Low-altitude photoelectric imaging recognition system with atmosphere turbulence fuzzy compensation
By using dual-light imaging and deep learning models to compensate for atmospheric turbulence, the high cost of adaptive optics systems in near-ground applications has been solved, enabling clear imaging and accurate identification of low-altitude targets, and making it suitable for scenarios such as low-altitude target surveillance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-03-13
AI Technical Summary
Existing adaptive optics systems are costly and bulky for low-altitude near-ground target detection, making them difficult to apply in civilian fields. Furthermore, atmospheric turbulence causes image blurring and jitter, affecting the accuracy of low-altitude target identification.
By employing a dual-light imaging device, a meteorological parameter detection device, a turbulence blur compensation module, and an image correction module, combined with a deep learning model, the effects of atmospheric turbulence are compensated in real time, thereby achieving image clarity and target recognition.
It achieves clear imaging and accurate identification of low-altitude targets in near-ground environments, improving detection performance and making it suitable for scenarios such as low-altitude target surveillance, general aviation management, border patrol, and ecological monitoring.
Smart Images
Figure CN121661602A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of low-altitude detection technology, and in particular relates to a low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation. Background Technology
[0002] Just as cars rely on roads to travel, low-altitude aircraft rely on air routes for navigation. The lower the altitude, the more complex the air routes become. Therefore, a low-altitude air route system should be established, similar to the roads used by cars, forming a network of low-altitude air routes. All aircraft must fly on designated air routes. The photoelectric surveillance network on urban air routes is like the traffic monitoring network at city intersections, acquiring real-time images and target attributes of all aircraft on the air routes. Combined with electromagnetic, communication, and positioning methods, the real-time status of all aircraft on the air routes can be monitored. When necessary, images can be captured and evidence can be collected to achieve accurate detection and verifiable evidence. Only then can the level of low-altitude supervision and enforcement be quantified, and only then can the enforcement actions of regulatory departments be operational.
[0003] In low-altitude airway environments, factors such as atmospheric turbulence and temperature gradient changes can cause severe blurring and jitter in optical imaging, significantly reducing the target recognition distance and accuracy of photoelectric detection equipment. This poses a significant challenge to low-altitude target surveillance (such as general aviation management, border patrol, ecological monitoring, and emergency rescue scenarios). To address this issue, it is necessary to construct a routine low-altitude airway surveillance photoelectric target detection and recognition system, and to build a photoelectric detection and image restoration solution with strong anti-interference capabilities to achieve stable target monitoring and accurate identification under complex atmospheric conditions.
[0004] Typically, within an altitude range of a few meters to tens of kilometers above the ground, turbulence intensity is related to altitude; the higher the altitude, the lower the turbulence intensity, and vice versa. Therefore, turbulence intensity is greater near the ground compared to other altitudes. Imaging systems used for detecting and identifying low-altitude flying targets are generally installed and operate at altitudes of a few meters to tens of meters above the ground, a near-ground operating environment with low altitude and high turbulence intensity. Therefore, it is necessary to address the impact of near-ground turbulence on image quality. Adaptive optics, as a method capable of real-time detection and correction of dynamic aberrations, enables optical systems to possess a certain degree of resistance to dynamic interference. These three elements constitute a closed-loop system to achieve real-time detection, reconstruction, and correction of distorted wavefronts, which is an important method for correcting atmospheric turbulence dynamic aberrations.
[0005] However, traditional adaptive optics systems suffer from the following problems: Existing adaptive optics technologies are primarily used in astronomical observations and laser communications for long-range imaging, effectively compensating for wavefront distortion caused by atmospheric turbulence and improving space imaging and communication performance. However, there is little analysis of the performance of adaptive optics systems for near-ground applications. In other words, there is virtually no practical application or research of adaptive optics systems for detecting distant targets near the ground. This is because traditional adaptive optics are very bulky, heavy, and expensive. Near-ground photoelectric detection devices are typically small and lightweight, making it impossible to incorporate complex adaptive optics systems. Furthermore, a typical adaptive optics system includes a wavefront sensor, a wavefront controller, and a wavefront corrector, resulting in a large and expensive system that is difficult to apply in civilian fields. Summary of the Invention
[0006] In view of the problems existing in current optical near-ground detection equipment, the present invention aims to achieve clear imaging of low-altitude photoelectric detection and identification system under near-ground medium and long distance conditions, and proposes a low-altitude photoelectric imaging and identification system with atmospheric turbulence ambiguity compensation.
[0007] This invention includes:
[0008] A dual-light imaging device is used to simultaneously acquire raw images of low-altitude targets in the visible light band and the long-wave infrared band.
[0009] Atmospheric parameter detection device is used to collect wind speed, temperature and humidity along the detection path in real time and output corresponding multi-dimensional meteorological time series data;
[0010] The turbulence blur compensation module is connected to the atmospheric parameter detection device. It processes the multidimensional meteorological time series data based on a deep learning multidimensional time series model, retrieves the atmospheric refractive index structure constant in real time, calculates the current optical degradation based on the atmospheric refractive index structure constant, and generates image compensation parameters.
[0011] An image correction module, connected to the dual-light imaging device and the turbulence blur compensation module, is used to perform real-time deblurring on the original image according to the image compensation parameters and output the restored image.
[0012] The integrated image recognition module, connected to the image correction module, is used to perform weak target detection and target classification in the restored image, and output the attribute and status information of low-altitude targets.
[0013] Optionally, the dual-light imaging device includes a visible light detector and a long-wave infrared detector, which are installed on the same optical axis or parallel to the optical axis to achieve continuous day and night observation.
[0014] Optionally, the atmospheric parameter detection device is at least one of microwave radar or lidar, and it works in conjunction with a ground meteorological station to synchronously acquire wind speed, temperature, and humidity.
[0015] Optionally, the deep learning multidimensional time series model in the turbulence fuzziness compensation module is one of LSTM, improved LSTM, VMD-LSTM hybrid model or EMD-LSTM hybrid model.
[0016] Optionally, the image correction module employs a deconvolution or Wiener filtering algorithm based on the atmospheric point spread function (PSF) to deblur the original image frame by frame.
[0017] Optionally, the weak target detection algorithm in the integrated image recognition module is a sparse feature extraction algorithm based on the combination of curvature feature points and improved target index.
[0018] Optionally, the integrated image recognition module further includes a trajectory association submodule, used for trajectory tracking and identity preservation of the same target in multiple consecutive restored images.
[0019] Optionally, the turbulence fuzziness compensation module further includes: a data preprocessing submodule, used to perform outlier removal, missing value filling, normalization, and variational mode decomposition (VMD) on multidimensional meteorological time series data to generate stationary subsequences that are input into the deep learning multidimensional time series model.
[0020] Optionally, the deep learning multidimensional time series model uses a sliding time window to construct supervised learning samples, with a window length T = 12 to 24 and a step size of 1. It uses the wind speed, temperature, and humidity sequences of the past T times to predict the atmospheric refractive index structure constant values of the current and future k times, thereby achieving multi-step advance compensation.
[0021] Optionally, the deep learning multidimensional time series model introduces a Dropout layer and an early stopping mechanism, with the Dropout rate set to 0.2 to 0.5, to prevent overfitting of the atmospheric refractive index structure constant prediction and to ensure real-time performance.
[0022] Through the above technical solution, the present invention utilizes an atmospheric parameter detection system, a turbulence fuzziness compensation and image processing module to form a complete low-altitude photoelectric identification system, which can obtain clear images on the detection path. Furthermore, the present invention employs real-time measurement of atmospheric data and model identification technology, which can accurately detect atmospheric turbulence and significantly improve detection performance. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the system structure according to an embodiment of this application;
[0024] Figure 2 This is a flowchart illustrating the atmospheric turbulence structure parameter identification process in an embodiment of this application.
[0025] Figure 3 This is a schematic diagram of the weak target detection algorithm in an embodiment of this application. Detailed Implementation
[0026] The technical solutions in the embodiments of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them.
[0027] Low-altitude sounding involves near-ground, medium-to-long-range imaging, and ground turbulence is a significant factor limiting clear imaging, potentially affecting the accuracy of the sounding system. By accurately detecting core atmospheric elements to deduce turbulence values, image blurring can be compensated for, allowing each image to be clearly restored. This enables weak target detection and precise target identification, significantly reducing the impact of atmospheric turbulence on the imaging system.
[0028] like Figure 1 As shown in the figure, this application discloses a low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation, which includes a dual-light imaging device, an atmospheric parameter detection device, a turbulence ambiguity compensation module, an image correction module, and a comprehensive image recognition module.
[0029] The dual-light imaging device includes a visible light imaging unit 1 and an infrared thermal imaging unit 2. The visible light imaging unit includes a zoom lens and an imaging detector for detecting and identifying low-altitude, slow-moving, and small flying targets during the day. The infrared thermal imaging unit includes a long-wave infrared lens and an uncooled detector for detecting and identifying low-altitude, slow-moving, and small flying targets at night. The two different wavelength imaging detectors are selected for the characteristics of low-altitude targets during the day and at night, respectively, and are used to detect and identify low-altitude, slow-moving, and small targets under different weather conditions.
[0030] This embodiment uses two detectors with different wavelengths—visible light and infrared thermal imaging—to detect and identify low-altitude targets. Firstly, the visible light band offers high resolution and a high frame rate, enabling the detection, identification, and classification of low-altitude, slow-moving, and small targets even in good lighting conditions. At close to medium ranges, it can promptly distinguish the type and attributes of the target. Secondly, long-wave infrared thermal imaging, with its longer wavelength, provides better penetration at night and remains effective even in environments with poor visibility. This allows for the detection and assisted identification of low-altitude targets from a distance.
[0031] The difference between visible light and long-wave infrared wavelengths allows a dual-band system to identify and analyze low-altitude targets under different weather conditions, providing information on the target's position and attributes, such as its trajectory and target type, at different stages of its approach. Therefore, combining the two helps to more accurately determine the actual situation.
[0032] The atmospheric parameter detection device includes a microwave / lidar system for high-resolution real-time detection of low-altitude atmospheric conditions. The microwave / lidar system detects three parameters in high resolution and real-time: low-altitude wind speed, temperature, and humidity. These parameters determine the real-time changes in low-altitude atmospheric turbulence. With the assistance of real-time atmospheric parameters, atmospheric turbulence can be accurately inferred. This solution is suitable for near-ground applications such as low-altitude target detection.
[0033] Furthermore, the atmospheric parameter detection device uses a lidar for precise detection of wind speed and temperature and humidity. The wind lidar 5 adopts a mature continuous wave laser Doppler wind measurement system, which can acquire wind speed and direction at a height of 1000 meters or less vertical to the ground with high temporal and spatial resolution. The atmospheric temperature and humidity lidar 6 adopts the Raman scattering principle. The backscattering intensity signal generated by the beam passing through the atmosphere is received by the detector and then used to retrieve the real-time atmospheric temperature and humidity.
[0034] The turbulence fuzziness compensation module 3 is connected to the atmospheric parameter detection device. It processes the multidimensional meteorological time series data based on a deep learning multidimensional time series model, retrieves the atmospheric refractive index structure constant in real time, calculates the current optical degradation based on the atmospheric refractive index structure constant, and generates image compensation parameters.
[0035] In a preferred example, the compensation module includes an atmospheric refractive index identification module 7, which uses the wind, temperature and humidity data measured in real time by the atmospheric parameter detection device, combined with historical data from the local meteorological station, and employs a support vector machine learning mechanism to identify the atmospheric turbulence model.
[0036] The core function of the Atmospheric Refractive Index Identification Module 7 is to estimate the atmospheric refractive index structure constant. This constant is a crucial parameter describing the vertical structure of the atmosphere and is significant for fields such as imaging target detection, optical communication, navigation systems, and meteorology. Meteorological data such as wind speed, temperature, and humidity directly affect changes in the atmospheric refractive index. Deep learning models can be used to predict the atmospheric refractive index structure constant using these multi-dimensional time-series data. Turbulence model identification will be performed on real-time, refined data of wind, temperature, and humidity in the lower atmosphere.
[0037] The specific principle and process of predicting the atmospheric refractive index structure constant based on deep learning (LSTM of multidimensional time series and improved hybrid model) in this embodiment are as follows:
[0038] Atmospheric refractive index structure constant C n 2It is a core parameter characterizing the intensity of atmospheric optical turbulence. Its spatiotemporal distribution is nonlinearly influenced by multidimensional meteorological factors such as temperature, humidity, wind speed, and air pressure. Traditional empirical models or single machine learning methods struggle to accurately capture this complex temporal dependency. LSTM (Long Short-Term Memory) networks, with their gating mechanism, can effectively address the long-term dependency problem in time-series data. Furthermore, improved hybrid models based on LSTM (incorporating data decomposition, optimization algorithms, and multi-model fusion strategies) can further enhance C... n 2 Prediction accuracy.
[0039] In this embodiment, an improved hybrid model based on LSTM using multidimensional time-series data is used to predict C. n 2 The overall framework can be divided into four core components: data layer, preprocessing layer, model layer, and prediction and validation layer. (See also...) Figure 2 The specific logic is as follows:
[0040] The data layer acquires multi-dimensional meteorological time-series data (temperature / humidity / wind speed / air pressure, etc.).
[0041] The preprocessing layer performs outlier removal, missing value imputation, normalization, and data decomposition.
[0042] In the model layer, the LSTM base model is selected, hyperparameters are optimized, and modules are fused.
[0043] The prediction and validation layer performs single-step / multi-step prediction, error evaluation, and finally generalization validation.
[0044] Final output: C n 2 Predicted values and accuracy analysis.
[0045] Furthermore, the quality of multidimensional time-series data directly determines model performance, specifically for C. n 2 Preprocessing for prediction needs to focus on addressing issues such as data non-stationarity, noise interference, and dimensional differences.
[0046] 1. Basic preprocessing includes:
[0047] Outlier removal: The 3σ criterion and box plot method are used to identify and remove extreme outliers (such as sudden changes in wind speed and temperature jumps) in sensor observation data.
[0048] Missing value imputation: For intermittent missing meteorological data, linear interpolation, spline interpolation, or LSTM interpolation methods are used (using adjacent time series features to predict missing values).
[0049] Normalization / Standardization: Perform Min-Max normalization (mapped to [0,1]) or Z-score standardization on meteorological variables of different dimensions (such as temperature: ℃, air pressure: hPa, wind speed: m / s) to eliminate the influence of dimensional differences on the model gradient descent.
[0050] 2. Data Decomposition
[0051] C n 2 Time series data contains trend terms, periodic terms, and random noise terms, exhibiting strong non-stationarity. Therefore, decomposition techniques are needed to break it down into stationary subsequences to reduce the learning difficulty of the model.
[0052] Optionally, Empirical Mode Decomposition (EMD) / Ensemble Empirical Mode Decomposition (EEMD): This involves transforming the original C... n 2 Meteorological time series data are decomposed into several intrinsic mode functions (IMFs) and residual terms to solve the mode aliasing problem of EMD;
[0053] Optionally, variational mode decomposition (VMD) can be used to split data into a specified number of stationary mode components (IMF) through adaptive decomposition. Compared with EMD, it has stronger mathematical theoretical support and can effectively separate high-frequency noise and low-frequency trends in turbulent data.
[0054] After decomposition, an LSTM sub-model is constructed separately for each stationary IMF component. Finally, the prediction results of each sub-model are merged to improve the overall accuracy.
[0055] Therefore, the implementation method of this embodiment is as follows:
[0056] Input: Construct a multidimensional temporal feature matrix X=[x1, x2, ..., x m ], where x i For the time series of the i-th meteorological variable (e.g., 2-meter temperature, 10-meter wind speed, relative humidity, etc.), a sliding time window method (e.g., window length T=12, step size=1) is used to transform the time series data into supervised learning samples: (X {t-T+1:t} C n 2 (t)), that is, using multidimensional meteorological data from the previous T time points to predict the current time C. n 2 ;
[0057] Output: Single-step prediction is C at the current time. n 2 Value, multi-step prediction of C for the next k times. n 2 sequence
[0058] Network structure: Input layer (dimension = number of meteorological variables × window length), enters LSTM hidden layer (1 to 3 layers, number of hidden units can be set to 64 / 128 / 256), passes through Dropout layer (suppresses overfitting, dropout_rate=0.2 to 0.5), then to fully connected layer, and finally to output layer (dimension = 1, single-step prediction).
[0059] The training strategy used in this embodiment: Loss function: Root Mean Square Error (RMSE) or Mean Absolute Percentage Error (MAPE) (adapted to C) n 2 Numerical range characteristics); Optimizer: Adam optimizer (adaptive learning rate, fast convergence speed), initial learning rate set to 0.001~0.01; Training set / validation set / test set partitioning: partitioned in chronological order (e.g. 7:1:2) to avoid data leakage.
[0060] In a preferred example, the compensation module further includes a compensation parameter generation module 8, which calculates the current optical degradation based on the atmospheric refractive index structure constant and generates image compensation parameters.
[0061] like Figure 1 As shown, the image processing module 4 in this embodiment includes an image correction module 9 and a comprehensive image recognition module 10, wherein:
[0062] The image correction module 9 is connected to the dual-light imaging device and the turbulence blur compensation module, and is used to perform real-time deblurring on the original image according to the image compensation parameters and output the restored image.
[0063] The integrated image recognition module 10 is connected to the image correction module and is used to perform weak target detection, track association and target classification recognition in the restored image, and output the attribute and status information of low-altitude targets.
[0064] Furthermore, the integrated image recognition module 10 includes:
[0065] The weak target detection submodule performs real-time detection of low-speed, small targets in the corrected and restored high-resolution visible light / infrared image, and associates the acquired target points to form a track. When the target is large at medium to close range, the clearly restored image enters the target classification and recognition submodule to classify and identify the target and confirm its type and attributes.
[0066] The trajectory association submodule is used to track the trajectory and maintain the identity of the same target in multiple consecutive frames of restored images.
[0067] The target classification and recognition submodule identifies and tracks targets in real time in the high-resolution visible light / infrared image restored by the correction module, identifies and classifies targets at different distances, and confirms target attributes.
[0068] In a preferred example, the principle of the weak target detection algorithm is as follows: Figure 3 As shown, this algorithm first obtains the curvature feature point map of the image through curvature filtering, and then uses an improved target index algorithm to sparsify the relatively dense feature point regions to obtain the final curvature feature point map of the image. Simultaneously, it extracts the four-directional grayscale curve of each feature point as its descriptor. This algorithm overcomes the problem that traditional feature point detection algorithms struggle to stably extract weak target feature points in complex backgrounds, employing a curvature feature point detection algorithm based on curvature and an improved target index.
[0069] In summary, this application utilizes a simplified near-ground adaptive optics compensation system, which can compensate for the blurring of near-ground mid-to-long-range imaging by accurately measuring atmospheric turbulence, thus greatly improving the performance of low-altitude target detection and identification.
[0070] The foregoing has described one embodiment of the present invention in detail, but the description is merely a preferred embodiment of this application and should not be construed as limiting the scope of this application. All equivalent variations and modifications made within the scope of this application should still fall within the patent coverage of this application.
Claims
1. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation, characterized in that, include: A dual-light imaging device is used to simultaneously acquire raw images of low-altitude targets in the visible light band and the long-wave infrared band. Atmospheric parameter detection device is used to collect wind speed, temperature and humidity along the detection path in real time and output corresponding multi-dimensional meteorological time series data; The turbulence blur compensation module is connected to the atmospheric parameter detection device. It processes the multidimensional meteorological time series data based on a deep learning multidimensional time series model, retrieves the atmospheric refractive index structure constant in real time, calculates the current optical degradation based on the atmospheric refractive index structure constant, and generates image compensation parameters. An image correction module, connected to the dual-light imaging device and the turbulence blur compensation module, is used to perform real-time deblurring on the original image according to the image compensation parameters and output the restored image. The integrated image recognition module, connected to the image correction module, is used to perform weak target detection and target classification in the restored image, and output the attribute and status information of low-altitude targets.
2. The low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 1, characterized in that, The dual-light imaging device includes a visible light detector and a long-wave infrared detector, which are installed on the same optical axis or parallel to the optical axis to achieve continuous day and night observation.
3. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 1 or 2, characterized in that, The atmospheric parameter detection device is at least one of microwave radar or lidar, and it works in conjunction with a ground meteorological station to synchronously acquire wind speed, temperature, and humidity.
4. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 1, characterized in that, The deep learning multidimensional temporal model in the turbulence fuzziness compensation module is one of LSTM, improved LSTM, VMD-LSTM hybrid model or EMD-LSTM hybrid model.
5. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 1 or 4, characterized in that, The image correction module uses a deconvolution or Wiener filtering algorithm based on the atmospheric point spread function (PSF) to deblur the original image frame by frame.
6. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 5, characterized in that, The weak target detection algorithm in the integrated image recognition module is a sparse feature extraction algorithm based on the combination of curvature feature points and improved target index.
7. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 6, characterized in that, The integrated image recognition module also includes a trajectory association submodule, which is used to track the trajectory and maintain the identity of the same target in multiple consecutive frames of restored images.
8. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 5, characterized in that, The turbulence fuzziness compensation module further includes a data preprocessing submodule, used to perform outlier removal, missing value filling, normalization, and variational mode decomposition on multidimensional meteorological time series data to generate stationary subsequences that are input into the deep learning multidimensional time series model.
9. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 1, characterized in that, The deep learning multidimensional time series model uses a sliding time window to construct supervised learning samples. The window length is T = 12 to 24 and the step size is 1. It uses the wind speed, temperature and humidity sequence of the past T times to predict the atmospheric refractive index structure constant value of the current and future k times, so as to achieve multi-step advance compensation.
10. A low-altitude photoelectric imaging and recognition system with atmospheric turbulence ambiguity compensation according to claim 9, characterized in that, The deep learning multidimensional time series model introduces a Dropout layer and an early stopping mechanism, with the Dropout rate set to 0.2–0.5 to prevent overfitting of the atmospheric refractive index structure constant prediction and to ensure real-time performance.