Oil film detection method and system based on hierarchical self-organizing network and scale adaptation

Abstract

The present invention proposes an oil film detection method and system based on a hierarchical self-organizing network and scale adaptation, which relates to the technical field of oil film detection. The technical points of the present invention include: pre-processing the acquired original radar image; using a trained oil film recognition model based on an improved growing hierarchical neural gas network to identify the oil film area of the pre-processed radar image to obtain the oil film candidate area; using an improved multi-scale threshold segmentation algorithm to segment the oil film candidate area to obtain the final oil film detection result. The present invention significantly improves the adaptability, robustness and computational efficiency of the detection model by introducing a local feature enhancement mechanism, dynamic hierarchical adjustment, adaptive error attenuation, sparse constraint optimization, and a real-time adaptive learning mechanism. The present invention can be integrated into a shipborne radar system, has high computational efficiency, can meet the real-time processing requirements of shipborne equipment, and is suitable for practical application scenarios such as offshore oil spill monitoring and emergency response.

Description

Technical Field

[0001] The present invention relates to the technical field of oil film detection, and in particular to an oil film detection method and system based on a hierarchical self-organizing network and scale adaptation. Background Art

[0002] Currently, remote sensing technology is widely used in oil spill monitoring, primarily using optical remote sensing and microwave radar. Optical remote sensing can capture information across multiple wavelengths, providing rich spectral data that facilitates accurate identification of oil slicks. However, optical remote sensing is susceptible to interference from lighting and weather conditions, making it difficult to achieve real-time monitoring of the spread of oil slicks. Unlike optical remote sensing, microwave radar generates surface echo images by transmitting electromagnetic waves to the sea surface and receiving the reflected scattered signals. In these images, oil-spilled sea surface areas have relatively weaker backscatter signals compared to areas not covered by the oil slick, resulting in dark spots on the microwave radar surface echo images. Microwave radar has become one of the most effective oil spill monitoring methods due to its all-weather, real-time, and efficient monitoring capabilities, as well as its ability to withstand adverse sea conditions to a certain extent.

[0003] Traditional methods rely on threshold segmentation, making them difficult to handle with complex backgrounds, uneven lighting, and variable oil film morphology. Existing supervised learning algorithms can classify data, but they require data labeling during the learning process. Furthermore, while traditional threshold segmentation methods can achieve localized segmentation optimization, they are inadequate for multi-scale features. Summary of the Invention

[0004] In view of the above problems, the present invention proposes an oil film detection method and system based on hierarchical self-organizing network and scale adaptation, which improves the accuracy of oil film detection by integrating dynamic self-organizing network and multi-scale threshold technology.

[0005] According to one aspect of the present invention, a method for oil film detection based on a hierarchical self-organizing network and scale adaptation is proposed, the method comprising:

[0006] Preprocessing the acquired raw radar image;

[0007] The oil film area of the pre-processed radar image is identified using the trained oil film recognition model based on the improved growing hierarchical neural gas network to obtain the candidate oil film area.

[0008] The oil film candidate area is segmented using an improved multi-scale threshold segmentation algorithm to obtain the final oil film detection result.

[0009] Furthermore, the preprocessing includes: performing lateral operator convolution on the original radar image, using the Otsu algorithm to perform vertical noise detection, noise smoothing, speckle removal, and contrast enhancement.

[0010] Furthermore, the training process of the oil film recognition model based on the improved growth hierarchical neural gas network includes:

[0011] Initialize neurons, the reference vector of each neuron is a sample point randomly selected from the training set S; the error variable is initialized to zero and the age of the edge is zero;

[0012] In the growth phase, each iteration randomly selects an input sample point, that is, a neuron, calculates the Euclidean distance between the neuron and all other neurons, and finds the neuron closest to the neuron. and subproximal neurons , increase the age of all connections from neuron q; update the nearest neuron Error variables, neurons and its neighbors’ reference vectors;

[0013] If neuron q and neuron s are not connected, create a new connection, otherwise reset the connection age to zero; delete connections whose age exceeds the threshold, remove unconnected neurons; and insert new neurons;

[0014] If the number of neurons in a graph is less than or equal to 2, prune it, otherwise generate its corresponding subgraph; recursively generate a subgraph for each neuron, and use the receptive fields of multiple neurons as the training set for training until the preset maximum layer depth is reached;

[0015] Calculate the average quantization error of the backup graph and the current graph, and compare the average quantization error. If the stopping condition is met, the growth stops and enters the convergence stage;

[0016] During the convergence phase, the error variables of all neurons are decayed through an adaptive decay mechanism;

[0017] Output the clustering results of each layer, extract the oil film candidate area, and the training is completed.

[0018] Furthermore, the reference vector is updated according to the following formula: :

[0019]

[0020] Where w i (t) represents the reference vector of neuron i at time t; η(t) represents the learning rate that decreases over time; α represents the global feature weight, which ranges from [0, 1]; β represents the local feature weight, which ranges from [0, 1]; Represents a sample point randomly selected from the training set S local features.

[0021] Furthermore, the reference vector is updated according to the following formula: :

[0022]

[0023] Where min means minimization; W represents the reference vector set of all neurons; S is the training set; Represents the best matching unit of the input sample point x(t) in the reference vector set of all neurons; The weight parameter representing the sparsity constraint; represents the neuron reference vector The L1 norm of ; N represents the total number of neurons.

[0024] Furthermore, the formula for attenuating the error variables of all neurons through the adaptive attenuation mechanism is:

[0025]

[0026] Where, represents the error variable of neuron i at time t; γ(t) represents the dynamic attenuation factor,

[0027]

[0028] Where γ0 is the initial decay constant, ranging from 0 to 1. Δ Data Complexity represents the rate of change of the complexity of the input data, which is obtained by calculating the difference in information entropy of the input data at adjacent time steps. Represents the attenuation adjustment parameter, which is used to control the dynamic attenuation factor Sensitivity to changes in data complexity.

[0029] Furthermore, the segmenting of the oil film candidate area using an improved multi-scale threshold segmentation algorithm includes:

[0030] For the oil film candidate area, local thresholds are calculated at different neighborhood scales:

[0031]

[0032] Where, Represents the pixel at the kth scale The local threshold of represents the local mean; represents the local standard deviation; is the sensitivity coefficient; Represents pixel points The local feature significance of

[0033]

[0034] in, Represents pixel points The gradient vector of ; max means taking the maximum value;

[0035] The local thresholds at each scale are averaged to generate the final threshold:

[0036] Where, represents the final threshold after fusion; M represents the total number of scales; w k Represents the pixel at the kth scale The spatial weight of

[0037] According to the final threshold The oil film candidate area is segmented, the oil film boundary is extracted, and the final oil film detection result is obtained.

[0038] Furthermore, the sensitivity coefficient in the improved multi-scale threshold segmentation algorithm is The calculation formula is:

[0039]

[0040] Where k0 represents the initial sensitivity coefficient; represents the sensitivity adjustment parameter, which is used to control the degree to which the sensitivity changes with local complexity; Represents pixel points The entropy value of the neighborhood grayscale value; MaxComplexit represents the maximum entropy value of all neighborhood grayscale values.

[0041] Furthermore, the pixel point at the kth scale in the improved multi-scale threshold segmentation algorithm The spatial weight w k The calculation formula is:

[0042]

[0043] Where, Indicates the center position of the kth scale, representing the core position of the scale in the image space; Represents pixel points The distance from the center of the kth scale; σ k represents the scale-dependent standard deviation.

[0044] According to another aspect of the present invention, an oil film detection system based on a hierarchical self-organizing network and scale adaptation is proposed, the system comprising:

[0045] An image preprocessing module configured to preprocess the acquired raw radar image;

[0046] an oil film region recognition module configured to use a trained oil film recognition model based on an improved growing hierarchical neural gas network to perform oil film region recognition on the preprocessed radar image to obtain candidate oil film regions;

[0047] The oil film segmentation module is configured to segment the oil film candidate area using an improved multi-scale threshold segmentation algorithm to obtain a final oil film detection result.

[0048] The beneficial technical effects of the present invention are:

[0049] Identifying oil slicks is a crucial step in oil spill identification, and incorporating feature information during the experimental process can help improve identification accuracy. This paper proposes an oil slick detection method and system based on a hierarchical self-organizing network and scale adaptation. This method uses an improved growing hierarchical neural gas network for unsupervised clustering of oil slicks and employs an improved multiscale adaptive threshold segmentation algorithm for precise oil slick extraction. This paper innovatively improves the traditional growing hierarchical neural gas network model and the multiscale adaptive threshold segmentation algorithm, combining their unique advantages for efficient oil slick identification. By introducing a local feature enhancement mechanism, dynamic hierarchical adjustment, adaptive error attenuation, sparse constraint optimization, and a real-time adaptive learning mechanism, the model's adaptability, robustness, and computational efficiency are significantly enhanced. Regarding the multiscale adaptive threshold segmentation process, the introduction of local feature significance analysis, multiscale threshold fusion based on spatial weights, adaptive sensitivity adjustment, and threshold optimization based on significance analysis further enhances the ability to accurately extract oil slick boundaries and improves adaptability to complex backgrounds and changing environments. Through these improvements and optimizations, the model's parameters can be dynamically optimized based on the threshold processing results, significantly improving oil slick detection accuracy in complex scenarios.

[0050] The present invention can be integrated into shipborne radar systems, has high computational efficiency, can meet the real-time processing requirements of shipborne equipment, and is suitable for practical application scenarios such as offshore oil spill monitoring and emergency response. The present invention improves the accuracy of shipborne radar systems in detecting marine oil spill films under complex sea conditions, provides efficient technical support for marine environmental protection and emergency response to oil spill accidents, and has significant social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0052] Figure 1 This is a flow chart of an oil film detection method based on a hierarchical self-organizing network and scale adaptation according to an embodiment of the present invention;

[0053] Figure 2 This is another flow chart of an oil film detection method based on a hierarchical self-organizing network and scale adaptation according to an embodiment of the present invention;

[0054] Figure 3 is an example diagram of a radar image according to an embodiment of the present invention;

[0055] Figure 4 is an example diagram of a radar image after preprocessing in an embodiment of the present invention;

[0056] Figure 5 This is an example image of an oil film candidate area obtained by using an oil film recognition model based on an improved growing hierarchical neural gas network in an embodiment of the present invention;

[0057] Figure 6 This is an example of an image of the final oil film detection result obtained by using the improved multi-scale threshold segmentation algorithm in an embodiment of the present invention;

[0058] Figure 7 This is an example image of the oil film detection result after deleting spots in an embodiment of the present invention;

[0059] Figure 8 It is a structural diagram of an oil film detection system based on a hierarchical self-organizing network and scale adaptation described in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0061] Those skilled in the art will appreciate that embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present disclosure may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software. It should be understood that any number of elements in the figures is for illustrative purposes only and not limiting, and any nomenclature is for distinction only and does not have any limiting meaning.

[0062] The present invention proposes an oil film detection method and system based on hierarchical self-organizing network and scale adaptation, which provides an intelligent processing method for oil spill identification and is used for shipborne microwave radar oil spill monitoring tasks.

[0063] The embodiment of the present invention proposes an oil film detection method based on a hierarchical self-organizing network and scale adaptation, such as Figures 1 and 2 As shown, the method includes:

[0064] S1. Preprocessing the acquired original radar image;

[0065] S2. Using the trained oil film recognition model based on the improved growing hierarchical neural gas network to identify the oil film area in the preprocessed radar image, and obtain the oil film candidate area;

[0066] S3. Segment the oil film candidate area using an improved multi-scale threshold segmentation algorithm to obtain a final oil film detection result.

[0067] The method starts from S1. In S1, the acquired raw radar image is preprocessed.

[0068] According to an embodiment of the present invention, the collected original radar image is as follows: Figure 3 As shown. The original radar image is preprocessed as follows: the image is convolved with a horizontal operator, the vertical noise detection is performed using the Otsu algorithm, and then the original radar image is converted into high-quality output data through noise smoothing, speckle removal, and contrast enhancement operations, providing a clear and less noisy basis for subsequent oil film detection. It should be noted that the preprocessing process is an existing preprocessing process for radar images, so it will not be repeated in this embodiment. The preprocessed image is as shown in FIG. Figure 4 shown.

[0069] Then, in S2, the trained oil film recognition model based on the improved growing hierarchical neural gas network is used to identify the oil film area of the preprocessed radar image to obtain the oil film candidate area.

[0070] According to an embodiment of the present invention, the training process of the original Growing Hierarchical Neural Gas Network (GHNG) is as follows.

[0071] 1) Initialization phase: Starting from the root graph, initialize two neurons, and the reference vector of each neuron Randomly extract from the training set S, the position of the initial neuron is determined by the input data distribution; initialize the error variable is zero, the edge age is initialized to zero.

[0072] When a grayscale image is input, the grayscale value of each pixel is used as an input sample, and the model completes the task by learning the distribution and characteristics of these grayscale values; all pixels in an image can form a training set or sample set, and each pixel is used as an independent sample.

[0073] 2) In the growth phase, the GHNG model dynamically adds neurons through an error-driven learning mechanism; the specific steps are as follows: Calculate the input sample point (i.e. a neuron) with all other neurons, find the neuron closest to it and subproximal neurons , increase the age of all connections from q:

[0074]

[0075]

[0076] Update the nearest neuron The error variable will be The error variable e accumulated to q q middle:

[0077]

[0078] Update neurons and its neighbors' reference vectors:

[0079]

[0080] in, is the learning rate, defined as: ; control the learning rate of q and neighbors respectively.

[0081] 3) If q and s are not connected, create a new connection; otherwise, reset the connection age to zero. Delete connections whose age exceeds the threshold and remove disconnected neurons.

[0082] 4) Insert new neurons (every λ steps): Select the neuron r with the largest error and its neighbor z, insert a new neuron k, and the prototype is Connect k with r and z, and attenuate e r and e z (i.e., the error variable of neuron r and its neighbor neuron z).

[0083] 5) Hierarchical training and recursive subgraph generation: The GHNG model generates a hierarchical structure through recursive training. If the number of neurons in a graph is H≤2, it is pruned; otherwise, a subgraph corresponding to each neuron i is generated, and the receptive field of each neuron is Defined as: in, is the training set of the current graph; for each neuron , recursively generate subgraphs, and use The training is performed as a training set until the preset maximum layer depth is reached.

[0084] 6) Growth control mechanism: GHNG model uses parameters Control the growth of the graph; the specific steps are as follows: Calculate the mean quantization error (MQE) of the backup graph and the current graph:

[0085] in, and are the neuron sets of the backup graph and the current graph respectively; compare the average quantization error, and if the following conditions are met, the growth stops and enters the convergence stage:

[0086] 7) Error decay and convergence phase: In the convergence phase, the error variables of all neurons are reduced by a constant attenuation:

[0087]

[0088] in, is the decay constant.

[0089] 8) Transformation invariance of the hierarchical structure: The hierarchical structure of the GHNG model is invariant to uniform scaling and rigid transformation of the input data. The transformed neuron prototype and error variable are:

[0090]

[0091] in, is the scaling factor; is an orthogonal matrix; is the translation vector.

[0092] 9) Output the clustering results of each layer and extract the candidate oil film areas.

[0093] In this embodiment, the original growing hierarchical neural gas network is improved as follows.

[0094] 1) Optimize the update process of the reference vector to enable the model to better adapt to the needs of oil film detection in complex sea conditions and improve detection accuracy and robustness. Two different improvement strategies are adopted.

[0095] 11) A local feature enhancement mechanism is introduced to combine local feature information when updating the reference vector to enhance the model's robustness to complex backgrounds and noise. The improved reference vector update formula is:

[0096]

[0097] Where w i (t) represents the reference vector of neuron i at time t; η(t) represents the learning rate that decreases over time, , is the initial learning rate, λ is the decay coefficient, which can be adjusted according to actual conditions; α represents the global feature weight, ranging from [0, 1], which is used to balance the role of global features in updating the reference vector; β represents the local feature weight, ranging from [0, 1], which is used to control the influence of local features on the reference vector update; Represents the local features of the input image x(t). Local features include edge features, texture features, etc., which can be extracted through edge detection operators (such as the Sobel operator) or texture analysis methods (such as the gray-level co-occurrence matrix) to capture local detail information such as the edge and texture of the input sample.

[0098] 12) Introducing sparse constraints in the reference vector update phase reduces redundant neurons and reduces computational complexity. The improved reference vector update formula is:

[0099]

[0100] Where min means minimization; W represents the reference vector set of all neurons; S is the training set; represents the best matching unit of the input sample x(t) in the reference vector set of all neurons, where represents the best matching unit of the input sample x(t), that is, the neuron q with the closest Euclidean distance to the sample x(t); The weight parameter representing the sparse constraint. The larger its value, the higher the sparsity requirement for the neuron reference vector. The appropriate value can be determined by experimental methods such as cross-validation. value; represents the neuron reference vector The L1 norm is used to measure the sparsity of the vector, which makes the reference vectors of some neurons approach zero, thereby reducing redundant neurons; N represents the total number of neurons.

[0101] The local feature enhancement mechanism or sparse constrained optimization ensures that the model can accurately identify oil films on the one hand, and improves the operating efficiency of the model on the other hand; during the training and optimization process of the model, the local feature enhancement mechanism can make the reference vector better adapt to the local feature distribution of the input data, while the sparse constrained optimization helps to screen out neurons and features that are more important for oil film detection, so that the optimization direction of the model is more focused on features related to the oil film, thereby improving the optimization efficiency and effect.

[0102] 2) During the convergence phase, the error variable is passed through a constant The attenuation improvement is: the error variable is attenuated through an adaptive mechanism to improve the convergence speed and stability of the model. The formula for updating the error variable is:

[0103]

[0104] Where, represents the error variable of neuron i at time t; γ(t) represents the dynamic attenuation factor,

[0105]

[0106] Where γ0 is the initial decay constant, which ranges from [0 to 1] and can be set based on experimental experience. ΔDataComplexity represents the rate of change of input data complexity, which is obtained by calculating the difference in information entropy of input data at adjacent time steps to reflect the dynamic changes in data complexity. Specifically, the probability distribution of the current image pixel value is calculated at each iteration step to obtain the information entropy. ΔData Complexity is the difference between the information entropy of the current step and the previous step. Represents the attenuation adjustment parameter, which is used to control the dynamic attenuation factor The sensitivity changes with the complexity of the data. This parameter can be adjusted experimentally to achieve the best convergence effect. Figure 5 Shown is an image of the oil film candidate region obtained using the improved growing hierarchical neural gas network.

[0107] Then, S3 is executed to segment the oil film candidate area using an improved multi-scale threshold segmentation algorithm to obtain a final oil film detection result.

[0108] According to an embodiment of the present invention, a multi-scale threshold segmentation algorithm is used to further refine the oil film boundary, eliminate noise and improve detection accuracy. The specific steps are as follows:

[0109] a. Multi-scale threshold calculation: For the oil film candidate area image output by the GHNG model, calculate the local threshold T at different neighborhood scales:

[0110] in, and are the mean and standard deviation of the scale neighborhood, is the sensitivity coefficient. Different sensitivity parameters have different effects. Set the sensitivity parameter as needed and calculate the threshold at different scales through this formula combined with local statistical characteristics.

[0111] b. Threshold fusion: average the thresholds of each scale to generate the final threshold:

[0112] c. Binarization segmentation: According to The image is binarized to accurately extract the oil film boundary and eliminate background noise.

[0113] In this embodiment, the original multi-scale threshold segmentation algorithm is improved as follows.

[0114] 1) Introduce local feature saliency analysis and dynamically adjust the threshold to adapt to the local feature strength. The details are as follows:

[0115]

[0116] Where, Represents the pixel at the kth scale The local threshold of represents the local mean; represents the local standard deviation; is the sensitivity coefficient; Represents pixel points The local feature significance of

[0117]

[0118] in, Represents pixel points The gradient vector of the pixel is calculated by calculating the ratio of the gradient vector norm to the maximum norm to measure the significance of the local feature of the pixel. The larger the value, the more significant the local feature. Max means taking the maximum value.

[0119] 2) The sensitivity coefficient k is dynamically adjusted according to the complexity of local features. The formula is:

[0120]

[0121] Where k0 represents the initial sensitivity coefficient; represents the sensitivity adjustment parameter, which is used to control the degree to which the sensitivity changes with local complexity; Represents pixel points The entropy value of the neighborhood grayscale value is used to measure the local complexity by calculating the entropy value of the neighborhood grayscale. The larger the entropy value, the higher the local complexity. Max Complexity represents the maximum entropy value of all neighborhood grayscale values and is used to normalize the local complexity.

[0122] 3) Introduce spatial weights in multi-scale threshold fusion to consider the spatial distribution of pixels. The formula is:

[0123]

[0124] Where, Represents the final threshold after fusion, which is obtained by the threshold at M scales The threshold value of each scale is obtained by weighted summation of rows. According to the pixel The degree of spatial association with the scale (i.e., spatial weight ) to contribute to the final threshold; M represents the total number of scales; w k Represents the pixel at the kth scale The spatial weight of is defined as:

[0125]

[0126] Where, Indicates the center position of the kth scale, representing the core position of the scale in the image space; Represents pixel points The distance from the center of the kth scale; σ k represents the scale-dependent standard deviation, controlling the rate at which the weight changes with distance, σ k The smaller it is, the faster the weight decays with increasing distance. The exponential term in the numerator represents the pixel For the spatial correlation degree of the k-th scale threshold, the closer the distance, the larger the index value, indicating that the scale has a greater impact on the pixel point. The greater the influence weight of ; in the denominator, the sum of the exponential terms of all M scales is used to normalize the exponential value of the numerator, so that The value range of is between 0 and 1, ensuring that the sum of the weights of each scale is 1, thereby reasonably distributing the contribution of each scale in the fusion.

[0127] The spatial weight is introduced into the multi-scale threshold fusion, so that the fusion process pays more attention to the scale threshold that is spatially close to the current pixel. By allocating the spatial weight, the accuracy of the fused threshold in depicting the oil film boundary is improved. Figure 6 The oil film detection results obtained using the improved multi-scale threshold segmentation algorithm are shown.

[0128] Furthermore, optionally, the image element area threshold can be used to delete small spots and spots in the ship wake area to obtain the final oil film. This process is also a prior art. The final oil film detection result image is shown in the following example: Figure 7 shown.

[0129] Another embodiment of the present invention provides an oil film detection system based on a hierarchical self-organizing network and scale adaptation, such as Figure 8 As shown, the system includes:

[0130] An image preprocessing module 810 is configured to preprocess the acquired raw radar image;

[0131] An oil film region recognition module 820 is configured to use a trained oil film recognition model based on an improved growing hierarchical neural gas network to perform oil film region recognition on the pre-processed radar image to obtain candidate oil film regions;

[0132] The oil film segmentation module 830 is configured to segment the oil film candidate area using an improved multi-scale threshold segmentation algorithm to obtain a final oil film detection result.

[0133] It should be noted that the function of the oil film detection system based on hierarchical self-organizing network and scale adaptation described in this embodiment can be described by the aforementioned oil film detection method based on hierarchical self-organizing network and scale adaptation. For the parts not described in detail in the system embodiment, please refer to the above method embodiment.

[0134] It should be noted that although several units, modules, or submodules are mentioned in the detailed description above, such division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above may be embodied in one module. Conversely, the features and functions of one module described above may be further divided and embodied by multiple modules.

[0135] Furthermore, although the operations of the method of the present invention are described in a particular order in the accompanying drawings, this does not require or imply that these operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0136] Although the spirit and principles of the present invention have been described with reference to several specific embodiments, it should be understood that the present invention is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features of these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. The oil film detection method based on hierarchical self-organizing network and scale adaptation is characterized by: include: Preprocessing the acquired raw radar image; The oil film area of the pre-processed radar image is identified using the trained oil film recognition model based on the improved growing hierarchical neural gas network to obtain the candidate oil film area. The training process of the oil film recognition model based on the improved growth hierarchical neural gas network includes: Initialize neurons, the reference vector of each neuron is a sample point randomly selected from the training set S; the error variable is initialized to zero and the age of the edge is zero; In the growth phase, each iteration randomly selects an input sample point, that is, a neuron, calculates the Euclidean distance between the neuron and all other neurons, and finds the neuron closest to the neuron. and subproximal neurons , increase the age of all connections from neuron q; update the nearest neuron Error variables, neurons and its neighbors’ reference vectors; If neuron q and neuron s are not connected, create a new connection, otherwise reset the connection age to zero; delete connections whose age exceeds the threshold, remove unconnected neurons; and insert new neurons; If the number of neurons in a graph is less than or equal to 2, prune it, otherwise generate its corresponding subgraph; recursively generate a subgraph for each neuron, and use the receptive fields of multiple neurons as the training set for training until the preset maximum layer depth is reached; Calculate the average quantization error of the backup graph and the current graph, and compare the average quantization error. If the stopping condition is met, the growth stops and enters the convergence stage; During the convergence phase, the error variables of all neurons are decayed through an adaptive decay mechanism; Output the clustering results of each layer, extract the oil film candidate area, and the training is completed; The oil film candidate area is segmented using an improved multi-scale threshold segmentation algorithm to obtain a final oil film detection result; the segmentation of the oil film candidate area using the improved multi-scale threshold segmentation algorithm includes: For the oil film candidate area, local thresholds are calculated at different neighborhood scales: ； Where, Indicates the k Pixels at each scale The local threshold of represents the local mean; represents the local standard deviation; is the sensitivity coefficient; Represents pixel points The local feature significance of ； in, Represents pixel points The gradient vector of max Indicates taking the maximum value; The local thresholds at each scale are averaged to generate the final threshold: ； Where, represents the final threshold after fusion; M Indicates the total number of scales; w k Indicates the k Pixels at each scale The spatial weight of According to the final threshold The oil film candidate area is segmented, the oil film boundary is extracted, and the final oil film detection result is obtained.

2. The oil film detection method based on hierarchical self-organizing network and scale adaptation according to claim 1 is characterized in that: The preprocessing includes: performing lateral operator convolution on the original radar image, using the Otsu algorithm to perform vertical noise detection, noise smoothing, speckle removal, and contrast enhancement.

3. The oil film detection method based on hierarchical self-organizing network and scale adaptation according to claim 1 is characterized in that: Update the reference vector according to the following formula : ； Where, Represents neurons i At the moment t The reference vector of η ( t ) represents the learning rate that decreases over time; α Represents the global feature weight, with a value range of [0, 1]; β Represents the local feature weight, with a value range of [0, 1]; Represents a sample point randomly selected from the training set S local features.

4. The oil film detection method based on hierarchical self-organizing network and scale adaptation according to claim 1 is characterized in that: Update the reference vector according to the following formula : ； Where, min represents minimization; W represents the reference vector set of all neurons; S is the training set; Represents the input sample points in the reference vector set of all neurons The best matching unit; The weight parameter representing the sparsity constraint; represents the neuron reference vector The L1 norm of ; N represents the total number of neurons.

5. The oil film detection method based on hierarchical self-organizing network and scale adaptation according to claim 3 or 4, characterized in that: The formula for attenuating the error variables of all neurons through the adaptive attenuation mechanism is: ； Where, Represents neurons i At the moment t The error variable; γ ( t ) represents the dynamic attenuation factor, ； in, is the initial decay constant, ranging from 0 to 1. Δ Data Complexity represents the rate of change of the complexity of the input data, which is obtained by calculating the difference in information entropy of the input data at adjacent time steps. Represents the attenuation adjustment parameter, which is used to control the dynamic attenuation factor Sensitivity to changes in data complexity.

6. The oil film detection method based on hierarchical self-organizing network and scale adaptation according to claim 1 is characterized in that: The sensitivity coefficient in the improved multi-scale threshold segmentation algorithm The calculation formula is: ； Where, k 0 represents the initial sensitivity coefficient; represents the sensitivity adjustment parameter, which is used to control the degree to which the sensitivity changes with local complexity; Represents pixel points The entropy value of the neighborhood grayscale value; MaxComplexit represents the maximum entropy value of all neighborhood grayscale values.

7. The oil film detection method based on hierarchical self-organizing network and scale adaptation according to claim 6 is characterized in that: The improved multi-scale threshold segmentation algorithm k Pixels at each scale Spatial weight w k The calculation formula is: ； Where, Indicates the k The center position of a scale represents the core position of the scale in the image space; Represents pixel points With the k The distance between the centers of the scales; σ k represents the scale-dependent standard deviation.

8. The oil film detection system based on hierarchical self-organizing network and scale adaptation is characterized by: include: An image preprocessing module configured to preprocess the acquired raw radar image; an oil film region recognition module configured to use a trained oil film recognition model based on an improved growing hierarchical neural gas network to perform oil film region recognition on the preprocessed radar image to obtain candidate oil film regions; The training process of the oil film recognition model based on the improved growth hierarchical neural gas network includes: Initialize neurons, the reference vector of each neuron is a sample point randomly selected from the training set S; the error variable is initialized to zero and the age of the edge is zero; In the growth phase, each iteration randomly selects an input sample point, that is, a neuron, calculates the Euclidean distance between the neuron and all other neurons, and finds the neuron closest to the neuron. and subproximal neurons , increase the age of all connections from neuron q; update the nearest neuron Error variables, neurons and its neighbors’ reference vectors; If neuron q and neuron s are not connected, create a new connection, otherwise reset the connection age to zero; delete connections whose age exceeds the threshold, remove unconnected neurons; and insert new neurons; If the number of neurons in a graph is less than or equal to 2, prune it, otherwise generate its corresponding subgraph; recursively generate a subgraph for each neuron, and use the receptive fields of multiple neurons as the training set for training until the preset maximum layer depth is reached; Calculate the average quantization error of the backup graph and the current graph, and compare the average quantization error. If the stopping condition is met, the growth stops and enters the convergence stage; During the convergence phase, the error variables of all neurons are decayed through an adaptive decay mechanism; Output the clustering results of each layer, extract the oil film candidate area, and the training is completed; The oil film segmentation module is configured to segment the oil film candidate area using an improved multi-scale threshold segmentation algorithm to obtain a final oil film detection result; the segmentation of the oil film candidate area using the improved multi-scale threshold segmentation algorithm includes: For the oil film candidate area, local thresholds are calculated at different neighborhood scales: ； Where, Indicates the k Pixels at each scale The local threshold of represents the local mean; represents the local standard deviation; is the sensitivity coefficient; Represents pixel points The local feature significance of ； in, Represents pixel points The gradient vector of max Indicates taking the maximum value; The local thresholds at each scale are averaged to generate the final threshold: ； Where, represents the final threshold after fusion; M Indicates the total number of scales; w k Indicates the k Pixels at each scale The spatial weight of According to the final threshold The oil film candidate area is segmented, the oil film boundary is extracted, and the final oil film detection result is obtained.

Images