A target detection method and system based on damped newton method optimization clustering

Abstract

The application provides a target detection method and system based on a damping Newton method optimized clustering, and relates to the technical field of target detection. The method comprises the following steps: acquiring an image containing a coastline and a sea target; performing coastline extraction and dynamic mask generation processing on the image to obtain an image of a region of interest containing only the sea target; performing two-class clustering on the target and the background in the image of the region of interest by using a K-means clustering algorithm optimized by a damping Newton method to obtain a classification result image of the sea target and the background; and performing binaryzation processing on the classification result image to obtain a final target detection result. The application significantly improves the accuracy and anti-interference ability of target detection, and can realize high-precision separation of small targets and seawater backgrounds.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a target detection method and system based on damped Newton's method for optimized clustering. Background Technology

[0002] Drone aerial photography or satellite remote sensing images are widely used in marine monitoring. However, nearshore areas, due to their complex backgrounds including land, beaches, and docks, severely interfere with the detection of small targets at sea (such as small fishing boats and buoys). Existing technologies have the following problems: 1) Coastline interference: Traditional global thresholding or edge detection methods struggle to distinguish between coastlines and small targets, easily leading to target fragmentation or false detections of land areas; 2) Weak features of small targets: Small boats occupy a small proportion of the image, their grayscale is close to that of seawater, and they are affected by lighting and waves. Deep learning methods rely on a large amount of labeled data, resulting in poor generalization; 3) Insufficient clustering robustness: Classic clustering algorithms such as K-means are sensitive to initial centers and are prone to getting trapped in local optima when grayscale distributions overlap, leading to classification errors. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a target detection method and system based on damped Newton's method optimized clustering.

[0004] According to one aspect of the present invention, a target detection method based on damped Newton's method optimized clustering is proposed, the method comprising:

[0005] S1. Acquire an image containing the coastline and maritime targets;

[0006] S2. Perform coastline extraction and dynamic mask generation processing on the image to obtain an image of the region of interest containing only maritime targets.

[0007] S3. The K-means clustering algorithm optimized by the damped Newton method is used to perform binary clustering of targets and background in the region of interest image to obtain the classification result map of maritime targets and background;

[0008] S4. Binarize the classification result image to obtain the final target detection result.

[0009] Further, S2 includes: converting the image to grayscale and performing Gaussian filtering for noise reduction; applying adaptive histogram equalization to the denoised image for contrast enhancement; then, performing edge detection on the image using the Canny edge detection operator; subsequently, performing morphological operations on the Canny edge detection image to connect breakpoints and remove noise; next, extracting all outer contours, selecting the longest continuous contour as the main coastline, and constructing a land-sea segmentation mask; and multiplying the land-sea segmentation mask with the denoised image to obtain the region of interest image.

[0010] Furthermore, the K-means clustering algorithm optimized by the damped Newton method described in S3 for binary classification of targets and backgrounds in the region of interest image includes: setting the pixel set of the region of interest image into two classes, target and background, and initializing two cluster centers; then, in each iteration, calculating the distance from each pixel sample to the current cluster center and assigning a label, then obtaining the objective function, and calculating the gradient and Hessian matrix of each cluster center based on the objective function, calculating the Newton direction based on the gradient and Hessian matrix, updating the cluster centers along the Newton direction, and adaptively adjusting the search step size according to the decrease of the objective function; stopping the iteration when the change in the cluster centers is less than the preset tolerance or the maximum number of iterations is reached, thus completing the binary classification of target and background.

[0011] Furthermore, the initialization of the two cluster centers described in S3 includes:

[0012] ;

[0013] ;

[0014] in, Represents the initial cluster centers of the target; The initial cluster centers represent the background; D represents the dataset consisting of the gray values ​​of all pixels in the region of interest image. , Offset:

[0015] ;

[0016] in, This represents the weight of the i-th pixel sample point. This represents the grayscale value of the i-th pixel sample point. This represents the reference grayscale baseline for the target in dataset D. The variance represents the low grayscale range; N represents the total number of image pixel samples. The threshold value for the low grayscale quantiles in dataset D; Represents the global grayscale value of dataset D; This is the scaling factor.

[0017] Furthermore, the objective function described in S3 is:

[0018] ;

[0019] in, Indicates the cluster center. Corresponding to the target, Corresponding background; Indicates the first Pixel-like sample set, This represents the clustering label of the i-th pixel sample point; This represents the average density of the i-th pixel sample point in the k-th class. This represents the grayscale value of the j-th pixel sample point. This represents the bandwidth of the Gaussian kernel function; This represents the average density of the k-th class of pixel samples. For the first Number of pixel samples; This represents the scaling factor that controls density sensitivity.

[0020] Furthermore, the adaptive adjustment of the search step size based on the descent magnitude of the objective function described in S3 includes:

[0021] If the objective function value in the current iteration is less than the objective function value in the previous iteration, or the search step size is less than a fixed threshold, then accept the current step size and terminate the search; otherwise, update the search step size according to the following formula:

[0022] ;

[0023] in, This represents the step size in the (t+1)th iteration; This represents the step size in the t-th iteration; Indicates the basic attenuation coefficient; This represents the change in the objective function. This represents the k-th cluster center in the t-th iteration. Indicates the Newton search direction for the k-th cluster center; It represents the variance of the change in the objective function value.

[0024] Furthermore, the calculation of the gradient and Hessian matrix of each cluster center based on the objective function, the calculation of the Newton direction based on the gradient and Hessian matrix, and the updating of the cluster centers along the Newton direction, as described in S3, include:

[0025] Calculate the gradient of the objective function with respect to each cluster center. :

[0026] ;

[0027] And the Hessian matrix for each cluster center :

[0028] ;

[0029] Based on gradient Hessian matrix Calculate the Newton search direction for the k-th cluster center. ;

[0030] Update cluster centers along the Newtonian direction:

[0031] ;

[0032] in, For the first During the nth iteration Cluster centers of pixel-like samples For the first At the +1st iteration, the... Cluster centers of pixel-like samples.

[0033] According to another aspect of the present invention, a target detection system based on damped Newton's method optimized clustering is proposed. The system is used to implement the aforementioned target detection method based on damped Newton's method optimized clustering; the system includes:

[0034] The image acquisition module is used to acquire images containing coastlines and maritime targets.

[0035] The region of interest extraction module is used to perform coastline extraction and dynamic mask generation processing on the image to obtain a region of interest image containing only maritime targets.

[0036] The binary classification module is used to perform binary classification and clustering of targets and background in the region of interest image using the K-means clustering algorithm optimized by the damped Newton method, and obtain the classification result map of maritime targets and background;

[0037] The target detection module is used to binarize the classification result image to obtain the final target detection result.

[0038] The embodiments of the present invention have the following technical effects:

[0039] This invention proposes a target detection method and system based on damped Newton's method optimized clustering, representing a lightweight end-to-end detection process. First, a mask for the visible region below the coastline is accurately constructed using the longest contour selection and column-by-column adaptation strategy, effectively eliminating land interference. Then, a binary clustering algorithm based on damped Newton's method optimized K-means is designed, utilizing gray-level prior initialization and second derivatives to accelerate convergence, achieving high-precision separation of small targets from the seawater background. This invention significantly improves the accuracy and anti-interference capability of target detection. Attached Figure Description

[0040] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 This is a flowchart of a target detection method based on damped Newton's method optimized clustering, as described in an embodiment of the present invention.

[0042] Figure 2 This is an example image containing a coastline and maritime targets in an embodiment of the present invention;

[0043] Figure 3 This is an example of a region of interest image extracted in an embodiment of the present invention;

[0044] Figure 4 This is an example of a binary classification result in an embodiment of the present invention;

[0045] Figure 5 This is an example of the final target detection result in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0047] This invention proposes a target detection method based on damped Newton's method optimized clustering, such as... Figure 1 As shown, the method includes:

[0048] S1. Acquire an image containing the coastline and maritime targets;

[0049] S2. Perform coastline extraction and dynamic mask generation processing on the image to obtain an image of the region of interest containing only maritime targets.

[0050] S3. The K-means clustering algorithm optimized by the damped Newton method is used to perform binary clustering of targets and background in the region of interest image to obtain the classification result map of maritime targets and background;

[0051] S4. Binarize the classification result image to obtain the final target detection result.

[0052] The method begins with S1, in which images containing the coastline and maritime targets are acquired; such as Figure 2 As shown.

[0053] Then, in S2, the image is processed by coastline extraction and dynamic mask generation to obtain an image of the region of interest containing only maritime targets.

[0054] According to an embodiment of the present invention, in order to suppress image noise and highlight coastline features, the input image is first converted to grayscale and denoised using Gaussian filtering; then, adaptive histogram equalization is used for contrast enhancement; subsequently, the Canny edge detection operator is used to perform edge detection on the image; then, morphological operations are performed on the Canny edge-detected image to connect breakpoints and remove noise; next, all outer contours are extracted, and the longest continuous contour is selected as the main coastline. A land-sea segmentation mask is constructed to accurately remove land areas above the coastline. Finally, the land-sea segmentation mask is multiplied by the original image or the denoised image to obtain the Region of Interest (ROI) image, as shown below. Figure 3 As shown.

[0055] Then, in S3, the K-means clustering algorithm optimized by the damped Newton method is used to perform binary clustering of targets and background in the region of interest image to obtain the classification result map of maritime targets and background.

[0056] According to an embodiment of the present invention, in the obtained ROI image Above, the K-means clustering algorithm optimized by the damped Newton method is used to perform binary classification of the target boat and the background.

[0057] Damped Newton's method is a second-order gradient optimization algorithm, a significant improvement on the classic Newton's method, primarily used to solve unconstrained optimization problems. Its core idea is to introduce a step size factor into the search direction of the standard Newton's method, determining the optimal step size through a one-dimensional search, thus overcoming the Newton's method's sensitivity to initial points and ensuring global convergence. The core idea is: during iterative optimization, first calculate the gradient of the objective function and the Hessian matrix to determine the Newton search direction; then, using a fixed step size or a fixed-proportion decay step size strategy, update the optimization variables along this direction, and repeat the iteration until the convergence condition is met. K-means clustering is a classic and widely used unsupervised learning algorithm. Its core task is to automatically discover the inherent structure and grouping (i.e., "clusters") in unlabeled data. Its goal is to divide the data into K distinct clusters, making data points within the same cluster as similar as possible, while maximizing the differences between data points in different clusters.

[0058] The K-means algorithm optimized using the damped Newton's method is based on the idea of ​​transforming the clustering process into a numerical optimization problem and using the damped Newton's method to improve the centroid update strategy, thus overcoming the tendency of the traditional K-means algorithm to get trapped in local optima. While the traditional K-means algorithm can gradually converge during iteration by updating the centroid mean, its update direction at each step is often fixed, moving towards the cluster mean. This strategy easily gets trapped in local optima when the data distribution is complex or the initial centroid selection is poor. Introducing the damped Newton's method essentially treats the process of finding the optimal centroid as solving an unconstrained optimization problem.

[0059] The process of using the K-means clustering algorithm optimized by the damped Newton method to perform binary classification clustering of targets and background in the region of interest image is as follows.

[0060] 1) Initialization:

[0061] Traditional K-means algorithms use random initialization to set cluster centers, lacking prior information guidance. If the cluster centers are far from the true classification, it can easily lead to iterative convergence to an incorrect local optimum. To improve the rationality and relevance of the initialization, this invention improves the initialization as follows: Setting the number of categories K=2, based on the prior knowledge that the grayscale of the target boat is usually lower than the background, the cluster centers are initialized as follows:

[0062] ;

[0063] ;

[0064] in, The initial cluster centers of the target are represented. The initial cluster centers represent the background; , Offset:

[0065] ;

[0066] in, The weights reflect how close a data point is to its minimum value. This represents the grayscale value of the i-th pixel sample point. Represents the reference grayscale baseline for the target in dataset D; The variance of the Gaussian weights is represented, and in this embodiment, it is set to the variance of the low grayscale range; N represents the total number of image pixels. Let D represent the low grayscale quantile threshold of dataset D. The low grayscale quantile threshold is calculated using the p-quantile method: Assume the prior area ratio of the target is p=15%. Sort the image pixels in ascending order of grayscale value, and take the grayscale value corresponding to the first time the cumulative number of pixels reaches 15% of the total number of pixels as the low grayscale quantile threshold; D represents the dataset consisting of the grayscale values ​​of all pixels in the input image. Represents the global grayscale value of dataset D; In this embodiment, the scaling factor is used. .

[0067] 2) Objective function:

[0068] The traditional K-means algorithm uses the intra-class squared error as its objective function, relying solely on grayscale distance and neglecting differences in pixel density distribution in space or the grayscale domain. When the target boat and seawater have highly overlapping grayscale values, it cannot distinguish between sparse small target regions and dense background regions, resulting in blurred cluster boundaries and a high false negative rate. This invention aims to distinguish regions with similar grayscale values ​​but different distribution densities, improving cluster boundary accuracy. The objective function is improved as follows:

[0069] Let the two types of cluster centers be... ,in Corresponding to a small boat, Corresponding background. Calculate the objective function. as follows:

[0070] ;

[0071] in, Indicates the first A set of pixel samples, where i represents the i-th pixel sample point. This represents the clustering label of the i-th pixel sample point; This represents the average density of the i-th pixel sample point in the k-th class. This represents the grayscale value of the j-th pixel sample point. This represents the bandwidth of the Gaussian kernel function, which is set to 0.1 in this embodiment; This represents the average density of the k-th class of pixel samples. For the first Number of pixel samples; This embodiment represents the scaling factor that controls density sensitivity. =5.

[0072] 3) Iterative optimization:

[0073] In each iteration, the grayscale value of each pixel sample point is first calculated. Up to the current number Cluster centers The distance is assigned to the nearest center:

[0074] ;

[0075] Subsequently, the objective function of the current cluster is calculated. Next, the gradient of the objective function with respect to each cluster center is calculated. :

[0076] ;

[0077] And the Hessian matrix for each cluster center :

[0078] ;

[0079] Based on gradient Hessian matrix Calculate the Newton search direction for the k-th cluster center. .

[0080] The original damped Newton method, as a second-order optimization method, uses a fixed, globally uniform step size adjustment, which cannot adapt to different iteration stages or convergence states of different cluster centers. An excessively large step size easily leads to oscillations and divergence, while an excessively small step size results in slow convergence, making stable convergence difficult in complex contexts and potentially skipping the optimal solution, thus reducing the algorithm's robustness and clustering accuracy. Therefore, this invention improves the fixed step size to an adaptive step size update strategy. This strategy dynamically adjusts the step size based on both the change in the objective function during a single iteration and the variance of recent iterations. It adjusts independently for each cluster center and each iteration step, automatically balancing the step size, rapidly decreasing it during oscillations and moderately increasing it during stable periods, without requiring manual parameter tuning. The improved adaptive step size update strategy is as follows.

[0081] Let the initial step size be... A backtracking search strategy is adopted: if the objective function value in the current iteration is less than the objective function value in the previous iteration, or the search step size is less than a fixed threshold (10... -10 If the current step size is not found, the search is terminated; otherwise, the step size is updated according to the following formula:

[0082] ;

[0083] in, This represents the step size in the (t+1)th iteration; This represents the step size in the t-th iteration; This represents the basic attenuation coefficient, in this embodiment... ; This represents the change in the objective function. This represents the k-th cluster center in the t-th iteration. This represents the variance of recent changes in the objective function. This represents the number of historical iterations, which is set to 3 in this embodiment.

[0084] Finally, update the cluster centers along the Newtonian direction:

[0085] ;

[0086] in, For the first +1 iterations Cluster centers of pixel-like samples.

[0087] Until the change in all cluster centers is less than the preset convergence tolerance. Or reach the maximum number of iterations This achieves accurate binary classification of the target boat and the background. The classification results for the target boat and the background are as follows: Figure 4 As shown.

[0088] Then, in S4, the classification result image is binarized to obtain the final target detection result.

[0089] According to an embodiment of the present invention, the image obtained by S3 classification is binarized to generate a clear segmentation result. Specifically, pixels belonging to the category "target boat" are set to white (i.e., value 1), and the remaining background areas are set to black (i.e., value 0), thereby forming a binary mask image where white represents the target boat and black represents the background. Specific detection results are as follows... Figure 5 As shown.

[0090] This invention proposes a target detection method based on damped Newton's method optimized clustering. By integrating Canny edge detection, morphological processing, and longest contour selection, a dynamic column-oriented mask is constructed to accurately remove complex land backgrounds in nearshore images. This effectively solves the problem of target information loss or background residue caused by the difficulty of adapting traditional fixed ROI methods to winding coastlines. In the binary clustering stage, the improved initialization method makes it easier for the algorithm to start iterating from a reasonable starting point, reducing classification bias caused by improper initial center selection. The introduction of a density-aware objective function makes the algorithm more sensitive to small targets with sparse pixel distribution but practical significance, helping to distinguish real targets in backgrounds with similar gray levels. The Newton optimization mechanism with adaptive search step size improves the stability of the iteration process while ensuring convergence speed, reducing dependence on parameter tuning. This invention overcomes problems such as incomplete background suppression and blurred target segmentation, providing reliable technical support for practical applications such as maritime monitoring, maritime traffic management, and illegal fishing identification.

[0091] This invention also proposes a target detection system based on damped Newton's method optimized clustering. The system is used to implement the target detection method based on damped Newton's method optimized clustering described in the above embodiments; the system includes:

[0092] The image acquisition module is used to acquire images containing coastlines and maritime targets.

[0093] The region of interest extraction module is used to perform coastline extraction and dynamic mask generation processing on the image to obtain a region of interest image containing only maritime targets.

[0094] The binary classification module is used to perform binary classification and clustering of targets and background in the region of interest image using the K-means clustering algorithm optimized by the damped Newton method, and obtain the classification result map of maritime targets and background;

[0095] The target detection module is used to binarize the classification result image to obtain the final target detection result.

[0096] The function of the target detection system based on damped Newton's method optimized clustering described in this embodiment of the invention can be explained by the aforementioned target detection method based on damped Newton's method optimized clustering. 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.

[0097] It should be noted that the terminology used in this invention is for describing specific embodiments only and is not intended to limit the scope of this application. As shown in this specification, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may include the plural. The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A target detection method based on damped Newton's method optimized clustering, characterized in that, include: S1. Acquire an image containing the coastline and maritime targets; S2. Perform coastline extraction and dynamic mask generation processing on the image to obtain an image of the region of interest containing only maritime targets. S3. A K-means clustering algorithm optimized using the damped Newton method is used to perform binary classification clustering of targets and background in the region of interest image to obtain a classification result image of maritime targets and background; including: setting the pixel set of the region of interest image into two classes, target and background, and initializing two cluster centers; then in each iteration, calculating the distance from each pixel sample to the current cluster center and assigning a label, then obtaining the objective function, and calculating the gradient and Hessian matrix of each cluster center based on the objective function, calculating the Newton direction based on the gradient and Hessian matrix, updating the cluster centers along the Newton direction, and adaptively adjusting the search step size according to the decrease of the objective function; stopping the iteration when the change in the cluster centers is less than the preset tolerance or the maximum number of iterations is reached, thus completing the binary classification of target and background; wherein, the initialization of two cluster centers includes: ； ； in, Represents the initial cluster centers of the target; The initial cluster centers represent the background; D represents the dataset consisting of the gray values ​​of all pixels in the region of interest image. , Offset: ； in, This represents the weight of the i-th pixel sample point. This represents the grayscale value of the i-th pixel sample point. This represents the reference grayscale baseline for the target in dataset D. The variance represents the low grayscale range; N represents the total number of image pixel samples. The threshold value for the low grayscale quantiles in dataset D; Represents the global grayscale value of dataset D; This is the scaling factor; The objective function is: ； in, Indicates the cluster center. Corresponding to the target, Corresponding background; Indicates the first Pixel-like sample set, This represents the clustering label of the i-th pixel sample point; This represents the average density of the i-th pixel sample point in the k-th class. This represents the grayscale value of the j-th pixel sample point. This represents the bandwidth of the Gaussian kernel function; This represents the average density of the k-th class of pixel samples. For the first Number of pixel samples; The scaling factor that controls density sensitivity; The adaptive adjustment of the search step size based on the descent rate of the objective function includes: If the objective function value in the current iteration is less than the objective function value in the previous iteration, or the search step size is less than a fixed threshold, then accept the current step size and terminate the search; otherwise, update the search step size according to the following formula: ； in, This represents the step size in the (t+1)th iteration; This represents the step size in the t-th iteration; Indicates the basic attenuation coefficient; This represents the change in the objective function. This represents the k-th cluster center in the t-th iteration. Indicates the Newton search direction for the k-th cluster center; The variance representing the change in the objective function value; S4. Binarize the classification result image to obtain the final target detection result.

2. The target detection method based on damped Newton's method optimized clustering according to claim 1, characterized in that, S2 includes: The image is then converted to grayscale and denoised using Gaussian filtering. Adaptive histogram equalization is applied to the denoised image to enhance contrast. Subsequently, the Canny edge detection operator is used to perform edge detection on the image. Then, morphological operations are performed on the Canny edge detection image to connect breakpoints and remove noise. Next, all outer contours are extracted, and the longest continuous contour is selected as the main coastline to construct a land-sea segmentation mask. The land-sea segmentation mask is multiplied by the denoised image to obtain the region of interest image.

3. The target detection method based on damped Newton's method optimized clustering according to claim 1, characterized in that, The calculation of the gradient and Hessian matrix of each cluster center based on the objective function, the calculation of the Newton direction based on the gradient and Hessian matrix, and the updating of the cluster centers along the Newton direction, as described in S3, include: Calculate the gradient of the objective function with respect to each cluster center. : ； And the Hessian matrix for each cluster center : ； Based on gradient Hessian matrix Calculate the Newton search direction for the k-th cluster center. ; Update cluster centers along the Newtonian direction: ； in, For the first During the nth iteration Cluster centers of pixel-like samples For the first At the +1st iteration, the... Cluster centers of pixel-like samples.

4. A target detection system based on damped Newton's method optimized clustering, characterized in that, The system is used to implement the target detection method based on damped Newton's method optimized clustering as described in any one of claims 1-3; the system includes: The image acquisition module is used to acquire images containing coastlines and maritime targets. The region of interest extraction module is used to perform coastline extraction and dynamic mask generation processing on the image to obtain a region of interest image containing only maritime targets. The binary classification module is used to perform binary classification and clustering of targets and background in the region of interest image using the K-means clustering algorithm optimized by the damped Newton method, and obtain the classification result map of maritime targets and background; The target detection module is used to binarize the classification result image to obtain the final target detection result.

Images