1,2,…,N) is a one-dimensional vector of length K, where N is the number of activation vectors and K is the number of neurons in the jth fully connected layer.
Claims
1. A method for extracting energy spectrum features by integrating spatial transformation projection and depth migration, characterized in that: The steps include: S1. Based on the task requirements and data characteristics, pre-training is performed on public datasets to learn general feature representations. The resulting deep convolutional neural network model is used as the source domain model, and the radionuclide energy spectrum data is used as the target domain. S2. Through vector-matrix-image cascade representation transformation, the characteristic vector of the gamma energy spectrum from the Euclidean space is represented as a characteristic matrix in the Banach space, and further mapped into a characteristic image in the RGB color space; S3. The radionuclide energy spectrum feature image in the RGB color space is used as the source domain model input, and the global and local features of the radionuclide energy spectrum are obtained through multi-level deep visual descriptor migration.
2. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 1 is characterized in that: In step S1, a small amount of labeled measured radionuclide gamma energy spectrum data obtained by measuring the radiation source through the detector is used as the target domain: an open experimental area that meets safety standards is selected as the fixed operating point of the experiment; and the measured energy spectrum samples are obtained by setting the measurement distance, measurement time, and measurement angle.
3. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 1 is characterized in that: The specific process of step S2 is as follows: S21, through the vector-matrix-image cascade representation transformation, the characteristic vector of the gamma energy spectrum from the Euclidean space is represented as a characteristic matrix in the Banach space; S22. Select a color map with high discrimination and good continuity to convert the feature matrix into a feature image in the RGB color space that is intuitive and easy to visually perceive.
4. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 3 is characterized in that: The specific process of step S21 is as follows: Let the radionuclide γ spectrum s be a wide-sense stationary random vector in Euclidean space, expressed as s = {s k }, where s k is the number of k-th energy events distributed in the energy domain; ideally, s k It is positively correlated with the duration of the measurement, that is: s k =a k ·t Where t is the measurement duration, α k is the parameter of the kth count value affected by environmental background noise, measurement angle and distance, and radiation source intensity; The matrix of radionuclide energy spectrum in Banach space can be expressed as: Where f is the mapping of the radionuclide gamma spectrum s from the vector form of Euclidean space to the matrix form of Banach space; p ij Represents the element in the i-th row and j-th column of the matrix; i = 1…m, j = 1…n; specifically, Where k is the index of the energy spectrum vector value, and n is the number of columns in the matrix.
5. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 3 is characterized in that: The specific process of step S22 is as follows: Each entry in the colormap is an RGB triplet that defines a color, with each component intensity ranging from [0, 1]. This accepts a normalized value, scaled to the number of colormap entries, and returns the corresponding entry. This process enhances data visualization and facilitates the introduction of image processing and analysis techniques to mine more potential feature information. The mapping of matrix data to RGB color space can be expressed as: Where Q is the image representation of the radionuclide γ spectrum in RGB color space, q ij is the element value of the image representation, m and n are the number of rows and columns of the image representation; g represents the function of color mapping using the color map; this function represents the mapping process from the normalized value to the actual RGB color, accepts a value normalized and scaled to the number of color map entries, and returns the corresponding entry, B c is the number of entries in the colormap; p ij is the element in the i-th row and j-th column of the energy spectrum matrix.
6. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 1 is characterized in that: The specific process of step S3 is as follows: S31. Fine-tune the source domain model based on target domain data; S32. Use the output of the convolutional layer and the fully connected layer of the source domain model as the image descriptor to perform activation vector migration. Source domain model construction and activation vector migration; S33, obtaining the activation map corresponding to the feature map, which contains key information of the input data at different levels and dimensions; S33. The fusion operation involves the combination, transformation or aggregation of activation maps, aiming to generate a new vector that can integrate the advantages of the original activation map, reduce redundancy, and generate a one-dimensional feature vector.
7. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 6 is characterized in that: In step S31, the source domain model is fine-tuned based on the target domain data as follows: some layers are fixed and the parameter gradients are not updated; the learning rate is 5e -5 , optimizer SGD stochastic gradient descent, loss function cross entropy, epochs is 20.
8. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 6 is characterized in that: In step S32, the outputs of the convolutional layer and the fully connected layer of the source domain model are used as image descriptors to perform activation vector migration source domain model construction and activation vector migration; the specific process is as follows: In public datasets The source domain model obtained through pre-training is: Where, For public datasets The sample space of Its corresponding category label space, f PM To use public datasets A pre-trained source domain model that enables reasoning from samples to class labels; Activation vectors extracted from one or more layers of the source domain model: A=φ(fPM,Q,l); Where A represents the extracted activation vector, f PM To use public datasets The pre-trained source domain model, Q is the input image, l represents the index (or identity) of the selected single layer or multiple layers; φ is a function that extracts the activation vector from one or more layers (convolutional layer or fully connected layer) of the source domain model.
9. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 6 is characterized in that: In step S33, the activation map corresponding to the feature map is obtained, which contains key information of the input data at different levels and dimensions. The specific process is as follows: like A=C l ={c i |i∈{1,2,…,M}}; Where, is a collection of convolutional layers, is a collection of fully connected layers; If l represents one or more layers in the convolutional layer, then the set of activation vectors C passed from the convolutional layer is defined as l , A represents the extracted activation vector, M is the number of activation vectors passed from the convolutional layer, c i is the activation vector from the i-th convolutional layer; c i is the activation vector of the convolutional layer, which consists of multiple feature maps, c (i) is the feature map generated by each convolution kernel in the i-th convolution layer after performing the convolution operation on the input data, d (i) is the number of features extracted in the i-th convolutional layer; like Where, if l represents one or more layers in the fully connected layer, then the set of activation vectors transferred from the fully connected layer is defined as A represents the extracted activation vector, N is the number of activation vectors passed from the fully connected layer, and f j is the activation vector from the jth fully connected layer.
10. The energy spectrum feature extraction method of the fusion of space conversion projection and depth migration according to claim 6 is characterized in that: In step S34, the fusion operation involves combining, transforming or aggregating the activation maps, aiming to generate a new vector that can integrate the advantages of the original activation maps, reduce redundancy, and generate a one-dimensional feature vector; The specific process is as follows: Traverse all elements of the feature map and perform global average pooling: Where, Representation feature map The value of a specific element at position (h,w); h o and w o are the height and width of the feature map, respectively, and are calculated by the following formula; Where h i and w i Respectively represent the height and width of the input image, h k and w k Represent the height and width of the convolution kernel respectively, pad and str represent padding and stride respectively; One-dimensional feature vector generation: like Where, is the set of convolutional layers; c i is the final convolutional layer feature; i stands for c i is the activation vector from the i-th convolutional layer; c (i) is the feature map generated by each convolution kernel in the i-th convolution layer after performing the convolution operation on the input data, d (i) is the number of features extracted in the i-th convolutional layer; Represents the dth extracted from the i-th convolutional layer (i) The final vector obtained after global average pooling of features; like Where, is the set of fully connected layers; f j is the fully connected layer feature finally obtained; j represents f j is the activation vector from the jth fully connected layer; each activation vector f j (j=1,2,…,N) is a one-dimensional vector of length K, where N is the number of activation vectors and K is the number of neurons in the j-th fully connected layer.