Ship detection method and system based on feature fusion and FVIM-HLOA for airborne infrared image

Abstract

The application provides an airborne infrared image ship detection method and system based on feature fusion and FVIM-HLOA, and relates to the technical field of target detection.The technical points of the application include: acquiring an airborne infrared image containing a ship at sea and pre-processing the image;extracting a multi-dimensional feature map from the pre-processed image;using a fuzzy variable index element heuristic algorithm to adaptively cluster the multi-dimensional feature map to generate a region of interest mask;using an improved angle lizard optimization algorithm to obtain an optimal segmentation threshold, and using the optimal segmentation threshold to perform multi-threshold adaptive optimal segmentation on the region of interest mask to obtain a final segmentation map.The application has the advantages of strong feature discrimination, efficient and accurate detection framework, strong optimization algorithm optimization capability, and strong environmental adaptability, and has good generalization ability and practical value.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically to an airborne infrared image ship detection method and system based on feature fusion and FVIM-HLOA. Background Technology

[0002] Infrared thermal imaging technology, with its unique advantages of being less affected by haze, operating day and night, and not relying on visible light illumination, has become a key technology for ship target monitoring, maritime search and rescue, and situational awareness. However, the target features of infrared ship images acquired by airborne platforms often suffer distortion, degradation, decoupling, or inaccuracy in complex marine background environments. Existing identification methods face the following three challenges: First, the dynamic and changing sea surface environment, including wave clutter, specular reflection from the sea surface, atmospheric turbulence, and non-uniformly distributed clouds, all introduce strong background noise and signal attenuation. These interference factors not only blur the target edges and incomplete outlines but also severely weaken the radiation contrast between the target and the background, making the infrared thermal features of the target appear as weak signals with low contrast and low signal-to-noise ratio in the image, easily drowned out by background noise; Second, due to imaging... Due to differences in distance, target structure, and perspective, ship targets exhibit dramatic scale changes, non-rigid deformation, and arbitrary orientation in images. In addition, the dense arrangement and mutual occlusion of multiple targets in scenarios such as ports make it difficult for traditional models that rely on fixed receptive fields or preset anchor frames to effectively cover all aspects, easily leading to missed or false detections. Thirdly, for long-range surveillance or small ship targets, they often occupy only a few to tens of pixels in the image, exhibiting typical "small targets" or "weak targets." These targets lack sufficient spatial resolution to present distinguishable texture, shape, and other structural information, and their thermal radiation characteristics are also relatively weak.

[0003] Therefore, improving the detection accuracy and efficiency of ships at sea in images has long been a research challenge in target detection. Summary of the Invention

[0004] In view of the above problems, this invention proposes an airborne infrared image ship detection method and system based on feature fusion and FVIM-HLOA.

[0005] According to another aspect of the present invention, a method for ship detection in airborne infrared images based on feature fusion and FVIM-HLOA is proposed. The method includes: acquiring an airborne infrared image containing a ship at sea and preprocessing it; extracting multi-dimensional feature maps from the preprocessed image; using a fuzzy variable index meta-heuristic algorithm to adaptively cluster the multi-dimensional feature maps to generate a region of interest mask; using an improved horn lizard optimization algorithm to obtain an optimal segmentation threshold, and using the optimal segmentation threshold to perform multi-threshold adaptive optimal segmentation on the region of interest mask to obtain a final segmentation map.

[0006] Furthermore, the preprocessing includes grayscale conversion.

[0007] Furthermore, the step of extracting multi-dimensional feature maps from the preprocessed image includes: extracting edge features and texture features from the preprocessed image respectively; and fusing the edge features and texture features across modalities to generate multi-dimensional feature maps.

[0008] Furthermore, the generation of the region of interest mask includes: clustering to obtain three types of pixels, namely background, suspected target, and target; removing background pixels to obtain a coarse-grained region of interest mask; and then multiplying the coarse-grained region of interest mask with the preprocessed image by pixels to obtain the region of interest mask.

[0009] Furthermore, the fitness function in the fuzzy variable index metaheuristic algorithm for:

[0010] ;

[0011] In the formula, Represents the set of pixels of class j. variance , For the set of pixels of class j The i-th pixel in Let be the pixel mean of the j-th pixel set.

[0012] Furthermore, obtaining the optimal segmentation threshold using the improved horned lizard optimization algorithm includes:

[0013] In population initialization, use The initial population generated is biased towards higher thresholds, as follows:

[0014] ;

[0015] In the formula, This represents the nth-dimensional threshold of the m-th candidate solution; The numbers are random numbers distributed according to the Beta distribution. These represent the lower and upper bounds of the search space, respectively.

[0016] objective function for:

[0017] ;

[0018] In the formula, t is the current iteration number; Indicates the inter-class variance of Otsu's multi-threshold method; This represents the normalized threshold mean reward term; This represents the threshold spacing penalty term; , These are the corresponding weighting coefficients;

[0019] Generate a random decision variable ξ ~Uniform(0,1). Represents a random variable that follows a uniform distribution;

[0020] If the random decision variable ξ is less than or equal to a preset threshold, then a local development strategy biased towards higher thresholds is selected, and its position update formula is:

[0021] ;

[0022] In the formula, This represents the individual's position at the (t+1)th iteration; This indicates the current optimal position of the individual; This represents the individual's position at the current iteration t; A is the direction control parameter. , Here, T is the convergence factor, T is the maximum number of iterations, and C is the step size control parameter. , The numbers are uniformly distributed random numbers; For directional offset terms, , It is a random number;

[0023] If the random decision variable ξ is greater than a preset threshold, then a global exploration strategy based on Levi's flight is adopted, with the position update formula as follows:

[0024] ;

[0025] In the formula, Indicates adaptive weights, L represents Levi's flight stride. , All are Lévy flight random variables. This indicates that the expression follows a pattern with a mean of 0 and a variance of . The normal distribution This indicates that the distribution follows a normal distribution with a mean of 0 and a variance of 1. The characteristic index of the Lévy distribution is represented. It is a very small constant; This represents the normalized direction vector. .

[0026] Furthermore, the multi-threshold Otsu inter-class variance The calculation formula is:

[0027] ;

[0028] In the formula, k is the number of thresholds; Let be the probability of being clustered in the j-th class. Let be the grayscale mean of the pixel set in the j-th cluster. The global grayscale mean;

[0029] Normalized threshold mean reward item The expression is:

[0030] ;

[0031] In the formula, This represents the j-th threshold; Indicates the weighting coefficient;

[0032] Threshold Spacing Penalty The expression is:

[0033] ;

[0034] In the formula, Indicates the minimum spacing between adjacent thresholds; This represents the minimum spacing threshold.

[0035] According to another aspect of the present invention, an airborne infrared image ship detection system based on feature fusion and FVIM-HLOA is proposed. The system is implemented based on the aforementioned airborne infrared image ship detection method based on feature fusion and FVIM-HLOA. The system includes:

[0036] An image acquisition module configured to acquire and preprocess airborne infrared images containing ships at sea;

[0037] The feature extraction module is configured to extract multi-dimensional feature maps from the preprocessed image;

[0038] The region of interest extraction module is configured to use a fuzzy variable index meta-heuristic algorithm to adaptively cluster multi-dimensional feature maps and generate region of interest masks.

[0039] The segmentation module is configured to obtain the optimal segmentation threshold using an improved horned lizard optimization algorithm, and to perform multi-threshold adaptive optimal segmentation on the region of interest mask using the optimal segmentation threshold to obtain the final segmentation map.

[0040] The beneficial technical effects of this invention are:

[0041] 1) Strong feature discrimination: By fusing Canny edge features and GLCM texture features, a multi-dimensional feature map is constructed, which effectively enhances the contrast between ship targets and complex sea conditions. This solves the problem of traditional methods being highly dependent on single features and easily affected by sea clutter, and significantly improves the distinguishability and detection robustness of targets.

[0042] 2) Highly Efficient and Accurate Detection Framework: A lightweight two-stage framework of "coarse extraction + fine segmentation" is employed. First, the Fuzzy Varying-Index Metaheuristic (FVIM) algorithm is used to adaptively and rapidly cluster the fused feature map, extracting Regions of Interest (ROIs) and significantly compressing the computational scope. Then, an improved Horned Lizard Optimization Algorithm (HLOA) is introduced within the ROIs for multi-threshold adaptive fine segmentation. This framework significantly improves processing efficiency while maintaining detection accuracy, making it suitable for real-time detection needs in resource-constrained scenarios such as airborne platforms.

[0043] 3) Strong optimization algorithm: The improved horned lizard optimization algorithm simulates various defensive behaviors of the horned lizard, introduces an initialization strategy biased towards high thresholds, combines multi-objective fitness functions (combining inter-class variance, threshold mean reward and threshold gap penalty) and simulated annealing acceptance criteria, effectively balances global exploration and local exploitation capabilities, avoids premature convergence of the algorithm, and can adaptively obtain better combinations of segmentation thresholds, which is especially suitable for the accurate extraction of small targets.

[0044] 4) Strong environmental adaptability: The method of this invention can maintain high detection accuracy and low false alarm rate under different sea conditions, different ship types, and complex scenarios such as the presence of land, sea clutter, cloud interference, and sea surface reflection, and has good generalization ability and practical value. Attached Figure Description

[0045] 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, not limitation, in which:

[0046] Figure 1 This is a flowchart of an airborne infrared image ship detection method based on feature fusion and FVIM-HLOA, as described in an embodiment of the present invention.

[0047] Figure 2 This is an example of a preprocessed grayscale image in an embodiment of the present invention;

[0048] Figure 3 This is an example of a Canny edge feature map in an embodiment of the present invention;

[0049] Figure 4 This is an example of the three-class segmentation results of FVIM in an embodiment of the present invention;

[0050] Figure 5This is an example of a coarse-grained region of interest mask in an embodiment of the present invention;

[0051] Figure 6 This is an example of a region of interest mask in an embodiment of the present invention;

[0052] Figure 7 This is an example of the final segmentation image in an embodiment of the present invention;

[0053] Figure 8 This is an example of a fine segmentation result in an embodiment of the present invention. Detailed Implementation

[0054] 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.

[0055] Those skilled in the art will recognize that embodiments of the present invention can be implemented as a system, apparatus, device, method, or computer program product. Therefore, this disclosure can be specifically implemented in the following forms: entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. It should be understood herein that any number of elements in the accompanying drawings is for illustrative purposes only and not as a limitation, and any naming is for distinction only and has no limiting meaning.

[0056] This invention introduces feature fusion technology to fuse edge and texture features across modalities, enhancing the distinction between targets and backgrounds. It employs the Fuzzy Variational Indicator Heuristic (FVIM) algorithm to adaptively cluster the fused feature map, enabling rapid coarse extraction of Regions of Interest (ROIs). Furthermore, it introduces an improved Holo-Lizard Optimization (HLOA) algorithm for multi-threshold adaptive fine segmentation of ROIs, achieving efficient and accurate identification of ship targets. This invention is applicable to detection scenarios involving different sea conditions and ship types.

[0057] This invention proposes an airborne infrared image-based ship detection method based on feature fusion and FVIM-HLOA, such as... Figure 1 As shown, the method includes:

[0058] S1. Acquire airborne infrared images containing ships at sea and preprocess them;

[0059] S2. Extract multi-dimensional feature maps from the preprocessed image;

[0060] S3. Adaptive clustering of multi-dimensional feature maps is performed using a fuzzy variable index meta-heuristic algorithm to generate region of interest masks;

[0061] S4. Use the improved horned lizard optimization algorithm to obtain the optimal segmentation threshold, and use the optimal segmentation threshold to perform multi-threshold adaptive optimal segmentation on the region of interest mask to obtain the final segmentation map.

[0062] First, in S1, airborne infrared images containing ships at sea are acquired and preprocessed.

[0063] According to an embodiment of the present invention, during the image acquisition phase, an infrared thermal imager mounted on a UAV is used to acquire infrared images of the sea surface. As an example, the original image size is 640×512 pixels. The main technical parameters of the infrared camera include: an uncooled vanadium oxide microbolometer sensor, a viewing angle of 57°, a sensor resolution of 160×120, a pixel pitch of 12μm, and a wavelength range of 8~14μm.

[0064] Since the images captured by the infrared camera are color images (such as three-channel JPG format), they need to be converted to grayscale images for subsequent processing. Figure 2 As shown; if the image itself is in grayscale format, then use it directly.

[0065] Then, in S2, multi-dimensional feature maps are extracted from the preprocessed image.

[0066] According to an embodiment of the present invention, edge features and texture features are fused across modalities to enhance the distinction between the target and the background.

[0067] Edge feature extraction employs the Canny operator. The Canny operator is a classic edge detection algorithm that achieves accurate edge extraction through multi-step mathematical processing. The specific implementation steps are as follows:

[0068] 1) Gaussian filtering: using a Gaussian kernel For the original image Perform convolution:

[0069] ;

[0070] in:

[0071] ;

[0072] In the formula, This is the Gaussian standard deviation, used to control the degree of smoothing. The larger the value, the stronger the noise suppression, but the more blurred the edges.

[0073] 2) Calculate gradient magnitude and direction: Use the Sobel operator to calculate the smoothed image. The gradient. Horizontal gradient. for:

[0074] ;

[0075] Vertical gradient for:

[0076] ;

[0077] Then calculate the gradient magnitude (i.e., edge strength):

[0078] ;

[0079] Next, the gradient direction is calculated. (i.e., the direction of the edge):

[0080] ;

[0081] Then, the gradient direction It is quantified into four directions: 0°, 45°, 90°, and 135°.

[0082] 3) Non-maximum suppression (edge ​​thinning): For each pixel According to its gradient direction Compare its gradient magnitude with that of its neighboring pixels. :like If it is the maximum value in the neighborhood, keep it; otherwise, suppress it (set it to 0).

[0083] 4) Dual-threshold detection (edge ​​connectivity): using two thresholds and Edge filtering: gradient magnitude The pixels are identified as strong edges (and directly retained); gradient magnitude Pixels are identified as non-edge (removed); then edge connections are performed, with gradient magnitudes at... Pixels connected to strong edges are considered weak edges (retained); otherwise, they are discarded. The final output binary image is the Canny edge feature map, such as... Figure 3 As shown.

[0084] Texture feature extraction employs the Gray-Level Co-occurrence Matrix (GLCM) method. GLCM is a classic texture feature extraction method that aims to quantify the texture structure and spatial distribution patterns of an image by statistically analyzing the probability of occurrence of gray-level value pairs with specific spatial relationships. The specific implementation steps are as follows:

[0085] 1) Construction of the gray-level co-occurrence matrix: statistically analyzing gray-level pairs in an image. The frequency of occurrence is used to construct a gray-level co-occurrence matrix. :

[0086] ;

[0087] In the formula, The gray-level co-occurrence matrix is ​​given by the direction θ and step size d. For pixels grayscale value; For a given direction and step length The calculated displacement vector; matrix dimension is N is the height of the image (number of rows of pixels); M is the width of the image (number of columns of pixels).

[0088] The statistically obtained gray-level co-occurrence matrix is ​​then converted into a probability distribution form to obtain the probability matrix. :

[0089] ;

[0090] In the formula, The gray-level co-occurrence matrix obtained by summing over all directions and step sizes. = .

[0091] 2) Texture feature extraction: local contrast Quantization measures the grayscale difference between pixel pairs:

[0092] ;

[0093] Further analysis of the correlation between pixel grayscale :

[0094] ;

[0095] In the formula, , They are respectively , Standard deviation; The average gray value in the row direction. The formulas for calculating the average gray level along the column direction are as follows:

[0096] ;

[0097] ;

[0098] Next, calculate the energy. :

[0099] ;

[0100] Then, a weighting factor related to the diagonal position is introduced to measure the local homogeneity of the image texture. :

[0101] ;

[0102] Edge features extracted by the Canny operator and four types of texture features extracted by GLCM are fused across modally (channel dimension concatenation) to generate a multi-dimensional (5-dimensional) feature map. This fused feature effectively integrates the structural and statistical information of the target, enhancing the robustness of the feature representation to complex sea conditions.

[0103] Then, in S3, a fuzzy variable index meta-heuristic algorithm is used to adaptively cluster the multi-dimensional feature map to generate a region of interest mask.

[0104] According to an embodiment of the present invention, the Fuzzy Variational Index Metaheuristic (FVIM) algorithm is used to adaptively cluster the fused feature maps. Each agent represents a candidate solution, i.e., 3 class centers. ,satisfy , .

[0105] To evaluate the quality of candidate solutions (i.e., a set of cluster centers), all pixels need to be classified based on the current cluster centers. For each normalized pixel value in the feature map... Calculate its distance to the three class centers Euclidean distance : Based on the nearest neighbor principle, pixels are... It is assigned to the class (or cluster) corresponding to the class center with the smallest distance.

[0106] The embodiments of the present invention use Represents pixels The obtained class tags have the following value selection rules: Then, all of them were given the same label. The pixels that constitute the first The pixel set of a class, denoted as ,Right now: .

[0107] The formula for agent position update (perturbation strategy) is as follows:

[0108] ;

[0109] In the formula, Let represent the cluster center of the j-th class in the (t+1)-th iteration; This represents the cluster center of the j-th class in the current global optimal solution; It is the intensity of the disturbance; It follows a standard normal distribution.

[0110] To evaluate the quality of candidate solutions (i.e., a set of cluster centers), the sum of within-class variances is used as the fitness function. The smaller the within-class variance, the more compact the pixels within the same class, and the better the clustering effect. The fitness function is defined as:

[0111] ;

[0112] In the formula, Represents the set of pixels of class j. The variance is calculated using the following formula:

[0113] ;

[0114] in, Let j be the number of pixels in the j-th class. The pixel mean of class j: .

[0115] Iterate through all agents (candidate solutions) and calculate the fitness value F for each agent. Select the agent with the smallest fitness value as the global optimal solution; its class center is F. : .

[0116] FVIM three-class segmentation results are as follows: Figure 4 As shown in the diagram, clustering yields three pixel classes: background, suspected target, and target. The region with the highest grayscale value is identified as the target class, the region with the lowest grayscale value as the background class, and the region with a grayscale value between the background and target classes as the suspected target class. Based on the clustering results, the background with the lowest grayscale value is removed, generating a coarse-grained Region of Interest (ROI) mask, as shown in the diagram. Figure 5 As shown; then, the coarse-grained region of interest (ROI) mask is multiplied pixel by pixel with the preprocessed image to obtain the precise region of interest mask, as shown. Figure 6 As shown.

[0117] Then, in S4, the improved horned lizard optimization algorithm is used to obtain the optimal segmentation threshold, and the optimal segmentation threshold is used to perform multi-threshold adaptive optimal segmentation on the region of interest mask to obtain the final segmentation map.

[0118] According to an embodiment of the present invention, the Horned Lizard Optimization Algorithm (HLOA) is a novel metaheuristic algorithm proposed in 2024. Its core idea is to simulate the various defensive behaviors of horned lizards in the natural environment in order to balance the global exploration and local exploitation capabilities in the optimization process.

[0119] Obtaining the optimal segmentation threshold using the improved horned lizard optimization algorithm includes:

[0120] 1) Population initialization: using The initial population is generated with a bias towards higher thresholds:

[0121] ;

[0122] In the formula, This represents the nth-dimensional threshold of the m-th candidate solution; The numbers are random numbers distributed according to the Beta distribution. These represent the lower and upper bounds of the search space, respectively. For example, , ;in, probability density function for:

[0123] ;

[0124] in, It is a random variable; This is the first shape parameter; This is the second shape parameter; This is a Beta function.

[0125] 2) Combining multi-objective optimization functions :

[0126] ;

[0127] In the formula, t is the current iteration number; , For the corresponding weighting coefficients, in this embodiment , ; The formula for calculating the inter-class variance of Otsu with multiple thresholds is as follows:

[0128] ;

[0129] In the formula, k is the number of thresholds; Let be the probability of being clustered in the j-th class. Let be the grayscale mean of the pixel set in the j-th cluster. The global grayscale mean;

[0130] The normalized threshold mean reward term is calculated using the following formula:

[0131] ;

[0132] in, 255 is the j-th threshold; 255 is the normalization factor. This represents the weighting coefficient, used to control the importance of the threshold mean reward term in the fitness function. In this embodiment... =0.3;

[0133] Threshold spacing penalty term:

[0134] ;

[0135] In the formula, Indicates the minimum spacing between adjacent thresholds; This represents the minimum spacing threshold, which is 15 in this embodiment.

[0136] In the objective optimization function, the normalized threshold mean reward term guides the algorithm to select a higher threshold, so that the segmentation result focuses on the ship target area with higher gray values ​​in the infrared image, thereby enhancing the target extraction capability and effectively suppressing the interference of low gray background (such as sea surface and sky); the threshold spacing penalty term penalizes the combination of thresholds that are too close, preventing the oversegmentation problem caused by the threshold spacing being too small, ensuring that each threshold corresponds to a gray range with actual physical meaning, thereby improving the stability of the segmentation result and its robustness to noise.

[0137] 3) Horned Lizard Defensive Behavior Position Update: Generate a random decision variable ξ ~Uniform(0,1). Let ξ represent a random variable that follows a uniform distribution; if the random decision variable ξ is less than or equal to a preset threshold (the preset threshold is a probability of 0.6), then a local development strategy biased towards the higher threshold is selected, i.e., strategy 1; the position update formula is:

[0138] ;

[0139] In the formula, This represents the individual's position at the (t+1)th iteration; This indicates the current optimal position of the individual; Let A represent the individual position at the current iteration t; A is the direction control parameter in the algorithm, and C is the step size control parameter in the algorithm. For the directional offset item:

[0140] ;

[0141] ;

[0142] ;

[0143] In the formula, Let T be the convergence factor and T be the maximum number of iterations. These are random numbers used to introduce random perturbations; These are uniformly distributed random numbers.

[0144] This strategy performs a local fine-grained search around the current optimal solution, adjusting the search step size through parameters A and C, and adaptively shrinking the search range by decreasing the convergence factor 'a' with each iteration. The strategy uses a probability of 0.6 to dominate the algorithm's search, primarily responsible for local refinement to ensure convergence accuracy.

[0145] If the random decision variable ξ is greater than a preset threshold, then a global exploration strategy incorporating Lévy flight is selected, i.e., Strategy 2 - Lévy flight combined with high-value bias; the position update formula is:

[0146] ;

[0147] In the formula, For adaptive weights; L is the Lévy flight step size. Normalized direction vector:

[0148] ;

[0149] ;

[0150] ;

[0151] in, and All are Lévy flight random variables. This indicates that the expression follows a pattern with a mean of 0 and a variance of . The normal distribution This indicates that the distribution follows a normal distribution with a mean of 0 and a variance of 1. The characteristic index of the Lévy distribution, also known as the stability index or tail index, controls the heavy-tailed property of the random step size distribution of Lévy flight. It is a very small constant, in this embodiment .

[0152] This strategy introduces Lévy flight to achieve long-distance jumps. Its heavy-tailed distribution characteristic enables the algorithm to escape local optima, and the adaptive weight w decreases with iteration to adjust the exploration intensity. This strategy is used to assist the search, mainly responsible for global exploration, and avoids the algorithm getting trapped in local optima.

[0153] The two strategies complement each other and adapt parameters to balance the exploration and development capabilities of the algorithm.

[0154] 4) Simulated annealing acceptance criterion: The Metropolis criterion is used to determine whether to accept a new solution.

[0155] ;

[0156] In the formula, ; For temperature parameters; This means that a random number uniformly distributed in the interval [0,1] is less than the value determined by the change in fitness. and temperature parameters Probability of accepting the decision .

[0157] Compared with the previous horned lizard optimization algorithm, the horned lizard optimization algorithm proposed in this invention has the following improvements in segmentation performance: 1) Better initialization strategy: The population is initialized with a Beta distribution biased towards high thresholds, making the initial solution more consistent with the distribution characteristics of high grayscale of ship targets in infrared images, thus improving search efficiency and the quality of the segmentation starting point; 2) More comprehensive fitness function: A combined multi-objective fitness function is constructed, which integrates inter-class variance, threshold mean reward, and threshold gap penalty, effectively avoiding over-segmentation and under-segmentation, and enhancing the distinction between targets and background; 3) More efficient search mechanism: Through the division of labor and cooperation between strategy 1 (local fine search) and strategy 2 (Lévy flight global exploration), combined with the simulated annealing acceptance criterion, the algorithm's ability to escape local optima is significantly improved, achieving better convergence accuracy in multi-threshold segmentation.

[0158] An improved horned lizard optimization algorithm is used to obtain the optimal segmentation threshold. Then, this optimal segmentation threshold is used to perform multi-threshold adaptive optimal segmentation on the region of interest mask to obtain the final segmentation image, as shown below. Figure 7 As shown.

[0159] Furthermore, the largest connected regions in the segmentation results (usually corresponding to land or large fixed interference objects) can be removed, and fine speckle noise can be filtered out through morphological opening and closing operations to obtain a fine segmentation result for the ship target, such as... Figure 8 As shown.

[0160] This invention also proposes an airborne infrared image ship detection system based on feature fusion and FVIM-HLOA. The system is implemented based on the airborne infrared image ship detection method based on feature fusion and FVIM-HLOA described in the above embodiments. The system includes:

[0161] An image acquisition module configured to acquire and preprocess airborne infrared images containing ships at sea;

[0162] The feature extraction module is configured to extract multi-dimensional feature maps from the preprocessed image;

[0163] The region of interest extraction module is configured to use a fuzzy variable index meta-heuristic algorithm to adaptively cluster multi-dimensional feature maps and generate region of interest masks.

[0164] The segmentation module is configured to obtain the optimal segmentation threshold using an improved horned lizard optimization algorithm, and to perform multi-threshold adaptive optimal segmentation on the region of interest mask using the optimal segmentation threshold to obtain the final segmentation map.

[0165] The functionality of the airborne infrared image ship detection system based on feature fusion and FVIM-HLOA described in this embodiment of the invention can be explained by the aforementioned airborne infrared image ship detection method based on feature fusion and FVIM-HLOA. Therefore, for the parts not detailed in the system embodiment, please refer to the above method embodiment, and they will not be repeated here.

[0166] It should be noted that although several units, modules, or sub-modules are mentioned in the detailed description above, this 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 can be embodied in one module. Conversely, the features and functions of one module described above can be further divided and embodied by multiple modules.

[0167] Furthermore, although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0168] 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 method for airborne infrared image ship detection based on feature fusion and FVIM-HLOA, characterized in that, include: Acquire airborne infrared images containing ships at sea and preprocess them; Extract multi-dimensional feature maps from the preprocessed image; A fuzzy variable index meta-heuristic algorithm is used to adaptively cluster multi-dimensional feature maps to generate region of interest masks; An improved horned lizard optimization algorithm is used to obtain the optimal segmentation threshold, and the optimal segmentation threshold is then used to perform multi-threshold adaptive optimal segmentation on the region of interest mask to obtain the final segmentation map; wherein, obtaining the optimal segmentation threshold using the improved horned lizard optimization algorithm includes: In population initialization, use The initial population generated is biased towards higher thresholds, as follows: ； In the formula, This represents the nth-dimensional threshold of the m-th candidate solution; The numbers are random numbers distributed according to the Beta distribution. These represent the lower and upper bounds of the search space, respectively. objective function for: ； In the formula, t is the current iteration number; Indicates the inter-class variance of Otsu's multi-threshold method; This represents the normalized threshold mean reward term; This represents the threshold spacing penalty term; , These are the corresponding weighting coefficients; Generate a random decision variable ξ ~Uniform(0,1). Represents a random variable that follows a uniform distribution; If the random decision variable ξ is less than or equal to a preset threshold, then a local development strategy biased towards higher thresholds is selected, and its position update formula is: ； In the formula, This represents the individual's position at the (t+1)th iteration; This indicates the current optimal position of the individual; This represents the individual's position at the current iteration t; A is the direction control parameter. , Here, T is the convergence factor, T is the maximum number of iterations, and C is the step size control parameter. , The numbers are uniformly distributed random numbers; For directional offset terms, , It is a random number; If the random decision variable ξ is greater than a preset threshold, then a global exploration strategy based on Levi's flight is adopted, with the position update formula as follows: ； In the formula, Indicates adaptive weights, L represents Levy's flight stride. , All are Lévy flight random variables. This indicates that the expression follows a pattern with a mean of 0 and a variance of . The normal distribution This indicates that the distribution follows a normal distribution with a mean of 0 and a variance of 1. The characteristic index of the Lévy distribution is represented. It is a very small constant; This represents the normalized direction vector. .

2. The airborne infrared image ship detection method based on feature fusion and FVIM-HLOA according to claim 1, characterized in that, The preprocessing includes grayscale conversion.

3. The airborne infrared image ship detection method based on feature fusion and FVIM-HLOA according to claim 1, characterized in that, The step of extracting multi-dimensional feature maps from the preprocessed image includes: extracting edge features and texture features from the preprocessed image respectively; and fusing the edge features and texture features across modes to generate multi-dimensional feature maps.

4. The airborne infrared image ship detection method based on feature fusion and FVIM-HLOA according to claim 1, characterized in that, The process of generating the region of interest mask includes: clustering to obtain three types of pixels, namely background, suspected target, and target; removing background pixels to obtain a coarse-grained region of interest mask; and then multiplying the coarse-grained region of interest mask with the preprocessed image pixel by pixel to obtain the region of interest mask.

5. The airborne infrared image ship detection method based on feature fusion and FVIM-HLOA according to claim 1, characterized in that, The fitness function in the fuzzy variable index metaheuristic algorithm for: ； In the formula, Represents the set of pixels of class j. variance , For the set of pixels of class j The i-th pixel in Let be the grayscale mean of the j-th pixel set.

6. The airborne infrared image ship detection method based on feature fusion and FVIM-HLOA according to claim 1, characterized in that, The multi-threshold Otsu inter-class variance The calculation formula is: ； In the formula, k is the number of thresholds; Let be the probability of being clustered in the j-th class. Let be the grayscale mean of the pixel set in the j-th cluster. The global grayscale mean; Normalized threshold mean reward item The expression is: ； In the formula, This represents the j-th threshold; Indicates the weighting coefficient; Threshold Spacing Penalty The expression is: ； In the formula, Indicates the minimum spacing between adjacent thresholds; This represents the minimum spacing threshold.

7. An airborne infrared image ship detection system based on feature fusion and FVIM-HLOA, characterized in that, The system is implemented based on the airborne infrared image ship detection method based on feature fusion and FVIM-HLOA as described in any one of claims 1-6; the system includes: An image acquisition module configured to acquire and preprocess airborne infrared images containing ships at sea; The feature extraction module is configured to extract multi-dimensional feature maps from the preprocessed image; The region of interest extraction module is configured to use a fuzzy variable index meta-heuristic algorithm to adaptively cluster multi-dimensional feature maps and generate region of interest masks. The segmentation module is configured to obtain the optimal segmentation threshold using an improved horned lizard optimization algorithm, and to perform multi-threshold adaptive optimal segmentation on the region of interest mask using the optimal segmentation threshold to obtain the final segmentation map.

Images