EMIS reconstruction method and system based on composite deep learning

Abstract

The present invention provides an EMIS reconstruction method and system based on composite deep learning, the method comprising: inputting electromagnetic wave scattering data; optimizing the contrast image using a complex valued convolutional neural network (CVCNNs) to perform preliminary contrast recovery; optimizing the preliminary contrast image output by the CVCNNs using a deep residual convolutional neural network (DRCNNs) to complete a feature fusion operation. The present invention solves the technical problems of low reconstruction accuracy, poor algorithm generalization ability, high computational efficiency and low cost, and multiple scattering effects in high contrast and strong scattering environments.

Description

Technical Field

[0001] The present invention relates to the field of radar image detection and recognition and artificial intelligence algorithms, and in particular to an EMIS reconstruction method and system based on composite deep learning. Background Art

[0002] In the traditional solution to the electromagnetic inverse scattering problem (EMIS), iterative calculation methods such as the Born iteration method and the Rytov approximation are mainly used. For example, the existing invention patent application document "A method for reconstructing medium targets by electromagnetic inverse scattering" with publication number CN105677937A uses the Maxwell equations and the constitutive equation to derive the forward and reverse electromagnetic scattering integral equations; guess the shape of the object to be measured, and discretize the integral equation using the method to obtain the matrix equation; in the discretized inverse scattering equation, the Gauss-Newton minimization method is used to regularize the ill-conditioned equation, and the regularization parameter used is obtained by the multiplicative regularization method; the Born iteration method is used to iterate, and after reaching the required accuracy, the wave number of the medium of the measured object is obtained, thereby realizing the reconstruction of the medium target. And the existing invention patent application document "An electromagnetic inverse scattering imaging method based on a deep cascade network without training" with publication number CN115018943A, the aforementioned existing method includes: 1. Data preparation stage, obtaining input data according to the measured scattering field; 2. Network structure construction stage, simulating the process based on the subspace optimization iterative algorithm, and designing the deep cascade network structure; 3. Loss function design, in the objective function, in addition to the data loss and state loss of the scatterer to be sought that does not contain the reference true value, adding the total current loss and dielectric constant loss of several other known scatterers containing prior information as regularization terms to improve the reconstruction quality; 4. By iteratively optimizing the deep cascade network, a single scatterer to be sought is modeled to obtain the reconstruction result of the single scatterer to be sought. The Born iteration method used in the aforementioned prior art is suitable for processing low-contrast situations. It is based on the first-order scattering approximation, but in high-contrast situations, its accuracy is greatly reduced due to the multiple scattering effect. In contrast, the Rytov approximation is suitable for scenes with slightly higher contrast, and it handles phase delay and amplitude changes by logarithmic transformation of the wave equation, but it is also limited by contrast and multiple scattering effects. These traditional methods perform poorly in high-contrast or strong scattering environments, and require a lot of computing resources and time when processing complex or large-size data sets. With the development of deep learning technology, new methods based on deep learning provide new possibilities for dealing with EMIS problems in high-contrast and complex scattering environments.

[0003] Disadvantages of existing technology:

[0004] In the field of electromagnetic inverse scattering (EMIS) research, although traditional methods such as Born iteration and Rytov approximation are effective in some cases, they face obvious technical challenges when dealing with high contrast and strong scattering environments:

[0005] High-contrast problem: In high-contrast scenes, the dielectric constant or conductivity of an object is very different from that of its surroundings. Traditional methods often cannot accurately reconstruct object properties in this case because the linear approximation or single scattering assumption they are based on is no longer valid.

[0006] Multiple scattering effect: In a strong scattering environment, electromagnetic waves may be reflected and refracted multiple times on multiple surfaces before reaching the receiver. Traditional methods are usually unable to effectively handle such multiple scattering, resulting in a decrease in the accuracy and resolution of the reconstructed image.

[0007] Computational efficiency and resource consumption: Traditional iterative methods usually require a lot of computing time and resources, especially when dealing with large-scale datasets or requiring high-resolution reconstruction. This not only increases the time cost, but also limits the feasibility of these methods in real-time or large-scale applications.

[0008] Limited generalization ability: Traditional methods may need to adjust parameters or algorithms in different types of data sets or different physical scenarios, showing limited generalization ability. This means that they may not be as effective in new or changed application scenarios.

[0009] In summary, the existing technology has technical problems such as low reconstruction accuracy in high contrast and strong scattering environments, poor algorithm generalization ability, high computational efficiency and low cost, and multiple scattering effects. Summary of the invention

[0010] The technical problem to be solved by the present invention is how to solve the technical problems of low reconstruction accuracy, poor algorithm generalization ability, high computational efficiency and low cost, and multiple scattering effect in high contrast and strong scattering environments in the prior art.

[0011] The present invention adopts the following technical solutions to solve the above technical problems: The EMIS reconstruction method based on composite deep learning includes:

[0012] S1. Collect and input electromagnetic wave scattering data to construct an integral equation for electromagnetic inverse scattering problem;

[0013] S2, using complex-valued convolutional neural networks CVCNNs, processing according to the integral equation of the electromagnetic inverse scattering problem to obtain a contrast pair image, and using deep residual convolutional neural networks DRCNNs to perform residual learning to optimize the contrast image to obtain a first contrast restored image;

[0014] S3. Using the deep residual convolutional neural network DRCNNs, the first contrast restored image is re-optimized to complete the feature fusion operation and obtain a second contrast optimized image.

[0015] The present invention integrates complex-valued convolutional neural networks (CVCNNs) and deep residual convolutional neural networks (DRCNNs). The complex-valued convolutional neural networks (CVCNNs) process complex data, which can effectively capture the amplitude and phase information of the scattering data, and provide an accurate basis for the preliminary reconstruction of the contrast image; the present invention uses deep residual convolutional neural networks (DRCNNs) to further optimize the preliminary contrast image, and uses its residual learning mechanism to improve the details and accuracy of the reconstructed image, which is suitable for high-contrast and complex scattering scenes. The network outputs the final optimized contrast image, provides more accurate and detailed results, and improves the accuracy of the reconstruction of the electromagnetic inverse scattering problem. The present invention improves the accuracy and efficiency of electromagnetic wave scattering data processing, and has significant advantages over traditional technologies in the application scenario of electromagnetic inverse scattering reconstruction.

[0016] In a more specific technical solution, in step S1, the electromagnetic wave scattering data includes: amplitude and phase information of the scattered wave.

[0017] In a more specific technical solution, in step S1, the integral equation of the electromagnetic inverse scattering problem is constructed using the following logic:

[0018] E scattered (r) = ∫ V G(r,r′)·χ(r′)·E incident (r′)dr′

[0019] In the formula, E scattered (r) is the scattered field at position r, G(r,r′) is the Green’s function, which represents the influence of the source at position r′ on position r, χ(r′) is the electromagnetic property of the scatterer, and E incident (r′) is the incident field and V is the volume of the scatterer.

[0020] In a more specific technical solution, step S2 includes:

[0021] S21, annotating and removing noise from the electromagnetic wave scattering data to obtain complex input data;

[0022] S22, using complex-valued convolutional neural networks (CVCNNs) to receive and process amplitude information and phase information in complex input data;

[0023] S23, using deep residual convolutional neural network DRCNNs, performing multi-layer convolution operations to extract electromagnetic wave scattering features;

[0024] S24. Generate a first contrast-restored image according to the electromagnetic wave scattering characteristics.

[0025] The present invention uses deep residual convolutional neural networks (DRCNNs) to optimize the preliminary contrast image output by CVCNNs. The present invention uses DRCNNs for residual learning, allowing the network to focus on learning the small differences between input and output, which is suitable for refining and enhancing contrast images. At the same time, the present invention uses CVCNNs for the complex characteristics of electromagnetic scattering data, which can simultaneously process amplitude and phase information, thereby improving the accuracy and quality of the final reconstructed image.

[0026] In a more specific technical solution, in step S21, the electromagnetic wave scattering data is converted into complex input data X using the following logic: complex :

[0027] X complex =X amplitude +i·X phase

[0028] Where, X amplitude and X phase are amplitude and phase data respectively.

[0029] In a more specific technical solution, in step S22, the following logic is used to simultaneously process the amplitude information and the phase information:

[0030] Z=f(W·X complex + b)

[0031] Where Z is the network output, W is the weight, b is the bias, f is the activation function, and X complex is a complex input.

[0032] In a more specific technical solution, in step S23, each convolution layer in the multi-layer convolution operation is performed using the following logic:

[0033]

[0034] In the formula, F ij is a point on the feature map, σ is the activation function, K mn is the convolution kernel, X (i+m)(j+n) is the input data, B is the bias, m and n: these two parameters represent the convolution kernel K mn In the image data X (i+m)(j+n) In the convolution operation, the convolution kernel slides along the height (m direction) and width (n direction) of the input image to calculate the feature map F ij Every point on .

[0035] In a more specific technical solution, step S3 includes:

[0036] S31, receiving a first contrast-restored image, and performing a contrast re-optimization operation on the first contrast-restored image using a deep residual convolutional neural network DRCNNs;

[0037] S32, residual learning using deep residual convolutional neural networks DRCNNs;

[0038] S33. Use deep residual convolutional neural network DRCNNs to perform deep learning to perform contrast refinement and optimization operations on the first contrast restored image, so as to obtain a second contrast optimized image.

[0039] The present invention uses DRCNNs for deep learning, refines and optimizes the initial contrast image, and uses more layers and residual connections to enhance feature extraction and correction capabilities. DRCNNs deepens the recognition and optimization of complex features through residual learning, especially in high-contrast and complex scattering scenes.

[0040] The present invention combines the preliminary feature extraction capability of CVCNNs with the deep feature optimization advantage of DRCNNs, significantly improving the processing efficiency and reconstruction accuracy of electromagnetic inverse scattering problems. This fusion strategy not only speeds up the processing speed and reduces the computational cost, but also enhances the generalization ability of the model when processing scattering data of different types and complexity, making it perform well in a variety of application scenarios.

[0041] In a more specific technical solution, in step S32, residual learning is performed using the following logic:

[0042] R l =H l (X l )-X l

[0043] In the formula, R l is the residual of the ιth layer, H l (X l ) is the output of the ιth layer, X l is the input.

[0044] In a more specific technical solution, the EMIS reconstruction system based on composite deep learning includes:

[0045] An electromagnetic inverse scattering problem integral equation construction module is used to collect and input electromagnetic wave scattering data to construct an electromagnetic inverse scattering problem integral equation;

[0046] A first contrast restoration module is used to use complex-valued convolutional neural networks CVCNNs to process the contrast pair image according to the electromagnetic inverse scattering problem integral equation, and to use deep residual convolutional neural networks DRCNNs to perform residual learning to optimize the contrast image to obtain a first contrast restored image. The first contrast restoration module is connected to the electromagnetic inverse scattering problem integral equation construction module;

[0047] The second contrast restoration module is used to use the deep residual convolutional neural network DRCNNs to re-optimize the first contrast restored image to complete the feature fusion operation and obtain the second contrast optimized image. The second contrast restoration module is connected to the first contrast restoration module.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] The present invention integrates complex-valued convolutional neural networks (CVCNNs) and deep residual convolutional neural networks (DRCNNs). The complex-valued convolutional neural networks (CVCNNs) process complex data, which can effectively capture the amplitude and phase information of the scattering data, and provide an accurate basis for the preliminary reconstruction of the contrast image; the present invention uses deep residual convolutional neural networks (DRCNNs) to further optimize the preliminary contrast image, and uses its residual learning mechanism to improve the details and accuracy of the reconstructed image, which is suitable for high-contrast and complex scattering scenes. The network outputs the final optimized contrast image, provides more accurate and detailed results, and improves the accuracy of the reconstruction of the electromagnetic inverse scattering problem. The present invention improves the accuracy and efficiency of electromagnetic wave scattering data processing, and has significant advantages over traditional technologies in the application scenario of electromagnetic inverse scattering reconstruction.

[0050] The present invention uses deep residual convolutional neural networks (DRCNNs) to optimize the preliminary contrast image output by CVCNNs. The present invention uses DRCNNs for residual learning, allowing the network to focus on learning the small differences between input and output, which is suitable for refining and enhancing contrast images. At the same time, the present invention uses CVCNNs for the complex characteristics of electromagnetic scattering data, which can simultaneously process amplitude and phase information, thereby improving the accuracy and quality of the final reconstructed image.

[0051] The present invention uses DRCNNs for deep learning, refines and optimizes the initial contrast image, and uses more layers and residual connections to enhance feature extraction and correction capabilities. DRCNNs deepens the recognition and optimization of complex features through residual learning, especially in high-contrast and complex scattering scenes.

[0052] The present invention combines the preliminary feature extraction capability of CVCNNs with the deep feature optimization advantage of DRCNNs, significantly improving the processing efficiency and reconstruction accuracy of electromagnetic inverse scattering problems. This fusion strategy not only speeds up the processing speed and reduces the computational cost, but also enhances the generalization ability of the model when processing scattering data of different types and complexity, making it perform well in a variety of application scenarios.

[0053] The present invention solves the technical problems existing in the prior art of low reconstruction accuracy in high contrast and strong scattering environments, poor algorithm generalization ability, high computational efficiency and low cost, and multiple scattering effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of the basic steps of the EMIS reconstruction method based on composite deep learning according to Example 1 of the present invention;

[0055] Figure 2 This is a data flow processing schematic diagram of an EMIS reconstruction method based on composite deep learning according to Embodiment 1 of the present invention;

[0056] Figure 3 Schematic diagram of specific steps of preliminary contrast restoration in Example 1 of the present invention;

[0057] Figure 4 This is a schematic diagram of specific steps for optimizing a preliminary contrast image according to Embodiment 1 of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Example 1

[0060] like Figure 1 and Figure 2 As shown, the EMIS reconstruction method based on composite deep learning provided by the present invention includes the following basic steps:

[0061] S1. Input electromagnetic wave scattering data;

[0062] In this embodiment, the electromagnetic inverse scattering problem involves inferring the electromagnetic properties of an object from the scattering data of electromagnetic waves. In this embodiment, the electromagnetic properties of the object include but are not limited to: shape, size, material composition and internal structure. The electromagnetic scattering problem has a wide range of applications in many fields, such as radar detection, medical imaging, geological exploration, etc.

[0063] In the scattering phenomenon, scattering occurs when electromagnetic waves encounter objects with different electromagnetic properties. The characteristics of the scattered waves (such as amplitude, phase, direction) will vary depending on the electromagnetic properties of the object;

[0064] In the collection operation of scattered data, the scattered data is the result of the interaction between electromagnetic waves and objects, and can be used to infer the characteristics of the objects. In this embodiment, the scattered data includes: but is not limited to: the amplitude and phase information of the scattered waves;

[0065] In this embodiment, the inverse scattering problem refers to inferring the electromagnetic properties of the original object from the collected scattering data. The inverse scattering problem is a highly complex and nonlinear problem because the same scattering data may correspond to a variety of different object properties.

[0066] In this embodiment, the integral equation of the electromagnetic inverse scattering problem is as follows:

[0067] E scattered (r) = ∫ V G(r,r ′ )·χ(r ′ 0·E incident (r ′ )dr ′

[0068] Among them, E scattered (r) is the scattered field at position r, G(r,r ′ ) is the Green function, which represents the position r ′ The influence of the source at position r, χ(r ′ ) is the electromagnetic property of the scatterer, E incident (r ′ ) is the incident field, V is the volume of the scatterer;

[0069] S2, using complex-valued convolutional neural networks (CVCNNs) to optimize the contrast image and perform preliminary contrast restoration;

[0070] In this embodiment, deep residual convolutional neural networks (DRCNNs) are used to optimize the preliminary contrast images output by CVCNNs. DRCNNs are characterized by the application of residual learning, which allows the network to focus on learning the small differences between input and output. This method is particularly suitable for refining and enhancing contrast images, improving the accuracy and quality of the final reconstructed image.

[0071] like Figure 3 As shown, in this embodiment, the step S2 of preliminary contrast restoration also includes the previous specific steps:

[0072] S21, electromagnetic scattering data preprocessing;

[0073] In the data preprocessing process of this embodiment, the electromagnetic scattering data is first preprocessed; in this embodiment, the preprocessing operation includes but is not limited to: standardization (scaling the data to a certain range) and noise removal, and converting the data into a complex form, which is formally expressed as:

[0074] X complex =X amplitude +i·X phase

[0075] Among them, X amplitude and X phase are amplitude and phase data, respectively;

[0076] S22, complex input processing;

[0077] In the complex input processing process of this embodiment, CVCNNs can receive and process complex inputs, that is, the network can process amplitude and phase information at the same time. In this embodiment, the processing process of processing amplitude and phase information at the same time can be expressed as the following formula:

[0078] Z=f(W·X complex +b)

[0079] Among them, Z is the network output, W is the weight, b is the bias, f is the activation function, X complex is a complex input;

[0080] S23, extracting electromagnetic wave scattering data features;

[0081] In the feature extraction process of this embodiment, the deep residual convolutional neural network (DRCNNs) extracts the features of the data through multi-layer convolution operations; in this embodiment, each layer of convolution can be expressed as the following formula:

[0082]

[0083] Among them, F ij is a point on the feature map, σ is the activation function, K mn is the convolution kernel, X (i+m)(j+n) is the input data, B is the bias;

[0084] S24, generating a preliminary contrast image;

[0085] In this embodiment, the final output of CVCNNs is a preliminary contrast image representing an estimate of the electromagnetic properties of the object under test;

[0086] S3. Use deep residual convolutional neural networks (DRCNNs) to optimize the preliminary contrast image output by CVCNNs to complete the feature fusion operation.

[0087] like Figure 4 As shown, step S3 of optimizing the preliminary contrast image further includes the following specific steps:

[0088] S31, receiving a preliminary contrast map;

[0089] In this embodiment, in the received preliminary contrast image, DRCNNs uses the output of CVCNNs as input to further process and optimize the contrast;

[0090] S32, residual learning;

[0091] In the residual learning process of this embodiment, the key feature of DRCNNs is residual learning, which allows the network to learn the difference between input and output. This can be expressed by the following formula:

[0092] R l =H l (X l )-X l

[0093] Where R l is the residual of the lth layer, H l (X l ) is the output of layer l, X l is the input;

[0094] S33, contrast refinement and optimization;

[0095] In this embodiment, DRCNNs refine and optimize the initial contrast image through deep learning, using more layers and residual connections to enhance feature extraction and correction capabilities;

[0096] S34: Output the final image.

[0097] The network outputs a final optimized contrast image, providing more accurate and detailed results.

[0098] In summary, the present invention integrates complex-valued convolutional neural networks (CVCNNs) and deep residual convolutional neural networks (DRCNNs). By processing complex data through complex-valued convolutional neural networks (CVCNNs), the amplitude and phase information of scattering data can be effectively captured, providing an accurate basis for the preliminary reconstruction of contrast images; the present invention further optimizes the preliminary contrast image using deep residual convolutional neural networks (DRCNNs), and improves the details and accuracy of the reconstructed image using its residual learning mechanism, which is suitable for high-contrast and complex scattering scenes. The network outputs the final optimized contrast image, providing more accurate and detailed results, and improving the accuracy of reconstruction of electromagnetic inverse scattering problems. The present invention improves the accuracy and efficiency of electromagnetic wave scattering data processing, and has significant advantages over traditional technologies in the application scenario of electromagnetic inverse scattering reconstruction.

[0099] The present invention uses deep residual convolutional neural networks (DRCNNs) to optimize the preliminary contrast image output by CVCNNs. The present invention uses DRCNNs for residual learning, allowing the network to focus on learning the small differences between input and output, which is suitable for refining and enhancing contrast images. At the same time, the present invention uses CVCNNs for the complex characteristics of electromagnetic scattering data, which can simultaneously process amplitude and phase information, thereby improving the accuracy and quality of the final reconstructed image.

[0100] The present invention uses DRCNNs for deep learning, refines and optimizes the initial contrast image, and uses more layers and residual connections to enhance feature extraction and correction capabilities. DRCNNs deepens the recognition and optimization of complex features through residual learning, especially in high-contrast and complex scattering scenes.

[0101] The present invention combines the preliminary feature extraction capability of CVCNNs with the deep feature optimization advantage of DRCNNs, significantly improving the processing efficiency and reconstruction accuracy of electromagnetic inverse scattering problems. This fusion strategy not only speeds up the processing speed and reduces the computational cost, but also enhances the generalization ability of the model when processing scattering data of different types and complexity, making it perform well in a variety of application scenarios.

[0102] The present invention solves the technical problems existing in the prior art of low reconstruction accuracy in high contrast and strong scattering environments, poor algorithm generalization ability, high computational efficiency and low cost, and multiple scattering effects.

[0103] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The EMIS reconstruction method based on composite deep learning is characterized by: The method comprises: S1. Collect and input electromagnetic wave scattering data to construct an integral equation for electromagnetic inverse scattering problem. In step S1, the integral equation for the electromagnetic inverse scattering problem is constructed using the following logic: AND scattered (r)=∫ V G(r,r′) x(r′) E incident (r′)dr′ In the formula, E scattered (r) is the scattered field at position r, G(r, r′) is the Green’s function, which represents the influence of the source at position r′ on position r, x(r′) is the electromagnetic property of the scatterer, and E incident (r′) is the incident field, V is the volume of the scatterer; S2, using complex-valued convolutional neural networks CVCNNs to process the electromagnetic inverse scattering problem integral equation to obtain a contrast image, and using deep residual convolutional neural networks DRCNNs to perform residual learning to optimize the contrast image to obtain a first contrast restored image; Wherein, the step S2 comprises: S21, performing labeling and noise removal operations on the electromagnetic wave scattering data to obtain complex input data; S22, using the complex-valued convolutional neural network CVCNNs to receive and process the amplitude information and phase information in the complex input data; S23, using the deep residual convolutional neural network DRCNNs to perform multi-layer convolution operations to extract electromagnetic wave scattering characteristics; S24, generating the first contrast-restored image according to the electromagnetic wave scattering characteristics; S3, using the deep residual convolutional neural network DRCNNs, re-optimizing the first contrast restored image to complete the feature fusion operation and obtain a second contrast optimized image; Wherein, the step S3 comprises: S31, receiving the first contrast-restored image, and performing a contrast re-optimization operation on the first contrast-restored image using the deep residual convolutional neural network DRCNNs; S32, using the deep residual convolutional neural network DRCNNs to perform residual learning; S33. Use the deep residual convolutional neural network DRCNNs to perform deep learning to perform contrast refinement and optimization operations on the first contrast restored image, thereby obtaining the second contrast optimized image.

2. The EMIS reconstruction method based on composite deep learning according to claim 1, characterized in that: In the step S1, the electromagnetic wave scattering data includes: amplitude and phase information of the scattered wave.

3. The EMIS reconstruction method based on composite deep learning according to claim 1, characterized in that: In the step S21, the electromagnetic wave scattering data is converted into the complex input data using the following logic: X complex =X amplitude +i·X phase In the formula, X amplitude and X phase are amplitude and phase data respectively.

4. The EMIS reconstruction method based on composite deep learning according to claim 1, characterized in that: In step S22, the amplitude information and the phase information are processed simultaneously using the following logic: Z=f(W·X complex +b) Where Z is the network output, W is the weight, b is the bias, f is the activation function, and X complex is a complex input.

5. The EMIS reconstruction method based on composite deep learning according to claim 1, characterized in that: In step S23, each convolution layer in the multi-layer convolution operation is performed using the following logic: In the formula, F ij is a point on the feature map, σ is the activation function, K mn is the convolution kernel, X (i+m)(j+n) is the input data, B is the bias, m and n represent the convolution kernel K respectively. mn In the image data X (i+m)(j+n) Position offset on .

6. The EMIS reconstruction method based on composite deep learning according to claim 1, characterized in that: In step S32, the residual learning is performed using the following logic: R l =H l (X l )-X l In the formula, R l is the residual of the lth layer, H l (X l ) is the output of layer l, X l is the input.

7. An EMIS reconstruction system based on composite deep learning, used to execute the EMIS reconstruction method based on composite deep learning according to any one of claims 1 to 6, characterized in that: The system comprises: An electromagnetic inverse scattering problem integral equation construction module is used to collect and input electromagnetic wave scattering data to construct an electromagnetic inverse scattering problem integral equation; A first contrast restoration module, used to use complex-valued convolutional neural networks CVCNNs to process the electromagnetic inverse scattering problem integral equation to obtain a contrast image, and use deep residual convolutional neural networks DRCNNs to perform residual learning to optimize the contrast image to obtain a first contrast restored image, wherein the first contrast restoration module is connected to the electromagnetic inverse scattering problem integral equation construction module; The second contrast restoration module is used to use the deep residual convolutional neural network DRCNNs to re-optimize the first contrast restored image to complete the feature fusion operation and obtain a second contrast optimized image. The second contrast restoration module is connected to the first contrast restoration module.

Images