Target detection method and system based on improved topological myxin algorithm and neural network

Abstract

The invention provides a target detection method and system based on an improved topological myxin algorithm and a neural network, and relates to the technical field of target detection. The method is characterized by comprising the following steps: extracting saliency and boundary proportion characteristics from original radar image data, and obtaining an oil spill region-of-interest image; obtaining an optimal segmentation threshold value by using an improved topological myxomycete optimization algorithm, and carrying out optimal segmentation on the oil spill region-of-interest image in combination with a neural network to obtain a final segmented image; wherein the improved topological myxomycete optimization algorithm comprises the steps of converting fitness into a pressure index, dynamically updating conductivity based on an evaporation-growth mechanism, calculating information flow among myxomycete individuals in combination with the pressure index and the conductivity, and updating a myxomycete center based on the information flow. The method weakens the influence of wave reflection and mist interference on the segmentation result, improves the precision and robustness of oil spill identification, balances the precision and real-time performance of oil spill segmentation under the dynamic sea condition, and can meet the real-time processing requirements of maritime affair supervision and other scenes.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and specifically to a target detection method and system based on an improved topological slime mold algorithm and neural network. Background Technology

[0002] In recent years, with the increasing frequency of maritime transport and marine resource development, the risk of ship oil spills has significantly increased. Oil spills not only cause huge economic losses but also have long-term and catastrophic impacts on marine ecosystems. Therefore, developing efficient and accurate marine oil spill monitoring technologies to achieve rapid identification and precise delineation of oil spill areas is of vital importance for maritime supervision, emergency response, and ecological protection.

[0003] Synthetic Aperture Radar (SAR) has become a core technology for marine oil spill monitoring due to its all-weather, all-day, and wide-area imaging capabilities. However, SAR oil spill image segmentation faces several technical challenges: First, Bragg scattering caused by sea waves creates complex sea clutter in the image, whose texture features are highly similar to thin oil slicks, leading to false alarms of "oil slick-like" images. Second, the grayscale, texture, and shape features of the oil spill area are affected by multiple factors such as oil slick thickness, sea state, and imaging parameters, exhibiting high non-uniformity and time-varying characteristics, lacking universal and robust feature representation. Third, the boundary between the oil slick and the background (especially the transition area between the oil slick edge and seawater, waves, and clouds) is often blurred, with a large number of mixed pixels, making it difficult for traditional segmentation methods based on fixed thresholds to accurately define the boundary.

[0004] To address these challenges, existing technologies primarily improve upon feature extraction and segmentation algorithms. In feature extraction, traditional methods rely heavily on grayscale, texture (such as GLCM), and polarization features. However, these features have limited discriminative power in complex sea conditions. Regarding image segmentation algorithms, traditional methods such as thresholding, region growing, and active contour models are widely used. Among these, threshold selection methods based on intelligent optimization algorithms, such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO), have attracted attention due to their global search capabilities. However, these algorithms still have significant limitations when processing SAR oil spill images: strong initialization dependency; mismatch between optimization mechanisms and the problem; and lack of consideration for problem constraints. Furthermore, some studies have attempted to introduce neural networks (such as Convolutional Neural Networks (CNNs) for end-to-end segmentation. While these methods perform well on specific datasets, they rely on large amounts of precisely labeled training data, while high-quality SAR oil spill labeled data is scarce and expensive to obtain. Simultaneously, deep models have high computational complexity and poor interpretability, making them difficult to deploy and apply in time-sensitive emergency response scenarios.

[0005] In summary, existing SAR oil spill image segmentation techniques still have significant shortcomings in terms of the adaptability and robustness of feature extraction, as well as the initialization quality, optimization efficiency, and boundary processing accuracy of segmentation algorithms. Summary of the Invention

[0006] In view of the above problems, this invention proposes a target detection method and system based on an improved topological slime mold algorithm and neural network.

[0007] According to one aspect of the present invention, a target detection method based on an improved topological slime mold algorithm and a neural network is proposed. The method includes: S1, extracting saliency and boundary proportion features from the original radar image data to obtain an oil spill region of interest image; S2, using the improved topological slime mold optimization algorithm to obtain the optimal segmentation threshold, and combining it with a neural network to optimize the segmentation of the oil spill region of interest image to obtain a final segmented image; including: S21, initializing the slime mold population and generating initial cluster centers using range constraints and ordered sorting mechanisms; S22, iteratively optimizing the position of individual slime molds, including: calculating the fitness of individual slime molds using the sum of squared errors within the cluster as a fitness function; converting the fitness into a pressure index; and based on evaporation-biological... The long-term mechanism dynamically updates the conductivity; the information flow between slime mold individuals is calculated by combining the pressure index and the conductivity; based on the information flow between slime mold individuals, the slime mold center is updated using information weighting, dynamic perturbation, and range constraints; S23, the slime mold centers after iteration are sorted, and the center corresponding to the slime mold with the best fitness is selected as the optimal cluster center; S24, based on the optimal cluster center, the kernel width parameter of the neural network is adaptively determined and a classification label is assigned to each pixel; the neural network is used to re-predict the category of each pixel to obtain the final segmented image; S3, the final segmented image is post-processed to remove small noise areas in the image to obtain the oil spill detection result image.

[0008] Further, step S1, which involves extracting saliency and boundary proportion features from the original radar image data to obtain the oil spill region of interest image, includes: setting a sliding window; calculating an adaptive binarization threshold based on the grayscale mean and standard deviation of the local window of the image; and using the adaptive binarization threshold to dynamically binarize each window of the original radar grayscale image; calculating the local fractal dimension of the binarized image within the sliding window and performing normalization processing; calculating the global image segmentation threshold based on the normalized fractal dimension and its local statistical features; generating a binary image using the global image segmentation threshold; extracting the largest connected region as the oil spill region of interest; and performing smoothing filtering on the boundary of the oil spill region of interest to obtain the oil spill region of interest image.

[0009] Furthermore, the formula for calculating the adaptive binarization threshold based on the gray-level mean and standard deviation of the local window of the image in step S1 is as follows:

[0010] ;

[0011] in, For adaptive binarization thresholding; , These are the mean and standard deviation of the grayscale values ​​of the original radar image, respectively. , , Empirical weights are used to balance global contrast and local contrast. To avoid taking the minimum value when dividing by zero;

[0012] The formula for calculating the local fractal dimension is as follows:

[0013] ;

[0014] in, L is the number of foreground pixels in the binarized image; L is the perimeter of the sliding window.

[0015] The formula for calculating the global image segmentation threshold is:

[0016] ;

[0017] in, This is the global segmentation threshold for the image, used to separate highly complex regions. The mean of the normalized fractal dimension within the local sliding window; The coefficient of variation is the local variation. These are adjustable weight parameters; Weighting parameters are used to control the coupling features of local texture uniformity and average complexity.

[0018] Furthermore, the formula for generating the initial cluster centers in step S21 is:

[0019] ;

[0020] in, Let m be the initial cluster center of the m-th slime mold; It is a 1×n random vector, where n is the preset number of clusters; , These are constraint parameters; This represents a function used to sort cluster centers in ascending numerical order.

[0021] Furthermore, the formula for calculating the fitness function in step S22 is as follows:

[0022] ;

[0023] in, This represents the fitness value of the m-th slime mold in the t-th iteration; The normalized pixel value in the range [0,1]; Indicates the cluster center index in the range 1-n; The distance from the m-th slime mold in the t-th iteration is the distance to the m-th slime mold. The set of all nearest pixels to each cluster; For the m-th slime mold in the t-th iteration, the... Cluster centers; This is a minor correction.

[0024] Furthermore, the formula for converting fitness into a stress index in step S22 is as follows:

[0025] ;

[0026] in, This represents the pressure of the m-th slime mold in the t-th iteration; is the reciprocal of the maximum fitness in the slime mold population; M is the size of the slime mold population.

[0027] Furthermore, the update formula for conductivity in step S22 is:

[0028] ;

[0029] in, This represents the conductivity between the m-th slime mold and the q-th slime mold in the t-th iteration; , These are the minimum and maximum values ​​of conductivity, respectively. Evaporation rate ; For growth rate, , t represents the decay coefficient, t represents the current iteration number, and max represents taking the maximum value.

[0030] Furthermore, the formula for calculating the information flow between slime mold individuals in step S22 is as follows:

[0031] ;

[0032] in, This represents the information flow between the m-th slime mold and the q-th slime mold in the t-th iteration; This represents the pressure of the m-th slime mold in the t-th iteration; This represents the pressure of the q-th slime mold in the t-th iteration; , Let represent the center vectors of the m-th slime mold and the q-th slime mold, respectively; This represents the distance attenuation coefficient.

[0033] Furthermore, the update formula for the slime mold center in step S22 is:

[0034] ;

[0035] in, Let be the center vector of the m-th slime mold in the (t+1)th iteration; Let m be the weighted average center of the m-th slime mold. ; Let be the random perturbation vector of the m-th slime mold. , For the local search radius, For a k-dimensional random vector, This represents the step size factor.

[0036] According to another aspect of the present invention, a target detection system based on an improved topological slime mold algorithm and a neural network is proposed. The system is used to implement the aforementioned target detection method based on the improved topological slime mold algorithm and neural network. The system includes: a region of interest extraction module configured to extract saliency and boundary proportion features from original radar image data to obtain an oil spill region of interest image; and an optimization segmentation module configured to obtain an optimal segmentation threshold using an improved topological slime mold optimization algorithm and combine it with a neural network to optimize the segmentation of the oil spill region of interest image to obtain a final segmented image; the system includes: initializing a slime mold population by generating initial cluster centers using range constraints and an ordered sorting mechanism; and iteratively optimizing the position of individual slime molds, including: calculating the fitness function using the sum of squared errors within the cluster. Fitness is converted into a pressure index; conductivity is dynamically updated based on the evaporation-growth mechanism; information flow between slime mold individuals is calculated by combining the pressure index and the conductivity; slime mold centers are updated using information weighting, dynamic perturbation, and range constraints based on the information flow between slime mold individuals; the slime mold centers after iteration are sorted, and the center corresponding to the slime mold with the best fitness is selected as the optimal cluster center; based on the optimal cluster center, the kernel width parameter of the neural network is adaptively determined and a classification label is assigned to each pixel; the neural network is used to re-predict the category of each pixel to obtain the final segmented image; an oil spill detection module is configured to post-process the final segmented image to remove small noise areas in the image to obtain the oil spill detection result image.

[0037] The beneficial technical effects of this invention are:

[0038] This invention proposes a target detection method and system based on an improved topological slime mold algorithm and a neural network. First, relying on the strong response characteristics of saliency and boundary ratio features to oil spill areas, the core region of the oil spill target is accurately located. Then, the improved topological slime mold optimization algorithm completes adaptive threshold search and optimization within the oil spill-specific grayscale range. Subsequently, an RBF neural network is introduced to construct a high-dimensional similarity feature space for the optimal cluster centers. Through supervised learning, the nonlinear correlation between pixels and multiple cluster centers is fitted to achieve targeted and accurate segmentation of the oil spill target. Among them, the improved topological slime mold optimization algorithm effectively solves the convergence and stability problem of threshold optimization in complex radiation environment by constructing a dynamic conductivity update mechanism and an oil spill grayscale adaptation strategy, providing reliable centers for the initial clustering of oil spill areas. On this basis, the RBF neural network is introduced, which utilizes its strong mapping ability for nonlinear features to construct a high-dimensional similarity feature space based on the optimal clustering centers output by the topological slime mold algorithm. The hard assignment error of boundary pixels is corrected through supervised learning. Especially for grayscale transition areas caused by wave reflection and fog interference, RBF can integrate the correlation information between pixels and multiple class centers to achieve progressive optimization from "coarse clustering" to "fine classification".

[0039] This invention significantly reduces the impact of wave reflection and fog interference on the segmentation results, effectively overcomes the segmentation bottleneck under complex sea conditions (such as wave clutter and cloud fog interference), greatly improves the accuracy and robustness of oil spill identification, and successfully balances the accuracy and real-time performance of oil spill segmentation under dynamic sea conditions. It can adapt to the real-time processing needs of scenarios such as maritime supervision and marine oil spill emergency response, and provides highly reliable technical support for scenarios such as maritime monitoring and oil spill cleanup scheduling. Attached Figure Description

[0040] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:

[0041] Figure 1 This is a flowchart illustrating a target detection method based on an improved topological slime mold algorithm and neural network, as described in an embodiment of the present invention.

[0042] Figure 2 This is an example of an oil spill area of ​​interest extraction map in an embodiment of the present invention;

[0043] Figure 3 This is an example of an SBR feature map in an embodiment of the present invention;

[0044] Figure 4 This is an example of the final segmented image obtained by combining the improved topological slime mold optimization algorithm with a neural network in an embodiment of the present invention;

[0045] Figure 5This is an example of an oil spill image obtained after post-processing in an embodiment of the present invention. Detailed Implementation

[0046] The principles and spirit of the invention will now be described with reference to several exemplary embodiments. It should be understood that these embodiments are given merely to enable those skilled in the art to better understand and implement the invention, and are not intended to limit the scope of the invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0047] This invention combines saliency and boundary ratio (SBR) features with an improved topological slime mold optimization algorithm. The aim is to propose a radar oil spill image segmentation method based on SBR features, an improved topological slime mold optimization algorithm, and an RBF neural network. This method adapts to the target segmentation requirements under different sea conditions by dynamically adjusting the segmentation threshold.

[0048] This invention proposes a target detection method based on an improved topological slime mold algorithm and a neural network, such as... Figure 1 As shown, the method includes:

[0049] S1. Extract saliency and boundary ratio features from the original radar image data to obtain the oil spill area of ​​interest image;

[0050] S2. Obtain the optimal segmentation threshold using an improved topological slime mold optimization algorithm, and combine it with a neural network to optimize the segmentation of the oil spill region of interest image to obtain the final segmented image; including: S21. Initialize the slime mold population and generate initial cluster centers using a range constraint and ordered sorting mechanism; S22. Iteratively optimize the position of individual slime molds, including: calculating the fitness of individual slime molds using the sum of squared errors within the cluster as a fitness function; converting the fitness into a pressure index; dynamically updating the conductivity based on the evaporation-growth mechanism; calculating the information flow between slime molds based on the pressure index and the conductivity; updating the slime mold centers based on the information flow between slime molds using information weighting, dynamic perturbation, and range constraints; S23. Sort the slime mold centers after iteration and select the center corresponding to the slime mold with the best fitness as the optimal cluster center; S24. Based on the optimal cluster center, adaptively determine the kernel width parameter of the neural network and assign a classification label to each pixel; use the neural network to re-predict the category of each pixel to obtain the final segmented image;

[0051] S3. Post-process the final segmented image to remove small noise areas in the image and obtain the oil spill detection result image.

[0052] The method begins with S1. In S1, saliency and boundary scale features are extracted from the raw radar image data.

[0053] According to an embodiment of the present invention, firstly, a sliding window of size 15×15 is set, and the threshold is dynamically adjusted in combination with the image mean and standard deviation to make the binarization result more adaptable to different conditions. Binarization threshold It is the adaptive threshold for dynamic binarization in the SBR feature extraction process, which transforms the continuous grayscale values ​​of the original radar grayscale image into a binary form of "foreground and background" based on a local sliding window. Its calculation formula is:

[0054] ;

[0055] in, For adaptive binarization thresholding; , These are the mean and standard deviation of the grayscale values ​​of the original radar image, respectively. , , These are empirical weights used to balance global and local contrast; in this embodiment, they are respectively set to values ​​of [values ​​to be filled in]. =0.53, b=0.2, =0.1; To avoid the minimum value obtained when dividing by zero, a value of 10 can be chosen. -6 .

[0056] Then, the self-similarity of the local region is quantified by statistically analyzing the logarithmic ratio of the number of foreground pixels within the local window to the window perimeter. A higher fractal dimension indicates a more complex local texture. The formula for calculating fractal dimension is:

[0057] ;

[0058] ;

[0059] in, For pixels The local fractal dimension reflects the local texture complexity. This represents the number of foreground pixels after binarization. Two-dimensional coordinate index of image pixels; coordinates The grayscale value corresponding to the pixel; Let L be a 15×15 local sliding window centered at (x,y); L is the perimeter of the sliding window.

[0060] Furthermore, the fractal dimension is mapped to a standard range to eliminate the influence of dimensions, facilitating subsequent threshold segmentation and comparison.

[0061] ;

[0062] in, is the normalized fractal dimension, with a value range of [0,1]. , These are the global minimum and maximum fractal dimensions, respectively.

[0063] Then, the global image segmentation threshold is determined based on the normalized fractal dimension:

[0064] ;

[0065] in, The calculated global segmentation threshold for the image is used to separate highly complex regions; Within a local window The mean; The coefficient of variation is the local variation. These are adjustable weight parameters, which are then used to appropriately match the image. In this embodiment, the weight parameter for controlling the coupling feature between local texture uniformity and average complexity is set to 0.2.

[0066] Then, the largest connected region is extracted based on the global image segmentation threshold: the area of ​​each connected region is calculated using the label matrix, and the largest region is retained to eliminate noise and small fragment interference. The extraction formula is:

[0067] ;

[0068] ;

[0069] in, To utilize global image segmentation thresholding The generated binary image; The label matrix represents the connected regions. This is a binary image of the connected region with the largest area. It is a unique label index for connected regions, used to distinguish different connected regions in a binary image. To filter the labels with the largest area .

[0070] Then, a moving average filter is applied to the extracted maximum connected region boundary to eliminate jagged noise and make the boundary smoother.

[0071] By extracting SBR features from the original data image, we can obtain the following: Figure 2 The oil spill region of interest map shown is then smoothed using a 5×5 window to obtain the following result: Figure 3 The SBR feature map shown.

[0072] Then, S2 is executed. In S2, the optimal segmentation threshold is obtained by using the improved topological slime mold optimization algorithm, and the oil spill region of interest image is optimized and segmented by combining the neural network to obtain the final segmented image.

[0073] According to an embodiment of the present invention, the specific steps include: S21, initializing a slime mold population and generating initial cluster centers using a range constraint and ordered sorting mechanism; S22, iteratively optimizing the positions of individual slime molds, including: calculating the fitness of individual slime molds using the sum of squared errors within clusters as a fitness function; converting the fitness into a stress index; dynamically updating the conductivity based on an evaporation-growth mechanism; calculating the information flow between slime molds based on the stress index and the conductivity; updating the slime mold centers based on the information flow between slime molds using information weighting, dynamic perturbation, and range constraints; S23, sorting the slime mold centers after iteration and selecting the center corresponding to the slime mold with the best fitness as the optimal cluster center; S24, adaptively determining the kernel width parameter of the neural network and assigning a classification label to each pixel based on the optimal cluster center; re-predicting the category of each pixel using the neural network to correct the boundary error of classification based on the optimal cluster center; S25, post-processing the re-predicted pixel classification results to remove small noise areas in the image and obtain the final segmented image.

[0074] In S21, the slime mold population is initialized, and an initial cluster center is generated using a range constraint and ordered sorting mechanism: Traditional Topological Slime Mold Model (TSM) relies on manually pre-setting a fixed solution space range during the initialization phase, generating scattered initial individual positions through random numbers. It does not constrain the distribution characteristics or orderliness of the cluster centers, only selecting the optimal individual from the initial population as the starting point for iteration. The formula is as follows: ; ;in, This represents the initial position of the m-th slime mold; , These are the lower and upper bounds of the solution space, which are pre-defined by the user. It is a random number that follows a uniform distribution in [0,1]. This is the initial globally optimal position; This is a general fitness function; Indicates the position of the m-th slime mold; This represents the initial set of locations of slime molds.

[0075] The initial centers in the above process are generated completely randomly, lacking order and range constraints. This can easily lead to over-concentration of centers or centers deviating from the grayscale range of the oil spill, resulting in poor compatibility with the grayscale distribution characteristics of the oil spill image. Excessive initial deviations require a large number of iterations to correct the center position, resulting in slow convergence and a high proportion of invalid searches. Furthermore, the light and dark logic of the cluster centers is not considered, and the disorder of the initial centers increases the complexity of subsequent clustering.

[0076] This invention employs a "range constraint + ordered sorting" mechanism to construct an initial cluster center generation mechanism adapted to the clustering requirements of oil spill images, ensuring uniform initial center distribution and conforming to pixel brightness logic. The formula for the initial cluster center of the m-th slime mold is:

[0077] ;

[0078] in, This is the initial center of the m-th slime mold; It is a 1×n random vector (i.e., elements ∈ [0, 1]), where n is the preset number of clusters; , As constraint parameters, in this embodiment, the values ​​are taken as 0.8 and 0.1 respectively; This is a function that sorts cluster centers in ascending order of their numerical values.

[0079] In this embodiment of the invention, the initial cluster centers are adapted to the grayscale distribution characteristics of the oil spill image by using range constraint parameters c and d, avoiding extreme or sensitive mixing of distributions and significantly reducing initial bias. The ordering function sout(.) ensures that the initial centers conform to the pixel brightness logic, reducing the correction cost of center sorting in subsequent iterations and significantly shortening the convergence time. Uniform and ordered initial cluster centers provide a diverse and high-quality starting point for subsequent iterations, reducing invalid searches and improving iteration efficiency.

[0080] In S22, the position of slime mold individuals is iteratively optimized: the fitness calculation of the traditional topological slime mold algorithm relies only on a single objective function, without considering the specific characteristics of clustering problems. It only maps the solution to a generally comparable value, guiding the iteration with the criterion that "the smaller the value, the better the solution," without quantifying cluster compactness and intra-cluster consistency. Its fitness calculation formula is: ;in, Let be the position of the slime mold individual in the t-th iteration; is the globally optimal fitness value; h is the sensitivity coefficient, controlling the steepness of the fitness curve. Traditional topological slime mold algorithms suffer from the core requirement of clustering unrelated oil spill image pixels during iterative optimization, resulting in poor adaptability and difficulty in accurately measuring clustering quality. This leads to ambiguous iteration directions and a tendency to get trapped in local optima.

[0081] This invention employs the sum of squared intra-cluster errors to quantify clustering quality, directly linking fitness to the compactness of pixel clusters in oil spill images. The optimized fitness calculation directly aligns with the clustering requirements of oil spill images, accurately quantifying intra-cluster compactness and providing a clear objective for minimizing intra-cluster errors during iteration. This avoids iteration direction deviations caused by generic fitness and improves the initial accuracy of cluster centers. The fitness formula for the m-th slime mold in the t-th iteration is:

[0082] ;

[0083] in, The normalized pixel value in the range [0,1]; Indicates a cluster center index in the range 1-n; The distance from the m-th slime mold in the t-th iteration is the distance to the m-th slime mold. The set of all nearest pixels to each cluster; For the m-th slime mold in the t-th iteration, the... Cluster centers; The square of the difference between the pixel and the center is used to amplify the degree of dispersion within the cluster; This is a minor correction.

[0084] Traditional topological slime mold algorithms lack a pressure transformation mechanism, directly using fitness values ​​as the basis for individual information interaction. This lack of standardization of individual quality within the population leads to significant differences in fitness values ​​among individuals, making individual quality differences difficult to visualize. Furthermore, it can result in a few individuals with extreme fitness values ​​dominating the iteration process, ultimately causing a loss of population diversity. This invention transforms fitness into a pressure index, standardizing the population scale. Higher fitness corresponds to higher pressure, thus achieving a standardized measurement of individual quality. Individual quality differences become intuitive and quantifiable, preventing a few individuals with extreme fitness from dominating the iteration, ensuring population diversity, and laying the foundation for subsequent multi-factor coupled information transmission. The pressure formula for the m-th slime mold in the t-th iteration is:

[0085] ;

[0086] in, is the reciprocal of the maximum fitness in the population; M is the size of the slime mold population.

[0087] Traditional topological slime mold algorithms directly use connection weights to achieve information transfer between individuals based on fitness values, combining only spatial distance and fitness as a single dimension, without integrating individual quality differences and connection strength. Its connection weights are updated as follows: ;in, Let be the connection weight between slime mold i and slime mold j in the t-th iteration. A value close to 1 indicates a connection core, and a value close to 0 indicates a connection redundancy. This is the distance sensitivity adjustment coefficient; the larger the value, the more obvious the difference in weight between the distance and the weight. Let be the Euclidean distance between slime bacterium i and slime bacterium j in the solution space during the t-th iteration, which is used to measure the compactness of the space; Let be the average Euclidean distance between all slime mold individuals in the t-th iteration; Let be the globally optimal slime mold position for the entire population in the t-th iteration; , Let i and j be the positions of slime mold i and slime mold j respectively in the t-th iteration.

[0088] The connection weight update step in the traditional topological slime mold algorithm suffers from a simplistic simultaneous information transmission logic, failing to consider the coupling relationships between individual mass differences (pressure differences), spatial distance attenuation, and conductivity. This leads to the transmission of invalid information by long-distance or low-quality slime molds, reducing iteration efficiency. Subsequently, a fixed inertia weight is used to update the conductivity. However, the conductivity initialization depends on the sensitivity coefficient and maximum distance, which does not adapt to the needs of different iteration stages. ; ;in, Let be the conductivity of slime mold i and slime mold j in the t-th iteration; This is a sensitivity adjustment coefficient, with a value of 2-5, used to control the distinguishability of connection strength; Initialize the maximum Euclidean distance within the space for the slime mold population; Let be the initial Euclidean distance between slime mold i and slime mold j in the solution space; This is an inertial weight, with a fixed value of 0.6-0.9, used to balance the influence of historical transmission rate and current cooperation status; Let be the optimal fitness value of the entire slime mold population in the t-th iteration. The processing in the traditional topological slime mold algorithm mentioned above results in fixed inertial weights and a lack of flexibility in adjusting the conductance, making it unable to adapt to the iterative pattern of "initial exploration and later convergence," and easily leading to problems such as insufficient initial exploration or slow convergence in the later stages.

[0089] Therefore, this embodiment of the invention dynamically adjusts conductivity through an "evaporation-growth mechanism," strengthening high-flow-rate connections and weakening low-flow-rate connections through evaporation. Conductivity adjustment adapts to the needs of the entire iteration cycle; initially, low evaporation and high growth ensure the exploration scope; later, high evaporation and low growth quickly focus on high-quality connections, avoiding extreme cases of excessively high or low conductivity, accelerating convergence speed, and improving the stability of threshold optimization. The improved conductivity update formula for slime molds is:

[0090] ;

[0091] in, This represents the conductivity between the m-th slime mold and the q-th slime mold in the t-th iteration; a larger value indicates a tighter connection. , These are the minimum and maximum values ​​of conductivity, respectively. Evaporation rate ; For growth rate, , The decay coefficient is closer to 1, and the change is more gradual. t represents the current iteration number; max represents taking the maximum value.

[0092] This invention further incorporates three factors—quality difference (pressure difference), spatial distance, and conductivity—to construct a flow formula, prioritizing the transmission of information from high-quality neighbors. The optimized flow calculation better aligns with actual clustering needs, effectively filtering out invalid information transmission from distant or low-quality slime molds, strengthening information interaction among high-quality neighbors, improving the targeting and efficiency of group iteration, and avoiding resource waste. Combining pressure indicators and conductivity, the information flow between slime molds m and n is:

[0093] ;

[0094] in, This represents the information flow between the m-th slime mold and the q-th slime mold in the t-th iteration; This represents the pressure of the m-th slime mold in the t-th iteration; This represents the pressure of the q-th slime mold in the t-th iteration; , Let represent the center vectors of the m-th slime mold and the q-th slime mold, respectively; Indicates the distance attenuation coefficient; It is the distance attenuation factor, which decreases as the Euclidean distance increases.

[0095] This invention further utilizes information flow between slime mold individuals, employing information weighting, dynamic perturbation, and range constraints to update slime mold centers. Traditional topological slime mold algorithms rely solely on randomly generated initial centers for center updates, combining the global optimal position, neighbor conductance, and a fixed step size factor during updates, without constraining the weight, distribution range, and logical order of high-quality information within the centers. Furthermore, the update process does not assign higher weights to the center information of high-quality slime molds (high pressure, low fitness), making the optimization direction susceptible to interference from inferior information and resulting in low center update accuracy. The lack of range constraints and logical ordering means that centers may exceed the pixel grayscale normalization range or violate the light-dark distribution logic of oil spill images from gray to dark, increasing the difficulty of subsequent clustering. Therefore, relying solely on fixed step sizes and random weight adjustments lacks a dynamic perturbation mechanism, resulting in a weak ability to balance exploration and convergence, and a tendency for center oscillations or getting trapped in local optima.

[0096] This invention constructs a slime mold center update mechanism through a combination of high-quality information weighting, dynamic perturbation, and logical constraints. This ensures that the slime mold center update aligns with both high-quality population information and the characteristics of oil spill images. It optimizes the slime mold center position by combining population information, random exploration, and constraints. Weighted averaging increases the center weight of high-quality slime molds, ensuring the correct optimization direction. Sorting and range constraints ensure that the center always conforms to pixel brightness logic. Dynamic perturbation and constraints significantly improve stability. The slime mold center update formula is:

[0097] ;

[0098] in, Let be the center vector of the m-th slime mold in the (t+1)th iteration; Let m be the weighted average center of the m-th slime mold. ; Let be the random perturbation vector of the m-th slime mold. , For the local search radius, For a k-dimensional random vector, This represents the step size factor.

[0099] In S23, the slime mold centers after the iteration are sorted, and the center corresponding to the slime mold with the best fitness is selected as the optimal cluster center: the center vector of slime mold, that is, the cluster center of image pixels, is sorted in ascending order from gray to dark according to the brightness logic of image pixels, so as to further sort the slime mold centers; the center with the smallest fitness is selected from M slime molds as the final result, and the optimal is selected from the population to reduce the risk of local optima.

[0100] ;

[0101] in, The optimal cluster center output by TSM; Let represent the fitness value of the m-th slime mold in the T-th iteration, where T is the maximum number of iterations.

[0102] In S24, based on the optimal cluster center, the kernel width parameter of the neural network is adaptively determined and a classification label is assigned to each pixel. The neural network is then used to re-predict the category of each pixel to correct the boundary error of classification based on the optimal cluster center. Specifically, the RBF neural network first 1) adaptively determines the kernel width parameter based on the optimal cluster center output by the improved TSM algorithm. By analyzing the average distance and distribution density of the cluster centers, the kernel width is made to accurately match the pixel grayscale distribution pattern of the oil spill image, avoiding feature measurement deviation caused by a fixed kernel width. Subsequently, 2) the network constructs a hidden layer structure to map the features of each pixel to a high-dimensional space, fully preserving the multi-dimensional correlation information between pixels and various cluster centers, especially strengthening the feature differences in grayscale transition areas such as wave reflection and fog interference, laying a solid foundation for subsequent accurate classification. This process makes the pixel similarity measurement more in line with the actual scene, greatly improving the feature differentiation ability between oil spills and the background under complex sea conditions, and providing sufficient basis for boundary correction. Based on the improved initial clustering results of the TSM algorithm, the RBF neural network assigns a clear classification label to each pixel and constructs a supervised target matrix to ensure that the network training direction is highly consistent with the initial clustering logic, avoiding deviation from the core requirements of oil spill segmentation during the training process. This supervised mode, which seamlessly integrates with the TSM results, inherits the accuracy of TSM in cluster center localization and provides a clear target orientation for network optimization, effectively avoiding the classification confusion that may occur in unsupervised training. Ultimately, a logical closed loop of "coarse clustering → fine classification" is achieved, ensuring the consistency between the segmentation results and the initial clustering. Finally, 3) after solving the output layer weights, the RBF neural network re-predicts the category of each pixel based on the mapping results of the high-dimensional feature space, focusing on correcting the "hard assignment" error of boundary pixels in the initial TSM clustering. For the blurred boundary areas under complex sea conditions, the network integrates the correlation information between pixels and multiple class centers to accurately distinguish between oil spill and background pixels, solving the problems of "jagged" boundary misclassification or over-smoothing in traditional methods, and effectively filtering dynamic interference such as sea clutter and cloud obstruction. Ultimately, this not only makes the boundaries of the oil spill area more closely resemble the actual shape, significantly reducing the rate of missed and incorrect classification, but also allows the classification results to accurately meet the actual needs of scenarios such as maritime supervision and oil spill cleanup scheduling.

[0103] Among them, the kernel width parameter is adaptively determined based on the optimal cluster center output by the improved TSM algorithm. as follows.

[0104] ;

[0105] ;

[0106] in, The average distance to the optimal center of TSM; Let be the Euclidean distance between the a-th and b-th centers.

[0107] Based on kernel width parameter The hidden layer structure is constructed by mapping each pixel feature to a high-dimensional space as follows: Multi-dimensional information is preserved through feature mapping, and the allocation error of boundary pixels is corrected. The (e,f)th element in the output hidden layer matrix G is:

[0108] ;

[0109] in, The closer the value is to 1, the more similar the pixel is to the center; It is an exponential function; This is the value of the e-th normalized pixel; Let f represent the f-th cluster center.

[0110] Using the initial classification of TSM as the supervision objective, RBF is provided with training direction. Its supervision signal is seamlessly integrated with the TSM results, ensuring continuity between optimization and previous clustering. A supervision objective matrix is ​​constructed. Each element in is:

[0111] ;

[0112] ;

[0113] in, To supervise the target matrix The (u,v)th element in the dataset (u is the pixel index, v is the category index); Let be the initial category label for the u-th element; Let be the feature vector of the u-th pixel; The v-th cluster center is output by TSM.

[0114] A pseudo-inverse weight calculation is employed to predict pixel categories based on high-dimensional features, correcting the initial classification error of TSM, especially for boundary pixels. An RBF neural network is used to progressively improve classification accuracy from coarse clustering to fine classification. The RBF output layer weights are calculated as follows: , Hidden layer matrix The pseudo-inverse; then the final optimized classification result is:

[0115] ;

[0116] in, This represents the row vector corresponding to the u-th pixel in the hidden layer matrix G; This represents the column vector corresponding to the v-th class in the output layer weight matrix O; This represents the output score of the RBF neural network for the class v of the u-th pixel.

[0117] Step S2 uses the improved topological slime mold optimization algorithm to perform initial clustering on the oil spill region of interest obtained from SBR feature extraction. This initial clustering is achieved through ordered constraint initial center generation, multi-factor coupled iterative optimization, and range and logical constraint center updates, accurately locating the core cluster centers of the oil spill and background, thus obtaining preliminary oil spill region segmentation results. Based on this, an RBF neural network is introduced for fine optimization: first, the kernel width is adaptively determined based on the optimal cluster centers output by the improved TSM algorithm, ensuring that the feature measurement fits the grayscale distribution of the oil spill image; then, high-dimensional feature mapping is used to enhance the feature differences in grayscale transition areas, and supervised training seamlessly integrated with the TSM results is used to calibrate the classification direction; finally, through weight solving and classification optimization, the "hard allocation" error of boundary pixels in the initial clustering is corrected, accurately distinguishing oil spill and background pixels. Through the synergistic effect of the improved topological slime mold optimization algorithm's clustering localization and the RBF neural network's fine correction, the oil spill region is further accurately extracted, resulting in the following... Figure 4 The final refined segmentation result of the oil spill area is shown.

[0118] Then, S3 is executed. In S3, the final segmented image is post-processed to remove small noise areas in the image, and the oil spill detection result image is obtained.

[0119] According to embodiments of the present invention, for Figure 4 Post-processing was performed to remove small speckle noise regions with an area less than 50, resulting in... Figure 5 Final oil spill detection results image.

[0120] This invention also proposes a target detection system based on an improved topological slime mold algorithm and neural network. The system is used to implement the target detection method based on the improved topological slime mold algorithm and neural network described in the above embodiments. The system includes:

[0121] The region of interest extraction module is configured to extract saliency and boundary ratio features from the original radar image data to obtain the oil spill region of interest image.

[0122] The optimized segmentation module is configured to obtain the optimal segmentation threshold using an improved topological slime mold optimization algorithm and combine it with a neural network to optimize the segmentation of the oil spill region of interest image to obtain the final segmented image. This includes: initializing the slime mold population and generating initial cluster centers using a range constraint and ordered sorting mechanism; iteratively optimizing the positions of individual slime molds, including: calculating the fitness of individual slime molds using the sum of squared errors within clusters as a fitness function; converting the fitness into a pressure index; dynamically updating the conductivity based on an evaporation-growth mechanism; calculating the information flow between slime molds using the pressure index and the conductivity; updating the slime mold centers based on the information flow between slime molds using information weighting, dynamic perturbation, and range constraints; sorting the slime mold centers after iteration and selecting the center corresponding to the slime mold with the best fitness as the optimal cluster center; adaptively determining the kernel width parameter of the neural network and assigning a classification label to each pixel based on the optimal cluster center; and re-predicting the category of each pixel using the neural network to obtain the final segmented image.

[0123] The oil spill detection module is configured to post-process the final segmented image to remove small noise areas in the image and obtain the oil spill detection result image.

[0124] The function of the target detection system based on the improved topological slime mold algorithm and neural network described in this embodiment of the invention can be explained by the aforementioned target detection method based on the improved topological slime mold algorithm and neural network. Therefore, for the parts not described in detail in the system embodiment, please refer to the above method embodiment, and they will not be repeated here.

[0125] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A target detection method based on an improved topological slime mold algorithm and neural network, characterized in that, include: S1. Extract saliency and boundary ratio features from the original radar image data to obtain the oil spill area of ​​interest image; S2. The optimal segmentation threshold is obtained by using the improved topological slime mold optimization algorithm, and the oil spill region of interest image is optimized and segmented by combining the neural network to obtain the final segmented image; include: S21. Initialize the slime mold population and generate initial cluster centers using a range constraint and ordered sorting mechanism; S22. Iteratively optimize the location of slime mold individuals, including: calculating the fitness of slime mold individuals using the sum of squared errors within the cluster as a fitness function; converting the fitness into a pressure index; dynamically updating the conductivity based on the evaporation-growth mechanism; calculating the information flow between slime mold individuals by combining the pressure index and the conductivity; and updating the slime mold center based on the information flow between slime mold individuals using information weighting, dynamic perturbation, and range constraints. S23. Sort the slime mold centers after the iteration and select the center corresponding to the slime mold with the best fitness as the optimal cluster center. S24. Based on the optimal clustering center, adaptively determine the kernel width parameter of the neural network and assign a classification label to each pixel; use the neural network to re-predict the category of each pixel, and then obtain the final segmented image; S3. Post-process the final segmented image to remove small noise areas in the image and obtain the oil spill detection result image.

2. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 1, characterized in that, Step S1, which involves extracting saliency and boundary proportion features from the original radar image data to obtain the oil spill region of interest image, includes: setting a sliding window; calculating an adaptive binarization threshold based on the grayscale mean and standard deviation of the local window of the image; and using the adaptive binarization threshold to dynamically binarize each window of the original radar grayscale image; calculating the local fractal dimension of the binarized image within the sliding window and performing normalization processing; calculating the global image segmentation threshold based on the normalized fractal dimension and its local statistical features; generating a binary image using the global image segmentation threshold; extracting the largest connected region as the oil spill region of interest; and performing smoothing filtering on the boundary of the oil spill region of interest to obtain the oil spill region of interest image.

3. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 2, characterized in that, The formula for calculating the adaptive binarization threshold based on the gray-level mean and standard deviation of the local window of the image in step S1 is as follows: ； in, For adaptive binarization thresholding; , These are the mean and standard deviation of the grayscale values ​​of the original radar image, respectively. , , Empirical weights are used to balance global contrast and local contrast. To avoid taking the minimum value when dividing by zero; The formula for calculating the local fractal dimension is as follows: ； in, L is the number of foreground pixels in the binarized image; L is the perimeter of the sliding window. The formula for calculating the global image segmentation threshold is: ； in, This is the global segmentation threshold for the image, used to separate highly complex regions. The mean of the normalized fractal dimension within the local sliding window; The coefficient of variation is the local variation. These are adjustable weight parameters; Weighting parameters are used to control the coupling features of local texture uniformity and average complexity.

4. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 1, characterized in that, The formula for generating the initial cluster centers in step S21 is: ； in, Let m be the initial cluster center of the m-th slime mold; It is a 1×n random vector, where n is the preset number of clusters; , These are constraint parameters; This represents a function used to sort cluster centers in ascending numerical order.

5. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 4, characterized in that, The formula for calculating the fitness function in step S22 is as follows: ； in, This represents the fitness value of the m-th slime mold in the t-th iteration; The normalized pixel value in the range [0,1]; Indicates the cluster center index in the range 1-n; The distance from the m-th slime mold in the t-th iteration is the distance to the m-th slime mold. The set of all nearest pixels to each cluster; For the m-th slime mold in the t-th iteration, the... Cluster centers; This is a minor correction.

6. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 5, characterized in that, The formula for converting fitness into a stress index in step S22 is as follows: ； in, This represents the pressure of the m-th slime mold in the t-th iteration; is the reciprocal of the maximum fitness in the slime mold population; M is the size of the slime mold population.

7. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 6, characterized in that, The formula for updating the conductivity in step S22 is: ； in, This represents the conductivity between the m-th slime mold and the q-th slime mold in the t-th iteration; , These are the minimum and maximum values ​​of conductivity, respectively. Evaporation rate ; For growth rate, , t represents the decay coefficient, t represents the current iteration number, and max represents taking the maximum value.

8. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 7, characterized in that, The formula for calculating the information flow between slime mold individuals in step S22 is as follows: ； in, This represents the information flow between the m-th slime mold and the q-th slime mold in the t-th iteration; This represents the pressure of the m-th slime mold in the t-th iteration; This represents the pressure of the q-th slime mold in the t-th iteration; , Let represent the center vectors of the m-th slime mold and the q-th slime mold, respectively; This represents the distance attenuation coefficient.

9. The target detection method based on the improved topological slime mold algorithm and neural network according to claim 8, characterized in that, The update formula for the slime mold center in step S22 is: ； in, Let be the center vector of the m-th slime mold in the (t+1)th iteration; Let m be the weighted average center of the m-th slime mold. ; Let be the random perturbation vector of the m-th slime mold. , For the local search radius, For a k-dimensional random vector, This represents the step size factor.

10. A target detection system based on an improved topological slime mold algorithm and neural network, characterized in that, The system is used to implement the target detection method based on the improved topological slime mold algorithm and neural network as described in any one of claims 1-9; the system includes: The region of interest extraction module is configured to extract saliency and boundary ratio features from the original radar image data to obtain the oil spill region of interest image. The optimized segmentation module is configured to obtain the optimal segmentation threshold using an improved topological slime mold optimization algorithm and combine it with a neural network to optimize the segmentation of the oil spill region of interest image to obtain the final segmented image. This includes: initializing the slime mold population and generating initial cluster centers using a range constraint and ordered sorting mechanism; iteratively optimizing the positions of individual slime molds, including: calculating the fitness of individual slime molds using the sum of squared errors within clusters as a fitness function; converting the fitness into a pressure index; dynamically updating the conductivity based on an evaporation-growth mechanism; calculating the information flow between slime molds using the pressure index and the conductivity; updating the slime mold centers based on the information flow between slime molds using information weighting, dynamic perturbation, and range constraints; sorting the slime mold centers after iteration and selecting the center corresponding to the slime mold with the best fitness as the optimal cluster center; adaptively determining the kernel width parameter of the neural network and assigning a classification label to each pixel based on the optimal cluster center; and re-predicting the category of each pixel using the neural network to obtain the final segmented image. The oil spill detection module is configured to post-process the final segmented image to remove small noise areas in the image and obtain the oil spill detection result image.

Images