Sea surface small target detection method based on environment varying dynamic feature selection

Abstract

The invention discloses a sea surface small target detection method based on environment varying dynamic feature selection. The method comprises the following steps: firstly, obtaining a radar echo signal; then, a simulation target echo is constructed; feature extraction is carried out; feature selection based on an ant colony algorithm; and finally, a controllable false alarm improved K-nearest neighbor method is adopted for judgment. Compared with the prior art, the detection probability higher than 80% is still kept in the low SCR, and a brand new solution is provided for engineering detection of weak and small sea surface targets.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, specifically relating to a method for detecting small sea surface targets by selecting dynamic features that vary with environmental changes. Background Technology

[0002] The detection of small targets on the sea surface against sea clutter backgrounds remains a long-standing core problem in radar target detection. The technical challenges stem primarily from three aspects: the non-stationary nature of high-resolution sea clutter leading to mismatches in traditional statistical models; weak target echoes, slow movement speeds, and small radar cross-sections; and the risk of confusion between wave-breaking sea clutter spikes and target Doppler characteristics. Technological evolution in small target detection against sea clutter backgrounds has mainly revolved around two paths: clutter suppression and feature space optimization.

[0003] Traditional feature detection methods mainly rely on single-domain feature extraction, i.e., independent analysis based on time-domain power statistics or frequency-domain Doppler shift characteristics. However, in small target detection scenarios against high-resolution sea clutter backgrounds, such single-dimensional feature detectors face significant performance degradation, with their detection probability exhibiting a precipitous drop under low signal-to-clutter ratio (SCR) conditions.

[0004] Multi-feature fusion strategies have become a research hotspot in recent years. A joint detection framework combining time-domain and frequency-domain features has been proposed, overcoming the limitations of single-domain approaches. By using a convex hull decision domain, it achieves a detection probability improvement of over 10% on the IPIX dataset, but its detection performance depends on signal retrieval rate (SCR). To detect targets with low SCR, researchers have proposed time-frequency detection methods. This method, by selecting an appropriate time-frequency distribution function, can effectively reveal the distribution characteristics of signal energy in the joint time-frequency domain, extracting discriminative features that are difficult to observe in single-domain analysis, thus achieving effective fusion of time-frequency domain information.

[0005] However, while multi-feature fusion strategies have the potential to improve detection performance, they also introduce new challenges. On the one hand, a large number of extracted time-domain, frequency-domain, and time-frequency-domain features inevitably exhibit high correlation or redundant information. A large number of redundant features not only fail to improve classification and discrimination capabilities but may also dilute the role of the few truly crucial strong features. On the other hand, high-dimensional feature spaces face the "curse of dimensionality," meaning that as the number of features increases, the required training sample size grows exponentially to ensure the model's generalization ability.

[0006] Therefore, to address the feature redundancy problem, there is an urgent need to propose a small sea surface target detector that can reduce feature dimensions while maintaining the physical interpretability of features and adaptability to dynamic environments. The core difference between this invention and existing small sea surface target detection patents lies in the dynamic feature selection mechanism, rather than simple feature fusion or fixed model optimization.

[0007] Traditional single-domain feature methods (such as fractal detectors) rely on single-dimensional features in the time or frequency domains, which can lead to a sharp drop in detection probability due to model mismatch in high clutter environments. This invention, however, avoids information loss by using multi-domain feature fusion (eight-dimensional features in the time, frequency, and time-frequency domains) combined with Ant Colony Optimization (ACO) for dynamic filtering.

[0008] Fixed multi-feature fusion patents (such as time-frequency three-feature detectors): Although they introduce multi-domain features, they use fixed combinations, failing to solve the feature redundancy problem. The curse of dimensionality leads to an exponential increase in the demand for training samples. This invention uses ACO to intelligently remove redundancy, reducing the feature dimension to 3-5 dimensions and reducing computation time to the millisecond level.

[0009] Machine learning-driven patents (such as Support Vector Machine (SVM) detectors) rely on large amounts of labeled data, but samples of small targets on the sea surface are scarce. This invention innovatively uses simulated target echoes to construct a training set, avoiding reliance on real data, and improves the model's generalization ability through ACO (Average Coherence to Optimization). Summary of the Invention

[0010] To overcome the shortcomings of existing technologies, this invention provides a method for detecting small sea surface targets using dynamic feature selection based on environmental changes. First, radar echo signals are acquired; then, simulated target echoes are constructed; next, feature extraction is performed; feature selection is based on an ant colony algorithm; finally, an improved K-nearest neighbor method with controllable false alarms is used for determination. Compared with existing technologies, this invention maintains a detection probability above 80% even with low SCR (Specialized Recognition Rate), providing a novel solution for the engineering detection of weak sea surface targets.

[0011] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Acquire radar echo signals; Step 2: Construct simulated target echo; Step 3: Feature extraction; Radar Receive Sequence x ( n Sea clutter time series were obtained through parallel branches. c ( n ) and the target echo generator s ( n After the two are merged, eight-dimensional feature extraction is performed simultaneously; Step 4: Feature selection based on ant colony algorithm; Step 5: Use the improved K-nearest neighbor method with controllable false alarms for determination.

[0012] Preferably, step 2 specifically comprises: The echo of a small target with uniform velocity and uniform acceleration is represented as: (1) in, P c Indicates the average power of clutter. a ( n () is a highly correlated non-negative random sequence used to characterize the amplitude fluctuation characteristics of the target; For radar operating wavelength, and The initial and final radial velocities of the target are not indicated by the symbol. The pulse repetition period, Assuming an initial phase that follows a uniform distribution in the range [0, 2π], N The simulated signal length; using SCR empirical parameters Set the normalized range [10 dB, 20 dB] to correspond to the interval [-10 dB, 20 dB]. -1 10 1 / 2 The uniform distribution of [ ] is used to adjust the SCR value of the simulated target; Preferably, step 3 specifically comprises: Radar Receive Sequence x ( n Sea clutter time series were obtained through parallel branches. c ( n ) and the target echo generator s ( n After the two are merged, eight-dimensional feature extraction is performed simultaneously; Step 3-1: Normalized Hurst Exponent (NHE). The amplitude time series of sea clutter exhibits multifractal characteristics in the time range of tens of milliseconds to several seconds. At the corresponding time scale, the following power-law relationship approximately holds: (2) in, It is the echo sequence received by the radar. H 2(x) represents the Hurst exponent of the sea clutter time series; It is a structure function, obtained by averaging the magnitudes of all intervals m in the sequence; Considering the spatial variation characteristics of the Hurst index of sea clutter, the differential characteristic NHE is: (3) in and They represent P The mean and standard deviation of the Hurst exponent for each reference unit; Step 3-2: Relative Average Amplitude (RAA); Length is N The average amplitude of the signal is defined as follows: (4) Where x represents the echo time series. and These represent the average echo of the cell to be estimated and the average echo of the surrounding reference cells, respectively. The relative average amplitude is calculated as follows: (5) Step 3-3: Relative Doppler Peak Height (RDPH); Differences between sea clutter and target-carrying echoes were observed from the Doppler amplitude spectrum (DAS) of the complex time series. ，x ( n The Doppler amplitude spectrum is calculated as follows: (6) in Doppler frequency, T r =0.001 is the pulse repetition frequency of the IPIX radar; The Doppler peak height (DPH) of DAS is calculated as follows: (7) in, Indicates the height of Doppler Peak. For reference, the range of Doppler elements, To maximize the Doppler bandwidth that the target may occupy; Represents the set of all Doppler reference elements. , The union of two sets. represent The number of all elements in the set; Using the DPH of the element to be estimated and the DPH of the surrounding reference elements, the RDPH is calculated as follows: (9) Where x p The echo representing the pure clutter reference cell, Indicates relative height to the Doppler peak; Steps 3-4: Relative vector entropy; The Doppler spectral vector entropy is expressed as: (10) in To normalize the Doppler amplitude spectrum; using the reference distance cell, the relative Doppler entropy RVE is defined as: (11) Steps 3-5: Ridge Integration (RI), Number of Connected Regions (NR), and Maximum Size of Connected Regions (MS). The time-frequency joint analysis of the time series of the cell to be detected and its neighboring reference cells is performed, and its normalized time-frequency distribution is defined as: (12) in, It is a smoothed pseudo-Wigner-Wylie distribution of the time series x; and They represent The mean and standard deviation; The characteristic RI is calculated as follows: (13) Steps 3-6: Generalized Likelihood Ratio Statistic (GLRT); An observation time interval is divided into K disjoint sub-intervals. The sea clutter time series is modeled as a piecewise spherical invariant random vector (PW-SIRV) with an inverse gamma distribution of power. The target echo is modeled as a sequence of unknown Doppler frequencies. The parameterized signal has a piecewise linear frequency modulated (PW-LFM) configuration, with a constant signal amplitude within each segment; the GLRT test statistic is calculated as follows: (14) in, η and λ These represent the scale parameter and shape parameter of the texture distribution, respectively; Δ t and Δ T These represent the pulse repetition interval of the radar and the duration of each sub-interval, respectively. Indicates the first k Normalized sample covariance matrix of Shanghai clutter in each sub-interval; Doppler frequency sequence in the PW-LFM model of the target echo. It can be estimated using the forward-backward income reference particle filter (FB-IRPF).

[0013] Preferably, step 4 specifically comprises: Step 4-1: Data preprocessing; The raw radar echo data is first standardized to eliminate the influence of differences in characteristic dimensions on the optimization process, using the Z-score standardization formula: (15) in x ij Indicates the first i The first sample j Original values ​​of each feature. and Corresponding to the first j The mean and standard deviation of the dimensional features. D =8 represents the total dimension of the features; after standardization z ij The characteristic space that constitutes a uniform dimension; Step 4-2: Construct feature prior knowledge based on Fisher scores; Fisher scores evaluate features by maximizing inter-class variance and minimizing intra-class variance, and their discriminant ratio is: (16) in , They represent the first i Features in Category k , k The mean and variance of =0,1; after normalization... Used as heuristic information to guide the search process; Step 4-3: Ant colony iterative optimization; 30 ants search the feature space in parallel. k Select features j The probability is determined by the pheromone concentration. With heuristic information Joint decision: (17) in α , β Adjust the pheromone and heuristic weights separately. δ To prevent smoothing terms with a denominator of zero, a dynamic threshold mechanism is introduced. If no feature is selected, a random feature is forcibly activated. The fitness function balances classification performance with the number of features. (18) Where S is the feature subset, and Accuracy is the accuracy of the K-NN classifier on the validation set. This is the sparsity penalty coefficient; Step 4-4: Elite Strategy Update: Pheromone on the globally optimal path is volatilized and enhanced. (19) in S is the volatility coefficient. ∗ The current optimal subset is used; after reaching the maximum number of iterations, the optimal subset S is output. ∗ ; Indicates the first t At +1 iteration, the features j pheromone concentration, Indicates the first t In the next iteration, the features j pheromone concentration, This represents the change in pheromone level of feature j during this iteration.

[0014] Preferably, step 5 specifically comprises: A simulated target echo generator was used to generate target samples of equal quantity to sea clutter data. Eight-dimensional features of both sea clutter and target samples were jointly extracted to construct a balanced training set. The Euclidean distance between the test sample and all training samples was calculated, and the closest sample was selected. k One neighbor; statistics k Number of clutter samples in the nearest neighbors n 0, introduce a decision threshold w :like n 0≥ w If it is determined to be clutter, it is determined to be a target; a fixed decision threshold is used. By dynamically adjusting the number of nearest neighbors k The value ensures a controllable false alarm rate; the false alarm probability is preset. P f Next, iterative optimization k Value makes the actual false alarm rate P fa satisfy , To find the minimum value, the optimal value is finally determined. During online testing, the parameters of each test sample are calculated. The number of clutter samples in the nearest neighbors, based on a threshold. A binary judgment was made.

[0015] Preferably, in step 1, the international benchmark dataset IPIX dataset is used as the measured radar echo signal.

[0016] Preferably, the =50Hz =5Hz.

[0017] Preferably, in steps 3-5, the feature extraction process for the number of connected regions NR and the maximum connected region size MS is as follows: First, dynamic thresholding is performed on NTFD, and the previous values ​​are retained in each time unit. L The largest pixel value is assigned a value of 1, and the remaining pixels are set to zero to generate a thresholded NTFD image after binarization. Then, region growing is performed based on the eight-neighbor connectivity criterion, the number of all connected regions is counted as the NR feature value, and the total number of pixels contained in the largest connected region is recorded as the MS feature value.

[0018] Preferably, the maximum number of iterations in step 4 T =50.

[0019] The beneficial effects of this invention are as follows: Based on the IPIX radar dataset, this study presents an experiment on small target detection on the sea surface. Eight-dimensional features were selected using an ant colony algorithm, and the detection performance of four polarization modes was systematically analyzed using a K-NN algorithm with controllable false alarms. The observation time was 0.512 seconds, and the false alarm probability was 0.001. Figure 3 It is a line graph showing the detection probability of five detectors under four polarization modes.

[0020] This invention achieves a breakthrough in the field of small target detection against complex sea clutter backgrounds. Systematic experiments based on the IPIX radar dataset demonstrate that, through a collaborative mechanism of ant colony optimization feature selection and a controllable false alarm K-NN classifier, the system achieves an average detection probability of 81% across four polarization modes, significantly outperforming traditional single-domain feature detectors. This advantage stems from two core technologies: First, the constructed time-domain-frequency-time-frequency domain eight-dimensional feature fusion framework systematically captures target discrimination information across multiple physical domains, maintaining over 70% detection capability, especially under extreme conditions of low SCR, through complementary feature representation. Second, the innovative ant colony optimization algorithm constructs feature prior knowledge based on Fisher scores and adaptively selects the most discriminative feature subset through a dynamic pheromone update mechanism, effectively solving the "curse of dimensionality" problem caused by traditional multi-feature fusion.

[0021] Experimental data further validated the system's practical engineering value. For example... Figure 4As shown, the ant colony algorithm intelligently reduces the original 8-dimensional feature set to 3-5 dimensions (average feature count 4.2 dimensions), reducing the computational load of feature extraction and classification by more than 40% and shortening the single detection time to the millisecond level. Crucially, although the number of features fluctuates dynamically, the detection probability remains stable within the 78%-85% range, demonstrating that the system can autonomously optimize feature combinations based on sea state changes: for example, automatically strengthening the weight of time-frequency domain features in strong sea spike interference scenarios, while emphasizing frequency domain features in stable sea states. This adaptive capability, combined with a K-NN classifier with controllable false alarm rate, enables the system to maintain a high detection probability while meeting the stringent real-time and reliability requirements of shipborne radar. Compared with existing technologies, this invention maintains a detection probability above 80% even with low SCR, providing a novel solution for the engineering detection of small targets on the sea surface. Attached Figure Description

[0022] Figure 1 This is a flowchart of feature selection based on the ant colony algorithm; Figure 2 This is a flowchart of a detector based on ant colony optimization for explicit feature selection; Figure 3 This is a diagram showing the performance comparison of the five detectors on IPIX data; Figure 4 It is the dimension of the feature space after feature selection. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] This invention aims to propose a small target detection system for the sea surface based on feature selection using an ant colony optimization strategy, in order to solve the following technical problems: (1) Due to the diverse motion characteristics of small targets on the sea surface and the complexity of their dynamic interaction with the sea surface, it is significantly difficult to obtain measured data covering all categories of targets. To address the lack of real target samples, this invention constructs a joint training set by simulating target echoes and measured sea clutter, thus avoiding reliance on real target samples; (2) In view of the problem that a single feature is difficult to effectively capture the information of small targets on the sea surface that are different from the background clutter in the context of strong sea clutter, the present invention constructs a multi-domain feature fusion framework, which systematically extracts representative eight-dimensional features from multiple levels and angles such as time domain, frequency domain and time-frequency domain, and enhances the distinguishability of weak target signals in extremely low SCR environment through complementary information representation.

[0025] (3) In view of the problem of feature information redundancy and high correlation caused by multi-feature fusion strategy in existing small target detection technology on the sea surface, this invention effectively eliminates redundant features through an innovative feature selection mechanism, significantly reduces computational complexity, optimizes feature combination, and thus improves the distinguishability of different categories (targets and clutter) data in the feature space.

[0026] (4) To address the stringent requirements for false alarm rate in the detection of small targets on the sea surface, this invention employs a specially improved controllable false alarm K-NN detection method. This classifier achieves accurate, reliable, and adaptive control of false alarms by explicitly setting the false alarm rate during the training phase and dynamically adjusting key parameters in the model, thereby meeting the core requirements of safety and reliability for practical systems.

[0027] Example: Step 1: Acquire radar echo signals; This invention uses the internationally authoritative benchmark dataset IPIX (Intelligent Pixel Processing X-band) as the measured radar echo data.

[0028] The IPIX radar dataset is an authoritative benchmark for sea clutter and target detection research, led by Professor Haykin's team in Canada. This data was acquired using an X-band fully polarimetric radar system on a Canadian coastal platform (30 meters above sea level) at low elevation angles, fully recording echo signals in four polarization modes (HH, HV, VH, VV). The dataset was acquired in two phases: in 1993, 10 datasets were obtained using low RCS floating bodies as targets in the Dartmouth Sea; and in 1998, another 10 datasets were obtained using anchored vessels as targets in the Grimsby Sea. Each dataset contains multiple range cells, where the main cell focuses on the target signal, the 2-3 neighboring influencing cells are affected by target scattering coupling effects, and the remaining reference cells provide clean sea clutter background data. Through this structured design, IPIX provides a crucial validation foundation for small target detection algorithms under complex sea conditions.

[0029] Step 2: Construct simulated target echo; Due to the diverse motion characteristics of small targets on the sea surface and the complexity of their dynamic interactions with the sea surface, obtaining measured data covering all categories of targets presents significant challenges. Therefore, this invention proposes a typical target echo simulation method based on prior clutter information.

[0030] The echo of a typical small target exhibiting uniform velocity and uniform acceleration can be represented as: (19) in, P c Indicates the average power of clutter.a ( n ( ) is a highly correlated non-negative random sequence used to characterize the amplitude fluctuation characteristics of the target. For radar operating wavelength, v 0 and v 1 represents the initial and final radial velocities of the target, respectively. The pulse repetition period, Assuming an initial phase that follows a uniform distribution in the range [0, 2π], N The simulated signal length. The SCR empirical parameters... Set the normalized range [10 dB, 20 dB] to correspond to the interval [-10 dB, 20 dB]. -1 10 1 / 2 The uniform distribution of [ ] is used to adjust the SCR value of the simulation target.

[0031] Step 3: Feature extraction; Radar Receive Sequence x ( n Sea clutter time series were obtained through parallel branches. c ( n ) and the target echo generator s ( n After the two are merged, eight-dimensional feature extraction is performed simultaneously; Step 3-1: Normalized Hurst Exponent (NHE). The amplitude time series of sea clutter exhibits multifractal characteristics in the time range of tens of milliseconds to several seconds. At the corresponding time scale, the following power-law relationship approximately holds: (20) in, It is the echo sequence received by the radar. H 2(x) represents the Hurst exponent of the sea clutter time series; It is a structure function, which is obtained by averaging the magnitudes of all intervals m in the sequence.

[0032] Considering the spatial variation characteristics of the Hurst index of sea clutter, the differential characteristic NHE is: (twenty one) in and They represent P The mean and standard deviation of the Hurst exponent for each reference unit; Step 3-2: Relative Average Amplitude (RAA); Length isN The average amplitude of the signal is defined as follows: (twenty two) Where x represents the echo time series. and These represent the average echo of the cell to be estimated and the average echo of the surrounding reference cells, respectively. The relative average amplitude is calculated as follows: (twenty three) Step 3-3: Relative Doppler Peak Height (RDPH); The difference between sea clutter and echoes with targets can be observed from the Doppler amplitude spectrum (DAS) of the complex time series. ，x ( n The Doppler amplitude spectrum is calculated as follows: (twenty four) in Doppler frequency, T r =0.001 is the pulse repetition frequency of the IPIX radar; The Doppler peak height (DPH) of DAS can be calculated as follows: (25) in, Indicates the height of Doppler Peak. For reference, the range of Doppler elements, To maximize the Doppler bandwidth that the target may occupy; Represents the set of all Doppler reference elements. , The union of two sets. represent The number of all elements in the set; Using the DPH of the element to be estimated and the DPH of the surrounding reference elements, the RDPH is calculated as follows: (26) Where x p The echo representing the pure clutter reference cell, This indicates the relative height of the Doppler peak. Steps 3-4: Relative Vector Entropy (RVE); The Doppler spectral vector entropy is expressed as: (27) in To normalize the Doppler amplitude spectrum; using the reference distance cell, the relative Doppler entropy RVE is defined as: (28) Steps 3-5: Ridge Integration (RI), Number of Connected Regions (NR), and Maximum Size of Connected Regions (MS). The time-frequency joint analysis of the time series of the cell to be detected and its neighboring reference cells is performed, and its normalized time-frequency distribution is defined as: (29) in, It is a smoothed pseudo-Wigner-Wylie distribution of the time series x; and They represent The mean and standard deviation; The characteristic RI is calculated as follows: (30) Steps 3-6: Generalized Likelihood Ratio Statistic (GLRT): A long observation time interval can be divided into K disjoint sub-intervals. Therefore, the sea clutter time series can be modeled as a piecewise spherical invariant random vector (PW-SIRV) with an inverse gamma distribution of power, and the target echo can be modeled as a sequence of unknown Doppler frequencies. A parameterized piecewise linear frequency modulated (PW-LFM) signal has a constant amplitude within each segment. The GLRT test statistic can be calculated as follows: (31) in, η and λ These represent the scale parameter and shape parameter of the texture distribution, respectively; Δ t and Δ T These represent the pulse repetition interval of the radar and the duration of each sub-interval, respectively. Indicates the first k Normalized sample covariance matrix of Shanghai clutter in each sub-interval; Doppler frequency sequence in the PW-LFM model of the target echo. It can be estimated using the forward-backward income reference particle filter (FB-IRPF).

[0033] Step 4: Feature selection based on ant colony algorithm; Step 4-1: Data preprocessing; The raw radar echo data is first standardized to eliminate the influence of differences in characteristic dimensions on the optimization process, using the Z-score standardization formula: (32) in x ij Indicates the first i The first sample j Original values ​​of each feature. and Corresponding to the first j The mean and standard deviation of the dimensional features. D =8 represents the total dimension of the features; after standardization z ij The characteristic space that constitutes a uniform dimension; Step 4-2: Construct feature prior knowledge based on Fisher scores; Fisher scores evaluate features by maximizing inter-class variance and minimizing intra-class variance, and their discriminant ratio is: (33) in , They represent the first i Features in Category k , k The mean and variance of =0,1; after normalization... Used as heuristic information to guide the search process; Step 4-3: Ant colony iterative optimization; 30 ants search the feature space in parallel. k Select features j The probability is determined by the pheromone concentration. With heuristic information Joint decision: (34) in α , β Adjust the pheromone and heuristic weights separately. δ To prevent smoothing terms with a denominator of zero, a dynamic threshold mechanism is introduced. If no feature is selected, a random feature is forcibly activated. The fitness function balances classification performance with the number of features. (35) Where S is the feature subset, and Accuracy is the accuracy of the K-NN classifier on the validation set. This is the sparsity penalty coefficient; Step 4-4: Elite Strategy Update: Pheromone on the globally optimal path is volatilized and enhanced. (36) in S is the volatility coefficient. ∗ The current optimal subset is used; after reaching the maximum number of iterations, the optimal subset S is output. ∗ ; Indicates the first t At +1 iteration, the features j pheromone concentration, Indicates the first t In the next iteration, the features j pheromone concentration, This represents the change in pheromone level of feature j during this iteration; Figure 1 This is a flowchart of feature selection based on the ant colony algorithm. Blue represents the start and end points, green represents the path construction of a single ant, yellow represents the construction of all ants in the same generation, and red represents iterative updates and optimizations.

[0034] Figure 2 This is a flowchart of a detector based on ant colony optimization and explicit feature selection. First, the radar receives the sequence. x ( n Sea clutter time series were obtained through parallel branches. c ( n ) and the target echo generator s ( n The two are then merged and seven-dimensional feature extraction is performed simultaneously. Subsequently, multi-objective feature selection is performed based on the ant colony optimization algorithm to select the feature subset with the greatest inter-class separability, and a new feature space after dimensionality reduction is constructed. This space data is divided into null hypotheses. H Training set corresponding to 0 (pure sea clutter scenario) T 0 and alternative assumptions H 1 (including target signal) corresponding to T 1. Finally, the input is a controllable false alarm K-nearest neighbor classifier. By dynamically adjusting the neighborhood radius and sample weights, the classifier can achieve the desired false alarm probability. P fa Achieving adaptive target detection and decision-making under constraints.

[0035] Step 5: Improved K-Nearest Neighbor Method for Controllable False Alarms; A simulated target echo generator was used to generate a typical target sample with an equal amount of sea clutter data. Eight-dimensional features of both the sea clutter and target samples were jointly extracted to construct a balanced training set. The Euclidean distance between the test sample and all training samples was calculated, and the closest sample was selected. k A neighboring country. Statistics k Number of clutter samples in the nearest neighbors n 0, introduce a decision threshold w :like n 0≥ w If it is not, it is considered clutter; otherwise, it is considered a target. A decision threshold is fixed based on experimental experience. The false alarm rate is controllable by dynamically adjusting the nearest neighbor number k. A preset false alarm probability is set. P f Next, iterative optimization k Value makes the actual false alarm rate P fa satisfy , To find the minimum value, the optimal value is finally determined. During online testing, its... The number of clutter samples in the nearest neighbors, based on a threshold. A binary judgment was made.

[0036] 1. Simulation results; (1) A comparison of the detection performance of the detector proposed in this invention with that of traditional three-feature detectors, time-frequency three-feature detectors, fractal detectors, controllable false alarm K-NN detectors, and SVM detectors on the IPIX radar database disclosed in the art, such as Figure 3 As shown.

[0037] (2) The number of dimensions of the feature space after feature selection based on the ant colony strategy proposed in this invention is as follows: Figure 4 As shown.

[0038] 2. Simulation results analysis; Based on the IPIX radar dataset, this study presents an experiment on small target detection on the sea surface. Eight-dimensional features were selected using an ant colony algorithm, and the detection performance of four polarization modes was systematically analyzed using a K-NN algorithm with controllable false alarms. The observation time was 0.512 seconds, and the false alarm probability was 0.001. Figure 3 It is a line graph showing the detection probability of five detectors under four polarization modes.

[0039] Experimental results show that, through a dynamic pheromone update mechanism and Fisher score guidance, the ant colony algorithm selects the feature subset that contributes most to inter-class separability. Combined with a K-nearest neighbor classifier constrained by false alarm rate, an average detection probability of 81% is achieved. The distribution of the number of selected features is as follows: Figure 4As shown, the ant colony algorithm intelligently reduces the original 8-dimensional feature set to 3-5 dimensions (average feature count 4.2 dimensions), reducing the computational load of feature extraction and classification by more than 40%. While the number of features fluctuates dynamically, the corresponding detection probabilities remain highly stable, indicating that the algorithm can intelligently optimize and balance feature dimensionality and information effectiveness. Experiments demonstrate that the feature selection process can adaptively eliminate redundant dimensions and strengthen the weight distribution of target-sensitive features, thereby maintaining detection stability under complex sea conditions.

[0040] The core innovation of this invention lies in introducing an ant colony optimization (ACO) strategy for feature selection, enabling dynamic and adaptive feature subset filtering, which significantly improves detection efficiency and robustness. Specific innovations include: ① Dynamic feature optimization mechanism Traditional methods (such as fixed multi-feature fusion) rely on predefined feature combinations, leading to redundant features diluting key discriminative information. This invention employs the Ant Colony Optimization (ACO) algorithm, simulating biological swarm intelligence, to search the feature space in parallel. Heuristic guidance: Based on Fisher scores, feature prior knowledge is constructed, the inter-class separability of features is quantified, and after normalization, it is used as heuristic information to guide the search, ensuring physical interpretability.

[0041] Probability-driven selection: 30 ants iterate in parallel, and the probability of each ant selecting a feature is determined by both pheromone concentration and heuristic information.

[0042] Adaptive Simplification: The fitness function balances classification accuracy and the number of features, updates pheromones through an elite strategy, and outputs the optimal feature subset.

[0043] ② Environmental adaptability: Ant colony optimization does not rely on fixed rules, but dynamically adjusts the combination of features according to sea conditions.

[0044] ③ Co-optimization with classifiers: Deep integration of feature selection and controllable false alarm K-NN classifier. The dimensionality-reduced feature subset is input into the K-NN classifier, and the false alarm rate is controllable by dynamically adjusting parameters, which meets the stringent safety requirements of shipborne radar.

Claims

1. A method for detecting small sea surface targets by selecting dynamic features that vary with environmental changes, characterized in that, Includes the following steps: Step 1: Acquire radar echo signals; Step 2: Construct simulated target echo; Step 3: Feature extraction; Radar Receive Sequence x ( n Sea clutter time series were obtained through parallel branches. c ( n ) and the target echo generator s ( n After the two are merged, eight-dimensional feature extraction is performed simultaneously; Step 4: Feature selection based on ant colony algorithm; Step 5: Use the improved K-nearest neighbor method with controllable false alarms for determination.

2. The method for detecting small sea surface targets by selecting dynamic features of environmental changes according to claim 1, characterized in that, Step 2 specifically involves: The echo of a small target with uniform velocity and uniform acceleration is represented as: (1) in, P c Indicates the average power of clutter. a ( n () is a highly correlated non-negative random sequence used to characterize the amplitude fluctuation characteristics of the target; For radar operating wavelength, and The initial and final radial velocities of the target are not indicated by the symbol. The pulse repetition period, Assuming an initial phase that follows a uniform distribution in the range [0, 2π], N The simulated signal length; using SCR empirical parameters Set the normalization range [10 dB, 20 dB] to correspond to the interval [-10 dB, 20 dB]. -1 10 1 / 2 The uniform distribution of [ ] is used to adjust the SCR value of the simulation target.

3. The method for detecting small sea surface targets by selecting dynamic features of environmental changes according to claim 2, characterized in that, Step 3 specifically involves: Radar Receive Sequence x ( n Sea clutter time series were obtained through parallel branches. c ( n ) and the target echo generator s ( n After the two are merged, eight-dimensional feature extraction is performed simultaneously; Step 3-1: Normalize the Hurst exponent; The amplitude time series of sea clutter exhibits multifractal characteristics in the time range of tens of milliseconds to several seconds. At the corresponding time scale, the following power-law relationship approximately holds: (2) in, It is the echo sequence received by the radar. H 2(x) represents the Hurst exponent of the sea clutter time series; It is a structure function, obtained by averaging the magnitudes of all intervals m in the sequence; Considering the spatial variation characteristics of the Hurst index of sea clutter, the differential characteristic NHE is: (3) in and They represent P The mean and standard deviation of the Hurst exponent for each reference unit; Step 3-2: Relative average amplitude; Length is N The average amplitude of the signal is defined as follows: (4) Where x represents the echo time series. and These represent the average echo of the cell to be estimated and the average echo of the surrounding reference cells, respectively. The relative average amplitude is calculated as follows: (5) Step 3-3: Relative Doppler peak height; The differences between sea clutter and echoes with targets were observed in the Doppler amplitude spectrum of the complex time series. ，x ( n The Doppler amplitude spectrum is calculated as follows: (6) in Doppler frequency, T r =0.001 is the pulse repetition frequency of the IPIX radar; The Doppler peak height of DAS is calculated as follows: (7) in, Indicates the height of Doppler Peak. For reference, the range of Doppler elements, To maximize the Doppler bandwidth that the target may occupy; Represents the set of all Doppler reference elements. , The union of two sets. represent The number of all elements in the set; Using the DPH of the element to be estimated and the DPH of the surrounding reference elements, the RDPH is calculated as follows: (9) Where x p The echo representing the pure clutter reference cell, Indicates relative height to the Doppler peak; Steps 3-4: Relative vector entropy; The Doppler spectral vector entropy is expressed as: (10) in To normalize the Doppler amplitude spectrum; using the reference distance cell, the relative Doppler entropy RVE is defined as: (11) Steps 3-5: Ridge accumulation, number of connected regions, and maximum connected region size; The time-frequency joint analysis of the time series of the cell to be detected and its neighboring reference cells is performed, and its normalized time-frequency distribution is defined as: (12) in, It is a smoothed pseudo-Wigner-Wylie distribution of the time series x; and They represent The mean and standard deviation; The characteristic RI is calculated as follows: (13) Steps 3-6: Generalized Likelihood Ratio Statistic (GLRT); An observation time interval is divided into K disjoint sub-intervals. The sea clutter time series is modeled as a piecewise spherically invariant random vector PW-SIRV with an inverse gamma distribution of power. The target echo is modeled as a sequence of unknown Doppler frequencies. The parameterized piecewise linear frequency modulated (PW-LFM) signal has a constant amplitude within each segment; the GLRT test statistic is calculated as follows: (14) in, η and λ These represent the scale parameter and shape parameter of the texture distribution, respectively; Δ t and Δ T These represent the pulse repetition interval of the radar and the duration of each sub-interval, respectively. Indicates the first k Normalized sample covariance matrix of Shanghai clutter in each sub-interval; Doppler frequency sequence in the PW-LFM model of the target echo. It can be estimated using a forward and backward gain reference particle filter.

4. The method for detecting small sea surface targets by selecting dynamic features of environmental variation according to claim 3, characterized in that, Step 4 specifically involves: Step 4-1: Data preprocessing; The raw radar echo data is first standardized to eliminate the influence of differences in characteristic dimensions on the optimization process, using the Z-score standardization formula: (15) in x ij Indicates the first i The first sample j Original values ​​of each feature. and Corresponding to the first j The mean and standard deviation of the dimensional features. D =8 represents the total dimension of the features; after standardization z ij The characteristic space that constitutes a uniform dimension; Step 4-2: Construct feature prior knowledge based on Fisher scores; Fisher scores evaluate features by maximizing inter-class variance and minimizing intra-class variance, and their discriminant ratio is: (16) in , They represent the first i Features in Category k , k The mean and variance of =0,1; after normalization... Used as heuristic information to guide the search process; Step 4-3: Ant colony iterative optimization; 30 ants search the feature space in parallel. k Select features j The probability is determined by the pheromone concentration. With heuristic information Joint decision: (17) in α , β Adjust the pheromone and heuristic weights separately. δ To prevent smoothing terms with a denominator of zero, a dynamic threshold mechanism is introduced. If no feature is selected, a random feature is forcibly activated. The fitness function balances classification performance with the number of features. (18) Where S is the feature subset, and Accuracy is the accuracy of the K-NN classifier on the validation set. This is the sparsity penalty coefficient; Step 4-4: Elite Strategy Update: Pheromone on the globally optimal path is volatilized and enhanced. (19) in S is the volatility coefficient. ∗ The current optimal subset is used; after reaching the maximum number of iterations, the optimal subset S is output. ∗ ; Indicates the first t At +1 iteration, the features j pheromone concentration, Indicates the first t In the next iteration, the features j pheromone concentration, This represents the change in pheromone level of feature j during this iteration.

5. The method for detecting small sea surface targets by selecting dynamic features that vary with environmental changes, as described in claim 4, is characterized in that... Step 5 specifically involves: A simulated target echo generator was used to generate target samples of equal quantity to sea clutter data. The 8-dimensional features of the sea clutter and target samples were jointly extracted to construct a balanced training set. Calculate the Euclidean distance between the test sample and all training samples, and select the closest one. k A close neighbor; statistics k Number of clutter samples in the nearest neighbors n 0, introduce a decision threshold w :like n 0≥ w If it is determined to be clutter, it is determined to be a target; a fixed decision threshold is used. By dynamically adjusting the number of nearest neighbors k The value ensures a controllable false alarm rate; the false alarm probability is preset. P f Next, iterative optimization k Value makes the actual false alarm rate P fa satisfy , To find the minimum value, the optimal value is finally determined. During online testing, the parameters of each test sample are calculated. The number of clutter samples in the nearest neighbors, based on a threshold. A binary judgment is made.

6. The method for detecting small sea surface targets by selecting dynamic features of environmental variation according to claim 5, characterized in that, In step 1, the international benchmark dataset IPIX dataset is used as the measured radar echo signal.

7. The method for detecting small sea surface targets by selecting dynamic features of environmental variation according to claim 6, characterized in that, The =50Hz =5Hz.

8. The method for detecting small sea surface targets by selecting dynamic features of environmental changes according to claim 7, characterized in that, In steps 3-5, the feature extraction process for the number of connected regions (NR) and the maximum connected region size (MS) is as follows: First, dynamic thresholding is performed on NTFD, and the previous data is retained in each time unit. L The largest pixel value is assigned a value of 1, and the remaining pixels are set to zero to generate a thresholded NTFD image after binarization. Subsequently, region growing was performed based on the eight-neighbor connectivity criterion. The total number of all connected regions was counted as the NR feature value, and the total number of pixels contained in the largest connected region was recorded as the MS feature value.

9. The method for detecting small sea surface targets by selecting dynamic features of environmental variation according to claim 8, characterized in that, Maximum number of iterations in step 4 T =50.

Images