A method and system for identifying oil spills using shipborne radar
By combining multi-scale feature fusion and adaptive weight adjustment with a restricted Boltzmann machine model and a population intelligent optimizer, the problems of high false alarm rate and insufficient robustness in marine oil spill monitoring are solved, and the fine classification and extraction of oil spill areas are realized.
Patent Information
- Application Number
- CN202611553880.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-09
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies for marine oil spill monitoring suffer from high false alarm rates and insufficient robustness, making it difficult to achieve stable application across different scenarios. Furthermore, traditional methods lack multi-scale feature fusion mechanisms and adaptive feature weight adjustment, resulting in limited feature representation capabilities.
Multi-scale local statistical features are used to extract regions of interest in oil spills. An integrated response map is generated by adaptive weight fusion. A restricted Boltzmann machine model and a population intelligent optimizer are combined for global optimization. K-means clustering is used to identify oil spill areas.
It enables accurate and robust identification of oil spill areas in radar images, reduces false alarm rate, and improves feature representation capability and algorithm stability.
Smart Images

Figure CN123071689A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oil spill identification technology, specifically to a shipborne radar oil spill identification method and system. Background Technology
[0002] In the field of marine oil spill monitoring, Synthetic Aperture Radar (SAR) and shipborne radar have become core monitoring methods due to their all-weather, 24 / 7 observation capabilities. Oil spill areas, covered by an oil slick, significantly suppress the generation of capillary waves on the sea surface, leading to a decrease in the radar backscattering coefficient, which manifests as typical dark spots in images. However, in actual monitoring, numerous oil spill-like interference phenomena exist: internal waves alter sea surface roughness to form strip-shaped dark areas; low-wind-speed areas exhibit uniform dark spots due to a lack of wave energy; and biofilm adhesion produces texture features similar to oil spills. These interference sources have a highly similar visual appearance to real oil spills in radar images, leading to two major challenges for traditional detection methods based on grayscale threshold segmentation or single texture features: firstly, a persistently high false alarm rate (FAR); and secondly, insufficient algorithm robustness, being sensitive to changes in different sensor parameters, sea surface wind speed, and other environmental factors, making stable application across different scenarios difficult. While existing technologies attempt to extract multi-directional texture features through Gabor filtering, construct gray-level co-occurrence matrices (GLCM) to quantify spatial correlations, or combine them with support vector machines (SVM) for classification, these methods are still limited by feature representation capabilities. The fixed kernel function of Gabor filtering is difficult to adapt to the diversity of oil spill morphology, GLCM is sensitive to noise and has high computational complexity, and SVM has the risk of overfitting when processing high-dimensional nonlinear data.
[0003] Current mainstream oil spill detection frameworks mostly employ a combination of feature engineering and classifiers. Essentially, they approximate the solution space by manually designing features, but this approach has fundamental flaws. Taking the Restricted Boltzmann Machine (RBM) as an example, while its unsupervised feature learning capability can automatically extract deep representations of the data, the training process relies on approximate inferences from the Contrastive Divergence (CD) algorithm, making the model prone to getting trapped in local optima. This is especially true when dealing with small sample oil spill data, where the discriminative power of the feature space significantly decreases. Furthermore, existing methods generally lack multi-scale feature fusion mechanisms: oil spills in synthetic aperture radar images may present as point-like (small-scale leaks), patchy (medium-scale), or strip-like (large-scale diffusion), and single-scale feature extraction cannot simultaneously capture these spatial structural differences. At the same time, traditional methods often use fixed weights or equal-weight fusion for each feature channel, failing to dynamically adjust the contribution of key features according to sea state differences, thus limiting feature representation capabilities. Summary of the Invention
[0004] To solve or alleviate one or more of the above problems, this invention proposes a shipborne radar oil spill identification method and system.
[0005] According to one aspect of the present invention, a method for identifying oil spills using shipborne radar is proposed, the method comprising: S1. Acquire pre-processed images from shipborne radar. S2. Extracting the oil spill region of interest from the preprocessed image to obtain the oil spill region of interest includes: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weights of each texture feature based on their distribution characteristics in the image, fusing them to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest. S3. Extract the gray values of non-background pixels in the oil spill interest area, perform deduplication and normalization, and use them as training samples to train the restricted Boltzmann machine model; use a population-based intelligent optimizer to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases. S4. Using a restricted Boltzmann machine model with optimal weights and biases, extract the hidden activation features of non-background pixels in the oil spill interest region. S5. Perform unsupervised clustering on the hidden layer activation features, and determine the oil spill area based on the clustering results.
[0006] Furthermore, the various texture features mentioned in S2 include at least two of the following: local standard deviation feature, local entropy feature, gradient contrast feature, and dark area feature; the local standard deviation feature is obtained by calculating the standard deviation of gray values in the pixel neighborhood; the local entropy feature is obtained by calculating the entropy of the gray level probability distribution in the pixel neighborhood; the gradient contrast feature is obtained by calculating the magnitude of the horizontal gradient and the vertical gradient; and the dark area feature is obtained by subtracting the local mean from the original gray value.
[0007] Furthermore, the method for determining the adaptive fusion weights described in S2 includes: Normalize each texture feature by its maximum value to obtain normalized features; The reciprocal of the global standard deviation of the r-th normalized feature is used as the initial weight; A direction adjustment factor is introduced, and the initial weights are multiplied by the direction factor to obtain the weights after direction correction; wherein, features negatively correlated with oil spill response take the positive direction factor, and features positively correlated with oil spill response take the negative direction factor; The weights after direction correction are L1 normalized to obtain the final weights.
[0008] Furthermore, the method for generating the comprehensive oil spill response map described in S2 includes: inverting the normalized features that are negatively correlated with the oil spill response and multiplying them with the corresponding weights; directly multiplying the normalized features that are positively correlated with the oil spill response with the corresponding weights; and then adding all the terms together to obtain the comprehensive response value of each pixel. The process of threshold segmentation and post-processing of the integrated response map to obtain the region of interest (ROI) for the oil spill includes: determining the optimal segmentation threshold using the Otsu's maximum inter-class variance method; performing binarization segmentation on the integrated response map to obtain the initial mask for the ROI; and setting the wave-free areas in the initial mask for the ROI as the background based on prior knowledge to obtain the ROI.
[0009] Furthermore, the parameters to be optimized in the restricted Boltzmann machine model described in S3 include the connection weights between the visible and hidden layers, the visible layer bias, and the hidden layer bias; the optimizer is the FNO optimizer, which causes individuals to move closer to the farthest better individual in the population and further away from the nearest worse individual in the population; the farthest better individual is the individual with the largest Euclidean distance from the current individual in the set of better individuals, and the nearest worse individual is the individual with the smallest Euclidean distance from the current individual in the set of worse individuals.
[0010] Furthermore, the global optimization described in S3 includes: initializing the population; for each individual, calculating the Euclidean distance in its set of better individuals and set of worse individuals respectively, and selecting the farthest better individual and the closest worse individual; iteratively performing position updates and using a greedy selection mechanism to retain only new individuals with better fitness; continuously recording the global optimal parameters during the iteration process, and mapping them to the optimal weights and optimal biases of the restricted Boltzmann machine model after the iteration ends.
[0011] Furthermore, the position update formula for the FNO optimizer described in S3 is: ; in, This is a dynamic focusing factor that decreases with each iteration. This indicates the t-th iteration. Indicates the maximum number of iterations; For random perturbation vectors, express d Uniformly distributed in dimensions r 1 and r Each component of 2 independently follows a uniform distribution on the interval [0,1]. For Gaussian random walk, express d 3D standard normal distribution; This indicates element-wise multiplication; Indicates the empty set; Let ζ be the current position of the ζ-th individual. For the farthest and better individuals, For the most recent worse individual; This represents the updated position of the ζ-th individual.
[0012] Furthermore, the fitness function of the FNO optimizer described in S3 is the reconstructed mean square error of the restricted Boltzmann machine model. : ; in, N Indicates the number of training samples; Indicates the first k One original input; k Indicates the sample index; The visible layer output reconstructed for the restricted Boltzmann machine model.
[0013] Furthermore, the unsupervised clustering in S5 uses the K-means clustering algorithm to divide the pixels in the oil spill interest region into multiple classes, calculate the original gray-scale mean of each class of pixels, and sort them in ascending order. The class with the lowest gray-scale mean is determined as the oil spill area.
[0014] According to another aspect of the present invention, a shipborne radar oil spill identification system is proposed, the system being used to implement the shipborne radar oil spill identification method described above; the system includes: Image acquisition module, used to acquire preprocessed images from shipborne radar; The region of interest extraction module is used to extract the oil spill region of interest from the preprocessed image to obtain the oil spill region of interest. The oil spill region of interest extraction includes: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weights of each texture feature according to the distribution characteristics of each texture feature in the image, fusing them to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest. The model training module is used to extract the gray values of non-background pixels in the oil spill interest region, perform deduplication and normalization, and use them as training samples to train the restricted Boltzmann machine model; a population-based intelligent optimizer is used to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases. The feature extraction module is used to extract the hidden activation features of non-background pixels in the oil spill interest region using a restricted Boltzmann machine model with optimal weights and optimal biases. The oil spill identification module is used to perform unsupervised clustering on the hidden layer activation features and determine the oil spill area based on the clustering results.
[0015] The present invention has the following technical effects: This invention proposes a method and system for identifying oil spills using shipborne radar. Specifically, it extracts regions of interest (ROIs) based on multi-scale local statistical features, comprehensively considering various features reflecting the characteristics of the oil spill area. A comprehensive feature map is generated through adaptive weighted linear fusion, and Otsu's maximum inter-class variance method is used for threshold segmentation and post-processing, accurately extracting ROIs and providing a solid foundation for subsequent processing. A population-based optimizer is employed to globally optimize the RBM parameters, guiding individuals towards better individuals and away from worse ones. Using reconstruction error as the fitness function, this method can escape local minima in the parameter space, obtaining better weights and biases, thus improving the feature representation capability of the RBM. Hidden features are extracted using the trained RBM and K-means clustering is performed, mapping the original high-dimensional data to a low-dimensional feature space. Based on the distribution characteristics of the oil spill area in the feature space, the category with the lowest gray value is retained as the oil spill area, achieving refined classification and extraction of the oil spill area. As can be seen from the oil film detection results, the hidden layer feature mapping results, and the final planar position display effect, the present invention has high accuracy and robustness in identifying oil spill areas in radar images, and can intuitively and effectively segment the oil film area. Attached Figure Description
[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0017] Figure 1 This is a flowchart of a shipborne radar oil spill identification method according to an embodiment of the present invention; Figure 2 This is an example of a preprocessed image from a shipborne radar in an embodiment of the present invention; Figure 3 This is an example of a comprehensive response diagram in an embodiment of the present invention; Figure 4 This is an example of an oil spill interest area in an embodiment of the present invention; Figure 5 This is an example of oil film detection results in an embodiment of the present invention; Figure 6 This is an example of the hidden layer feature mapping result in an embodiment of the present invention; Figure 7 This is an example of the final planar position display effect in an embodiment of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are part of this invention.
[0019] This invention aims to solve the problems of weak feature representation ability, easy model training getting trapped in local optima, and high false alarm rate in existing radar oil spill detection technology, and proposes a shipborne radar oil spill detection method and system.
[0020] like Figure 1 As shown in the figure, this invention proposes a method for identifying oil spills using shipborne radar, which includes the following steps: S1. Acquire pre-processed images from shipborne radar. S2. Extracting the oil spill region of interest from the preprocessed image to obtain the oil spill region of interest; including: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weights of each texture feature according to their distribution characteristics in the image, fusing them to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest. S3. Extract the gray values of non-background pixels in the oil spill interest area, perform deduplication and normalization, and use them as training samples to train the restricted Boltzmann machine model; use a population-based intelligent optimizer to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases. S4. Using a restricted Boltzmann machine model with optimal weights and biases, extract the hidden activation features of non-background pixels in the oil spill interest region. S5. Perform unsupervised clustering on the hidden layer activation features, and determine the oil spill area based on the clustering results.
[0021] The method begins with S1. In S1, preprocessed images from the shipborne radar are acquired.
[0022] According to embodiments of the present invention, the preprocessing of shipborne radar images includes: coordinate transformation, detection and suppression of co-channel interference, speckle noise filtering, grayscale correction, and contrast enhancement. Specifically, firstly, the original radar image data in polar coordinates is transformed to Cartesian coordinates to standardize the spatial structure. Based on this, convolution calculation is performed to initially enhance high-frequency features; co-channel interference is extracted using a grayscale threshold of 150 and suppressed using mean filtering. Subsequently, a grayscale threshold of 120 is set to extract bright pixels, and bright connected regions with an area less than 200 pixels are further identified in the upper 2 / 3 of the image area as speckle noise for filtering, combined with 20×20 window midpoint filtering to further eliminate discrete noise. To further improve image quality, the global grayscale distribution is optimized using a grayscale adjustment matrix, and local contrast enhancement is performed within the approximate rectangular area of the oil film target.
[0023] The acquired preprocessed images from the shipborne radar are as follows: Figure 2 As shown, this image reflects the echo intensity distribution of different targets within the radar detection range, and serves as the foundational data for subsequent target detection and identification.
[0024] Then, in S2, the preprocessed image is subjected to oil spill region of interest extraction to obtain the oil spill region of interest, including: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weight of each texture feature according to its distribution characteristics in the image, fusing to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest.
[0025] According to an embodiment of the present invention, since there are significant differences in texture features between the marine oil spill area and the surrounding seawater, in step S21, multiple complementary texture features sensitive to the oil spill area are extracted to characterize the characteristics of the oil spill area from different dimensions; then in step S22, based on the actual distribution of each feature in the image, an adaptive weighting strategy is adopted to determine the adaptive fusion weight of each feature; finally in step S23, the oil spill ROI is extracted by Otsu automatic threshold segmentation and post-processed to provide a data foundation for subsequent fine classification.
[0026] Among them, multiple texture features include at least two of the following: local standard deviation features, local entropy features, gradient contrast features, and dark area features.
[0027] Local standard deviation (LSD) reflects the degree of gray-level dispersion in local areas of an image and is an effective indicator of texture roughness. LSD is obtained by calculating the standard deviation of gray-level values within the neighborhood of a pixel. Specifically, oil splatter areas typically exhibit smoother texture characteristics, with lower local standard deviations for each pixel. Take a 7×7 neighborhood and calculate the standard deviation of gray levels within the neighborhood. As a feature of local standard deviation: ; in, It is a pixel. (i, j) A 7×7 neighborhood set, in which the total number of pixels is N 7, Indicates position grayscale value, It is the average gray value of the 7×7 neighborhood.
[0028] Local entropy features reflect the complexity of a texture; the higher the entropy, the more complex and information-rich the texture. Oily areas, being smooth, have low entropy values. Local entropy features are obtained by calculating the entropy of the gray-level probability distribution within a pixel's neighborhood. Specifically, for a 9×9 neighborhood, the local entropy is calculated. : ; in, For the neighborhood Inner gray level The probability distribution.
[0029] Gradient contrast features are obtained by calculating the magnitudes of the horizontal and vertical gradients. Gradient information reflects the edge intensity and structural changes of an image. Oil spill areas, due to their gentle internal gray-level changes, exhibit lower gradient magnitudes; while wave textures and oil-water boundaries in seawater areas show higher gradient responses. Specifically, the horizontal gradient is first calculated... and vertical gradient Then calculate the gradient magnitude. : ; .
[0030] Oil spill areas typically appear as darker grayscale features in remote sensing images. To enhance the response of dark areas, the dark area feature is defined as the difference between the local mean and the original grayscale. Specifically, a local mean with a window size of 15×15 is used as the background estimate, and then the difference is calculated with the original image. Dark Area Feature The calculation formula is: ; in, This is the local mean within a 15×15 window. This feature produces a positive response to dark areas, effectively enhancing the salience of the oil spill area.
[0031] Traditional fixed-weight methods struggle to adapt to the varying feature distributions across different images. This invention proposes an adaptive weight allocation strategy based on feature discriminative power. The core idea is that the higher the discriminative power of a feature, the greater its contribution to the fusion process. Discriminative power is measured by the standard deviation of the feature; a larger standard deviation indicates more significant variations in the feature across different regions of the image, resulting in stronger ability to distinguish between the target and the background.
[0032] The determination of adaptive fusion weights includes: S221, Since the dimensions and numerical ranges of each feature differ, normalization processing is required. Specifically, the first... r Features Normalized to the maximum value The initial weights are defined as the reciprocal of the standard deviation, and the calculation of the first weight... r Initial weights of each feature w r (0) for: ; in, It is the first r Normalized features global standard deviation r =1,2,…,4.
[0033] S222. Considering that the response directions of various features to oil spills are different (local standard deviation features) F 1. Local entropy characteristics F 2. Gradient contrast features F 3 is negatively correlated, a characteristic of dark areas. F 4 is positively correlated), so a direction adjustment factor is introduced. , set initial weight With the corresponding direction factor Multiply to obtain the weights after direction correction. Features negatively correlated with oil spill response are assigned positive directional factors, while features positively correlated with oil spill response are assigned negative directional factors. ; S223, Through L 1. Normalization yields the final weights : ; in, q This is the index for traversing all features.
[0034] The four normalized feature maps are linearly fused according to adaptive weights to generate the final comprehensive oil spill response map. S ,like Figure 3 As shown. To ensure that a higher response value indicates a greater likelihood that the pixel belongs to an oil spill area, the first three features are inverted. Specifically, the normalized features negatively correlated with the oil spill response are inverted and multiplied by their corresponding weights, while the normalized features positively correlated with the oil spill response are directly multiplied by their corresponding weights. Finally, all these terms are summed to obtain the comprehensive response value for each pixel. for: .
[0035] The optimal segmentation threshold is automatically determined using Otsu's maximum inter-class variance method, and the comprehensive response map is analyzed. Binarization segmentation is performed to extract the initial mask for the region of interest (ROI). Based on prior knowledge, the upper half of the initial ROI mask, the wave-free area (the first 350 rows), is forcibly set as the background. The final extracted oil spill ROI is shown below. Figure 4 As shown.
[0036] Then, in S3, the gray values of non-background pixels in the oil spill interest area are extracted, deduplicated, and normalized as training samples to train a restricted Boltzmann machine model. A population-based optimizer is used to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases.
[0037] According to an embodiment of the present invention, based on the ROI extraction results, a fine classification model based on RBM is introduced to perform multi-level grayscale clustering on the pixels in the candidate region, so as to achieve progressive identification of the oil spill area from coarse to fine.
[0038] The Restricted Boltzmann Machine (RBM) model is an energy-based generative stochastic neural network consisting of visible layer state vectors. and hidden state vector The structure consists of no connections within a layer and full connections between layers. Represents the set of real numbers. express A 3D real vector space. Set the number of visible layer units. (Input is normalized grayscale value), number of hidden layer units The parameters to be optimized include the connection weights between the visible and hidden layers, the visible layer bias, and the hidden layer bias, totaling [number missing]. Its energy function Defined as: ; in, For connecting the weight matrix; The visible layer bias vector; This is the hidden layer bias vector.
[0039] Given the visible layer state v Under the conditions, the first z Hidden layer units Activated ( The probability of (=1) for: ; in, Use the Sigmoid activation function; For the visible layer u The unit and the hidden layer z The connection weights of each unit; For the first z Bias terms for each hidden unit; For the visible layer u The state of each unit.
[0040] Similarly, the visible layer reconstruction probability is: ; in, For the visible layer u The bias term of each unit; Indicates the visible layer number. u One unit was activated.
[0041] Traditional Restricted Block Models (RBMs) typically employ the Contrastive Divergence (CD) algorithm for training, aiming to maximize the log-likelihood of the training samples. Its core limitations include a tendency to get trapped in local optima; biased gradient estimation leading to training divergence; and the fact that the gradient field does not belong to any objective function causing the update direction to deviate from the true gradient direction. To address these issues, this invention introduces Functional Non-Optimization (FNO) to replace CD for global optimization of the RBM network parameters.
[0042] The Farthest Better or Nearest Worse Optimizer (FNO) is a global optimization algorithm based on population intelligence. Its core idea is that each individual moves closer to the furthest better individual in the population to find a better global optimum, while simultaneously moving away from the nearest worse individual, thus avoiding getting trapped in local optima. Its training process is as follows: (1) Population initialization: Uniformly and randomly initialize the population within the search space. Each individual's location is used to calculate the location of each individual. fitness value And record the current global optimal solution.
[0043] (2) Select two key neighborhood individuals: for each individual Calculate its Euclidean distance to other individuals, and select the farthest better individual and the nearest worse individual; where the farthest better individual is the one with better fitness than the others. The one with the largest distance among the individuals; the one with the worst fitness is the one with the lowest fitness. The one with the smallest distance among the individuals.
[0044] (3) Perform two-stage position update: Based on the dynamic focus factor α, the ratio of exploration to development is controlled, and the process is divided into two stages: a. Exploration Phase: When an individual is located between the farthest better individual and the nearest worse individual, the algorithm skips the nearest worse individual and moves directly towards the farthest better individual to avoid getting trapped in local low-quality regions; b. Development Phase: The algorithm uses the Dynamic Focus Strategy (DFS) to generate a shrinkage coefficient and performs a fine search near the farthest better individual to gradually approach the global optimum. DFS controls the search step size through a dynamic focus factor α, which decreases from 1 to 0 with each iteration, achieving a smooth transition from global exploration to local development.
[0045] (4) Greedy selection mechanism: The fitness of the new position is compared with that of the old position. The position of an individual is updated only if the fitness of the new position is better; otherwise, the original position is retained. This ensures that the quality of the population will not degenerate.
[0046] (5) Iteration Termination: Repeat steps (2) to (4) until the maximum number of iterations is reached or other termination conditions are met, and finally output the global optimal position. and its fitness value .
[0047] This invention uses the aforementioned FNO optimization algorithm as the global optimizer for RBM, adapting its optimization object from the general objective function to the network parameters of the RBM. The fitness function is defined as the reconstruction mean square error of the RBM. The algorithm iteratively updates the population position, gradually approaching the global optimum. The FNO optimization algorithm does not rely on gradient information at all, driving the search solely through fitness comparisons and distance calculations among individuals in the population, fundamentally bypassing the biased gradient problem of CD. Through population-parallel search and a mechanism of "learning from distant optima and avoiding near inferior ones," the FNO optimization algorithm maintains diversity in the parameter space, reducing the risk of getting trapped in local optima. Furthermore, the FNO optimization algorithm's DFS strategy automatically adjusts the search step size, exploring with large steps in the early stages and fine-tuning with small steps in the later stages, eliminating the need for manual parameter tuning.
[0048] Specifically, let the first ζ The current location of each individual is The fitness function is Definition: A better set of individuals ; Indicates an index of other individuals in the population. Indicates the first The current position of each individual; the set of worse individuals. .
[0049] like If it is not an empty set, then the furthest better individual is... : ; in, Euclidean distance represents the spatial distance between two points.
[0050] like If it is not an empty set, then the worst nearest individual is... : .
[0051] The position update formula for this optimization algorithm is as follows: ; in, This is a dynamic focusing factor that decreases with each iteration. This indicates the t-th iteration. Indicates the maximum number of iterations; For random perturbation vectors, express d Uniformly distributed in dimensions r 1 and r Each component of 2 independently follows a uniform distribution on the interval [0,1]. For Gaussian random walk, express d3D standard normal distribution; This indicates element-wise multiplication; Indicates the empty set; This represents the updated position of the ζ-th individual.
[0052] The goal of FNO optimization is to minimize the mean squared error (MSE) of the reconstruction of the RBM. Specifically, for the input sample set... The fitness function MSE is defined as follows: ; in, N Indicates the number of training samples; Indicates the first k One original input; k Indicates the sample index; Visible layer output for RBM reconstruction: .
[0053] The process of using the FNO optimization algorithm to globally optimize the parameters of the restricted Boltzmann machine model is as follows: Extract all non-background pixels (grayscale values) from the oil spill area of interest image. The extracted grayscale values are deduplicated so that only one sample is retained for each different grayscale level, resulting in a grayscale level set; the grayscale values in the grayscale level set are then normalized to... As training samples; Initialize the FNO population: Population size Maximum number of iterations The search boundary is [-5,5]; For each individual, calculate the Euclidean distance in its set of better individuals and set of worse individuals respectively, and select the farthest better individual and the nearest worse individual; Iteratively perform position updates and use a greedy selection mechanism to retain only new positions with better fitness, continuously recording the globally optimal parameters; After the iteration stops, the globally optimal parameters are mapped to the optimal weights of the RBM in a fixed order. With optimal bias .
[0054] Then, in S4, the hidden activation features of non-background pixels in the oil spill region of interest are extracted using a restricted Boltzmann machine model with optimal weights and biases.
[0055] According to an embodiment of the present invention, the hidden activation features within the oil spill region of interest are extracted using a trained Restricted Boltzmann Machine (RBM). Specifically, for each non-background pixel within the oil spill region of interest in the image to be identified, the grayscale value is deduplicated and normalized, then input into the trained RBM model. The hidden activation features of that pixel are then calculated using the hidden activation function. ; In the formula, Indicates the first k Hidden layer features.
[0056] Through RBM nonlinear transformation, the original pixel grayscale values are mapped to a 5-dimensional hidden layer feature space, resulting in a more separable high-dimensional feature representation.
[0057] Based on this, in S5, unsupervised clustering is performed on the hidden layer activation features, and the oil spill area is determined according to the clustering results.
[0058] According to an embodiment of the present invention, the K-means clustering algorithm is used for unsupervised segmentation of hidden layer activation features. Based on the distribution characteristics of the oil spill area in the feature space, the number of cluster categories is set. The samples are divided into four categories, the original gray-scale mean of each category is calculated, and the samples are sorted in ascending order of gray-scale mean. The category with the lowest gray-scale mean is retained as the oil spill area, thereby achieving fine classification and extraction of the oil film.
[0059] Figure 5 The oil film detection results are shown, and the effective segmentation of the oil film area by the method of the present invention can be seen intuitively; Figure 6 The hidden layer feature mapping results reflect the clustering distribution of samples in the feature space; Figure 7 To further verify the accuracy and robustness of the method of the present invention in identifying oil spill areas in radar images, the final planar position display effect was used.
[0060] This invention also proposes a shipborne radar oil spill identification system, which is used to implement the shipborne radar oil spill identification method described in the above embodiments; the system includes: Image acquisition module, used to acquire preprocessed images from shipborne radar; The region of interest extraction module is used to extract the oil spill region of interest from the preprocessed image to obtain the oil spill region of interest. The oil spill region of interest extraction includes: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weights of each texture feature according to the distribution characteristics of each texture feature in the image, fusing them to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest. The model training module is used to extract the gray values of non-background pixels in the oil spill interest region, perform deduplication and normalization, and use them as training samples to train the restricted Boltzmann machine model; a population-based intelligent optimizer is used to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases. The feature extraction module is used to extract the hidden activation features of non-background pixels in the oil spill interest region using a restricted Boltzmann machine model with optimal weights and optimal biases. The oil spill identification module is used to perform unsupervised clustering on the hidden layer activation features and determine the oil spill area based on the clustering results.
[0061] It should be noted that the function of the shipborne radar oil spill identification system described in the embodiments of the present invention can be explained by the aforementioned shipborne radar oil spill identification method. Therefore, for the parts not described in detail in the system embodiments, please refer to the above method embodiments, and they will not be repeated here.
[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.
Claims
1. A method for identifying oil spills using shipborne radar, characterized in that, Includes the following steps: S1. Acquire pre-processed images from shipborne radar. S2. Extracting the oil spill region of interest from the preprocessed image to obtain the oil spill region of interest includes: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weights of each texture feature based on their distribution characteristics in the image, fusing them to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest. S3. Extract the gray values of non-background pixels in the oil spill interest area, perform deduplication and normalization, and use them as training samples to train the restricted Boltzmann machine model; use a population-based intelligent optimizer to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases. S4. Using a restricted Boltzmann machine model with optimal weights and biases, extract the hidden activation features of non-background pixels in the oil spill interest region. S5. Perform unsupervised clustering on the hidden layer activation features, and determine the oil spill area based on the clustering results.
2. The method according to claim 1, characterized in that, The various texture features mentioned in S2 include at least two of the following: local standard deviation feature, local entropy feature, gradient contrast feature, and dark area feature; the local standard deviation feature is obtained by calculating the standard deviation of gray values in the pixel neighborhood; the local entropy feature is obtained by calculating the entropy of the gray level probability distribution in the pixel neighborhood; the gradient contrast feature is obtained by calculating the magnitude of the horizontal gradient and the vertical gradient; and the dark area feature is obtained by subtracting the local mean from the original gray value.
3. The method according to claim 1, characterized in that, The method for determining the adaptive fusion weights described in S2 includes: Normalize each texture feature by its maximum value to obtain normalized features; The reciprocal of the global standard deviation of the r-th normalized feature is used as the initial weight; A direction adjustment factor is introduced, and the initial weights are multiplied by the direction factor to obtain the weights after direction correction; wherein, features negatively correlated with oil spill response take the positive direction factor, and features positively correlated with oil spill response take the negative direction factor; The weights after direction correction are L1 normalized to obtain the final weights.
4. The method according to claim 3, characterized in that, The method for generating the comprehensive oil spill response map described in S2 includes: inverting the normalized features that are negatively correlated with the oil spill response and multiplying them with the corresponding weights; directly multiplying the normalized features that are positively correlated with the oil spill response with the corresponding weights; and then adding all the terms together to obtain the comprehensive response value of each pixel. The process of threshold segmentation and post-processing of the integrated response map to obtain the region of interest (ROI) for the oil spill includes: determining the optimal segmentation threshold using the Otsu's maximum inter-class variance method; performing binarization segmentation on the integrated response map to obtain the initial mask for the ROI; and setting the wave-free areas in the initial mask for the ROI as the background based on prior knowledge to obtain the ROI.
5. The method according to claim 1, characterized in that, The parameters to be optimized in the restricted Boltzmann machine model described in S3 include the connection weights between the visible and hidden layers, the visible layer bias, and the hidden layer bias; the optimizer is the FNO optimizer, which moves individuals closer to the farthest better individual in the population and further away from the nearest worse individual in the population; the farthest better individual is the individual with the largest Euclidean distance from the current individual in the set of better individuals, and the nearest worse individual is the individual with the smallest Euclidean distance from the current individual in the set of worse individuals.
6. The method according to claim 5, characterized in that, The global optimization described in S3 includes: initializing the population; for each individual, calculating the Euclidean distance in its set of better individuals and set of worse individuals respectively, and selecting the farthest better individual and the closest worse individual; iteratively performing position updates and using a greedy selection mechanism to retain only new individuals with better fitness; continuously recording the global optimal parameters during the iteration process, and mapping them to the optimal weights and optimal biases of the restricted Boltzmann machine model after the iteration ends.
7. The method according to claim 6, characterized in that, The position update formula for the FNO optimizer described in S3 is: ; in, This is a dynamic focusing factor that decreases with each iteration. This indicates the t-th iteration. Indicates the maximum number of iterations; For random perturbation vectors, express d Uniformly distributed in dimensions r 1 and r Each component of 2 independently follows a uniform distribution on the interval [0,1]. For Gaussian random walk, express d 3D standard normal distribution; This indicates element-wise multiplication; Represents the empty set; Let ζ be the current position of the ζ-th individual. For the farthest and better individuals, For the most recent worse individual; This represents the updated position of the ζ-th individual.
8. The method according to claim 7, characterized in that, The fitness function of the FNO optimizer described in S3 is the reconstructed mean square error of the restricted Boltzmann machine model. : ; in, N Indicates the number of training samples; Indicates the first k One original input; k Indicates the sample index; The visible layer output reconstructed for the restricted Boltzmann machine model.
9. The method according to claim 1, characterized in that, The unsupervised clustering described in S5 uses the K-means clustering algorithm to divide the pixels in the oil spill interest area into multiple classes, calculate the original gray-scale mean of each class of pixels, sort them in ascending order, and determine the class with the lowest gray-scale mean as the oil spill area.
10. A shipborne radar oil spill identification system, characterized in that, The system is used to implement the shipborne radar oil spill identification method according to any one of claims 1-9; the system includes: Image acquisition module, used to acquire pre-processed images from shipborne radar; The region of interest extraction module is used to extract the oil spill region of interest from the preprocessed image to obtain the oil spill region of interest. The oil spill region of interest extraction includes: extracting multiple texture features to describe the difference between the oil spill area and seawater, determining the adaptive fusion weights of each texture feature according to the distribution characteristics of each texture feature in the image, fusing them to generate an oil spill comprehensive response map, and performing threshold segmentation and post-processing on the comprehensive response map to obtain the oil spill region of interest. The model training module is used to extract the gray values of non-background pixels in the oil spill interest region, perform deduplication and normalization, and use them as training samples to train the restricted Boltzmann machine model; a population-based intelligent optimizer is used to globally optimize the parameters of the restricted Boltzmann machine model to obtain the optimal weights and optimal biases. The feature extraction module is used to extract the hidden activation features of non-background pixels in the oil spill interest region using a restricted Boltzmann machine model with optimal weights and optimal biases. The oil spill identification module is used to perform unsupervised clustering on the hidden layer activation features and determine the oil spill area based on the clustering results.
