High-resolution radiometer intelligent imaging method and system
Through feature encoding strategy and parameter adaptive pulse coupled neural network, the diffraction limit and noise problems in microwave radiometer imaging are solved, high-resolution imaging is achieved, adaptability to diverse conditions and computational cost are reduced.
Patent Information
- Application Number
- CN202510909655.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-09-19
AI Technical Summary
Existing microwave radiometer imaging technology is limited by the diffraction limit, making it difficult to achieve high-resolution imaging in complex noisy environments. Existing methods also rely on large-scale training data or improper parameter settings, resulting in poor imaging results.
A feature coding strategy and parameter-adaptive pulse-coupled neural network are adopted to reduce the impact of noise and improve imaging resolution through data preprocessing, feature coding and parameter-adaptive pulse-coupled neural network processing.
It achieves high-resolution imaging in complex noisy environments, breaks through the diffraction limit, improves the signal-to-noise ratio and imaging quality, reduces computing costs, and adapts to diverse imaging conditions.
Smart Images
Figure CN120668268A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of microwave / millimeter wave radiation imaging detection technology, and in particular to a high-resolution radiometer intelligent imaging method and system. Background Art
[0002] Microwave radiometers are passive sensors that detect the thermal radiation of microwave signals. These signals are related to the physical temperature and electrical properties of the observed object and are also affected by the intervening atmosphere. These sensors primarily rely on real-aperture antennas (real-aperture scanning radiometers, RASRs) or interferometric antenna arrays (synthetic-aperture interferometric radiometers, SAIRs) and are widely used in astronomical observations and Earth remote sensing. Compared to active radar sensors, radiometers offer superior stealth and can operate in all-weather and near-all-weather conditions, making them valuable for target detection on land and at sea. Furthermore, radiometer imaging can penetrate clothing and detect concealed contraband, showing great potential in personnel security screening. However, a major challenge facing radiometer imaging is obtaining clear, contrast-rich images of the observed scene to ensure effective target extraction. This challenge is largely due to the diffraction limit, which restricts the angular resolution based on the physical size of the sensor antenna or array. For SAIRs in particular, the imaging process is formulated as an inverse problem, and inherent measurement noise can propagate during the reconstruction process, severely degrading the signal-to-noise ratio (SNR).
[0003] To address these challenges, researchers have proposed a variety of methods based on array signal processing and image processing. Traditional image reconstruction methods, such as the inverse discrete Fourier transform (IDFT), estimate brightness temperature images using visibility samples measured by SAIR. However, this IDFT inversion method fails to account for non-ideal factors such as antenna pattern deviation, fringe cleaning effects, and channel inconsistencies, resulting in limited reconstruction performance. To improve performance, researchers have introduced regularization techniques in Hilbert space, including Tikhonov regularization, truncated singular value decomposition, and band-limited regularization. While these methods help balance image reconstruction error and noise, they require careful selection of trade-off parameters. Improper parameter settings can lead to over-smoothing of the target or insufficient noise suppression. Furthermore, for RASR tasks, researchers have proposed image-domain methods that incorporate polarization information or regional features, such as edge contouring, reflection removal, and object segmentation. However, due to insufficient utilization of prior knowledge and inaccurate feature representation, the improvement achieved by these methods is limited. In recent years, deep learning (especially convolutional neural networks (CNNs) and Transformers) has been applied to SAIR imaging to achieve reconstruction by mapping visibility data to brightness temperature distributions. However, these methods rely heavily on large-scale, high-quality training datasets that match the test data. In certain remote sensing missions (especially in extreme environments such as deep space exploration and polar expeditions), obtaining paired real-label data is often not feasible, which highlights the urgent need for alternative learning methods. This limitation has prompted us to explore the application of self-supervised learning in the field of radiometer imaging. In view of this, how to break through the diffraction limit constraint and achieve high-resolution imaging of microwave radiometers in complex noisy environments remains a current research hotspot and a problem that needs to be solved urgently. Summary of the Invention
[0004] The purpose of the present invention is to provide a high-resolution radiometer intelligent imaging method and system to solve the problem of breaking through the diffraction limit constraint and achieving high-resolution imaging of microwave radiometers in complex noise environments.
[0005] To achieve the above objectives, the present invention provides a high-resolution radiometer intelligent imaging system, including: a data preprocessing module, a feature encoding strategy module and a parameter-adaptive pulse-coupled neural network; the feature encoding strategy module is based on the measurement noise of uncorrelated different spatial frequencies or baseline sampling data.
[0006] Furthermore, the processing steps of the data preprocessing module are as follows: if the radiometer is an interferometric radiometer, the visibility data obtained by spatial frequency or baseline sampling needs to be linearly mapped into a one-dimensional column vector; if the radiometer is a real aperture scanning radiometer, the average value of the multi-polarization radiation image needs to be calculated, and the spatial frequency component data is extracted from these average images through linear transformation of the partial Fourier matrix, and further linearly mapped into a one-dimensional column vector.
[0007] Furthermore, the spatial frequency or baseline sampled visibility data is expressed as:
[0008]
[0009] Furthermore, the parameter-adaptive pulse-coupled neural network includes a neighborhood modulation module and a pulse generator module.
[0010] Furthermore, the model expression of the modulation module of the neighborhood is:
[0011]
[0012] L q =V L ∑ kl W q,kl Y q,kl (n-1)
[0013] The model expression of the pulse generator block is:
[0014]
[0015] The parameter selection expression of the parameter adaptive pulse coupled neural network is:
[0016]
[0017] In one aspect, the present invention provides a high-resolution radiometer intelligent imaging method, comprising the steps of:
[0018] S1: Use the data preprocessing module to preprocess the data collected by the radiometer on the detection scene, and linearly map the preprocessed spatial frequency domain samples into a one-dimensional column vector;
[0019] S2: Use the feature encoding strategy module to encode the spatial frequency samples, thereby effectively reducing the impact of radiation noise;
[0020] S3: The encoded spatial frequency samples are aggregated into a feature map through position embedding, and then input into the parameter adaptive pulse coupled neural network for processing;
[0021] S4: Acquire high-resolution reconstructed images.
[0022] Furthermore, in step S2, the step of encoding the spatial frequency samples includes:
[0023] S21: Create a spatial frequency transformation matrix, which is expressed as:
[0024]
[0025] S22: Weight the spatial frequency samples. The weighted spatial frequency samples are expressed as:
[0026]
[0027] S23: Yes Perform a cross-correlation operation, expressed as:
[0028]
[0029] S24: To T q Perform summation and square root operations to achieve spatial frequency feature encoding, and its expression is:
[0030]
[0031] In summary, the present invention has the following beneficial effects compared to the prior art:
[0032] The present invention provides a high-resolution radiometer intelligent imaging method and system. This method introduces a feature encoding strategy based on the concept that the measurement noise of visibility data sampled at different spatial frequencies (or baselines) is uncorrelated. This strategy performs correlation processing on spatial frequency samples, thereby effectively mitigating the impact of radiation noise. Furthermore, a lightweight parameter-adaptive pulse-coupled neural network for radiometer imaging is designed. This network has the advantage of extracting high-resolution features of the target and suppressing noise at a low computational cost. Furthermore, the present invention can optimize computational efficiency through the following methods: The feature encoding strategy module can be optimized in two ways: first, pre-storing the spatial-frequency transformation matrix to improve the computational efficiency of spatial-frequency feature encoding using a "space-for-time" strategy; second, from a program optimization perspective, the Kronecker product can be implemented through loop parallelization or the use of built-in functions. For the models of the neighborhood-based modulation module and the pulse generator module, increasing the exponential decay time parameter (which serves as the step size during the iteration process) can simultaneously reduce the maximum number of iterations, thereby rapidly converging to the optimal solution. After these optimizations, the average runtime of the proposed method was reduced to approximately 0.15 seconds when simulated and verified on a MATLAB 2023a platform (PC with an Intel 3.70GHz CPU and 64GB of memory). To meet the higher real-time requirements of radiometer imaging tasks, embedded deployment on a DSP+FPGA chip architecture would be an effective solution to further accelerate the proposed method.
[0033] The present invention can dynamically adapt to diverse imaging conditions and achieves high-fidelity super-resolution imaging with an improved signal-to-noise ratio by integrating space-frequency coding and high-frequency reconstruction techniques. Compared with IDFT, Tikhonov regularization, and the original PCNN implementation methods, the present invention exhibits superior performance in retaining high-frequency information and breaks through the resolution limitations imposed by the antenna or array aperture size. It has been verified in scenarios such as long-range target imaging detection, remote sensing radio frequency interference identification, and concealed detection of prohibited items in security inspections.
[0034] This invention provides strong technical support for applications such as target detection, remote sensing monitoring, and security inspection and protection, and opens up new paths for the future development of radiometer imaging technology. Even with reduced hardware complexity or compressed measurement data, it can still provide effective solutions to these challenges. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0036] Figure 1 A flow chart of the high-resolution radiometer intelligent imaging system provided by the present invention;
[0037] Figure 2 Schematic diagram of spatial frequency sample feature coding provided by the present invention;
[0038] Figure 3 Schematic diagram of the parameter-adaptive pulse-coupled neural network provided by the present invention;
[0039] Figure 4 A flow chart of the high-resolution radiometer intelligent imaging method provided by the present invention;
[0040] Figure 5 The results of the high-resolution radiometer intelligent imaging method are compared with various existing methods;
[0041] Figure 6 This is the optical scene diagram of the airport hangar;
[0042] Figure 7 To reconstruct the brightness temperature map using the IDFT method;
[0043] Figure 8 To reconstruct the brightness temperature map using the Tikhonov method;
[0044] Figure 9 To reconstruct the brightness temperature map using the original PCNN method;
[0045] Figure 10The reconstructed brightness temperature image is obtained using the high-resolution radiometer intelligent imaging method and system provided by the present invention. DETAILED DESCRIPTION
[0046] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0047] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form can also include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0048] Unless otherwise specifically stated, the relative arrangement of the parts and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to actual proportional relationships. The techniques, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods and equipment should be considered as part of the authorized specification. In all examples shown and discussed herein, any specific values should be interpreted as being merely exemplary and not as limitations. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following figures, and therefore, once an item is defined in one figure, it does not need to be further discussed in subsequent figures.
[0049] See also Figures 1 to 3 As shown, on the one hand, the present invention provides a high-resolution radiometer intelligent imaging system, including: a data preprocessing module, a feature encoding strategy module and a parameter adaptive pulse coupled neural network; the feature encoding strategy module is based on the measurement noise of uncorrelated sampling data of different spatial frequencies (or baselines).
[0050] As a preferred embodiment, the input data format of the present invention is a one-dimensional column vector spatial frequency measurement value, so the following preprocessing is required (data preprocessing module): if the radiometer is an interferometric radiometer, the visibility data obtained by spatial frequency (or baseline) sampling needs to be linearly mapped into a one-dimensional column vector; if the radiometer is a real aperture scanning radiometer, the average value of the multi-polarization radiation image needs to be calculated, and the spatial frequency component data needs to be extracted from these average images through linear transformation of the partial Fourier matrix, and further linearly mapped into a one-dimensional column vector;
[0051] As a preferred method, regarding obtaining visibility data of spatial frequency (or baseline) sampling:
[0052] This theoretical framework enables the inversion and reconstruction of the brightness temperature (BT) distribution of the observed area using visibility function data obtained from measurements of the original scene. According to the Wiener-Schinchin theorem, the cross-correlation output (i.e., visibility function) between any two receiving antennas essentially characterizes the physical properties of electromagnetic wave interference within their overlapping detection area. For far-field observations of an interferometric radiometer, the output visibility data essentially represents the spatial frequency components of the target scene. For ease of analysis, the spatial frequency component v of the radiometer is expressed as:
[0053]
[0054] Where P represents the number of spatial frequency samples; Indicates plural form.
[0055] As a preferred option, Figure 3 As shown in the figure, this paper designs a lightweight parameter-adaptive pulse-coupled neural network for radiometer imaging. The network primarily consists of a neighborhood-based modulation module and a pulse generator module. The encoded spatial frequency samples are aggregated into a feature map through position embedding, which is then fed into the parameter-adaptive pulse-coupled neural network for processing, achieving high-resolution single-frame radiometer imaging.
[0056] Regarding the modulation module of the neighborhood, the expression of its model is:
[0057]
[0058] L q =V L ∑ kl W q,kl Y q,kl (n-1)
[0059] Among them, n represents the iteration index, α f represents the exponential decay time of the internal activity term, β represents the connection strength, and L q Indicates connection input, V L Indicates the amplitude of the connected input, W q,kl An element of the weight matrix representing the connection between neurons, Y q,kl (n-1) represents the state output. All W in the range of k∈[-K,K], l∈[-L,L] q,kl Construct a convolution kernel matrix W q .
[0060] Regarding the pulse generator module, the expression of its model is:
[0061]
[0062] Among them, α e (n) represents the exponential decay time of the dynamic threshold, V E Representation and dynamic threshold E q (n) related amplification factor, R represents a newly introduced parameter used to control the firing frequency of neurons in the network. It is worth noting that Y q,kl (n) is the result obtained at the n+1th iteration, Y q,00 (n) is equivalent to Y q * (n) is the result obtained at the nth iteration. In addition, the set {Y q,kl (n):k∈[-K,K]\{0},l∈[-L,L]\{0}} represents the state output corresponding to the neighboring pixels of the qth pixel.
[0063] The parameter selection expression of the parameter adaptive pulse coupled neural network is:
[0064]
[0065] In order to simplify parameter selection and maintain the performance of the network, the exponential decay time parameter α f (n) and α e (n) is integrated into a unified parameter, denoted as α; N max Indicates the total number of iterations.
[0066] See also Figure 4 As shown, another aspect of the present invention provides a high-resolution radiometer intelligent imaging method, comprising the steps of:
[0067] S1: Use the data preprocessing module to preprocess the data collected by the radiometer on the detection scene, and linearly map the preprocessed spatial frequency domain samples into a one-dimensional column vector;
[0068] S2: Use the feature encoding strategy module to encode the spatial frequency samples, thereby effectively reducing the impact of radiation noise;
[0069] As a preference, the step of encoding the spatial frequency samples includes:
[0070] S21: Create a spatial frequency transformation matrix. The spatial frequency transformation matrix can be expressed as:
[0071]
[0072] in, ξ q and u pThey represent spatial coordinates and spatial frequency coordinates respectively, Q represents the number of spatial samples, and P represents the number of spatial frequency samples.
[0073] S22: Weight the spatial frequency samples. The weighted spatial frequency samples can be expressed as:
[0074]
[0075] Here, ⊙ represents the Hadamard product.
[0076] S23: Yes Perform a cross-correlation operation, expressed as:
[0077]
[0078] in, represents the Kronecker product, (·) H represents the conjugate transpose.
[0079] S24: To T q Perform summation and square root operations to achieve spatial frequency feature encoding, and its expression is:
[0080]
[0081] Wherein, sum(·) represents the sum of all elements in the matrix; repeating steps S21 to S24 can obtain all spatial frequency feature coding samples T1,…T q ,…,T Q .
[0082] S3: The encoded spatial frequency samples are aggregated into a feature map through position embedding, and then input into the parameter adaptive pulse coupled neural network for processing;
[0083] S4: Acquire high-resolution reconstructed images.
[0084] Example:
[0085] Figure 5 The results of the high-resolution radiometer intelligent imaging method provided by the present invention are compared with various existing methods, among which ( Figure 5 a) is the result of earth simulation imaging; ( Figure 5 b) is the imaging result of the hangar experiment scene; ( Figure 5 c) is the result of radio frequency interference (RFI) source identification based on the SMOS satellite dataset; ( Figure 5 d) Detection results of concealed contraband using the W-band multi-polarization focal plane RASR system for security inspection applications.
[0086] Experiments show that: Figure 5 a and Figure 5 As shown in Fig. 2b, the imaging results of the present invention are better than those of other methods in terms of evaluation indicators. The imaging results of IDFT and Tikhonov regularization show obvious background clutter and blurred target contours. Although PCNN provides better results, the improvement of target contours seems to be negligible, and the problem of clutter interference still exists in the restored radiometric image. It is worth noting that the present invention can effectively filter out background clutter interference and depict the target shape with higher resolution. Figure 5 The SNR / SSIM of the baseline method IDFT in a is 5.67dB / 0.20, and the SNR / SSIM of the method of the present invention is 23.93dB / 0.91; Figure 5 The SNR / Entropy of the baseline method IDFT in b is 8.87dB / 5.99, and the SNR / Entropy of the method of the present invention is 25.21dB / 0.22. Figure 5 As shown in Figure c, the traditional IDFT not only restores most of the RF signal source, but also retains the natural background and oscillation artifacts, thus hindering the accurate identification of RF signals. Although PCNN can eliminate background interference, it suffers from missed detection and false alarm in the restored image. In contrast, the present invention excels in extracting RFI sources and suppressing background oscillations. Furthermore, compared with IDFT, both PCNN and the present invention exhibit higher average PR curves and stronger robustness in different observation scenarios, as verified by the width of the corresponding color bands. It is worth noting that the PR curve of the present invention is better than that of PCNN, indicating that it has stronger global detection capability in the RFI identification task. Figure 5 d shows a comparison of the imaging results of two traditional fusion processing methods (polarization sum average PSA and discrete wavelet transform DWT), PCNN and the present invention; obviously, both learning-based methods can effectively achieve image denoising and enhance the intensity contrast between hidden contraband and the human background; compared with PSA, DWT and PCNN methods, the present invention can more completely present the morphological characteristics of hidden contraband, has better noise reduction capability (SNR is 9.97dB), and has sharper contour edges.
[0087] like Figures 6 to 10 As shown, under the condition that the spatial frequency sample sampling rate is 20%, the experiment shows that: IDFT ( Figure 7 ), Tikhonov regularization method ( Figure 8 ) and the original PCNN method ( Figure 9 ) has a significantly deteriorated imaging quality, with increased artifacts and blurred target contours. It is noteworthy that the present invention can still maintain high-resolution target imaging even when the spatial frequency sample rate is 20%. Figure 10This further demonstrates that the proposed network architecture can efficiently extract high-resolution features of targets and suppress noise at low computational cost, maintaining excellent performance even with reduced hardware or compressed measurement data. The average runtime of the proposed method is reduced to approximately 0.15 seconds. To meet the higher real-time requirements of radiometer imaging tasks, embedded deployment on a DSP+FPGA chip architecture will be an effective solution to further accelerate the proposed method.
[0088] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.
Claims
1. A high-resolution radiometer intelligent imaging system, characterized in that: include: Data preprocessing module, feature encoding strategy module and parameter adaptive pulse coupled neural network; The feature encoding strategy module is based on the measurement noise of the uncorrelated different spatial frequencies or baseline sampling data.
2. The high-resolution radiometer intelligent imaging system according to claim 1, characterized in that: The processing steps of the data preprocessing module are as follows: if the radiometer is an interferometric radiometer, the visibility data obtained by spatial frequency or baseline sampling needs to be linearly mapped into a one-dimensional column vector; if the radiometer is a real aperture scanning radiometer, the average value of the multi-polarization radiation image needs to be calculated, and the spatial frequency component data is extracted from these average images through linear transformation of the partial Fourier matrix, and further linearly mapped into a one-dimensional column vector; the visibility data of the spatial frequency or baseline sampling is expressed as:
3. The high-resolution radiometer intelligent imaging system according to claim 1, characterized in that: The parameter adaptive pulse coupled neural network includes a neighborhood modulation module and a pulse generator module.
4. The high-resolution radiometer intelligent imaging system according to claim 3, characterized in that: The model expression of the modulation module of the neighborhood is: L q =V L ∑ kl W q,kl Y q,kl (n-1) The model expression of the pulse generator module is: The parameter selection expression of the parameter adaptive pulse coupled neural network is:
5. A high-resolution radiometer intelligent imaging method, characterized in that: Including steps: S1: Use the data preprocessing module to preprocess the data collected by the radiometer on the detection scene, and linearly map the preprocessed spatial frequency domain samples into a one-dimensional column vector; S2: Use the feature encoding strategy module to encode the spatial frequency samples, thereby effectively reducing the impact of radiation noise; S3: The encoded spatial frequency samples are aggregated into a feature map through position embedding, and then input into the parameter adaptive pulse coupled neural network for processing; S4: Acquire high-resolution reconstructed images.
6. The high-resolution radiometer intelligent imaging method according to claim 5, characterized in that: In step S2, the spatial frequency sample encoding step includes: S21: Create a spatial frequency transformation matrix, which is expressed as: S22: Weight the spatial frequency samples. The weighted spatial frequency samples are expressed as: S23: Yes Perform a cross-correlation operation, expressed as: S24: To T q Perform summation and square root operations to achieve spatial frequency feature encoding, and its expression is: