Highly dense ice fragmentation image segmentation method based on iterative MGAC and SAM model

Abstract

The application discloses a highly dense broken ice image segmentation method based on an iterative MGAC and SAM model, and the method comprises the following steps: performing sea ice instance segmentation on a pretreated image based on a SAM model to obtain a sea ice mask image corresponding to the pretreated image; acquiring an initial sea ice residual area image which is not recognized by the SAM model according to a binary image and the sea ice mask image; acquiring the initial sea ice residual area image under different gray threshold values to obtain an initial seed mask image; taking the initial sea ice residual area image and the initial seed mask image as inputs of a preset MGAC contour model, and acquiring an MGAC sea ice recognition result based on a multi-round iterative partition mechanism; and performing a set operation on the MGAC sea ice recognition result and the sea ice mask image to obtain a broken ice segmentation mask result. The application solves the problem of insufficient structure extraction precision and boundary integrity of a dense broken ice area in the prior art.

Description

Technical Field

[0001] This invention relates to the field of polar shipping technology, and in particular to a method for high-density ice fragmentation image segmentation based on iterative MGAC and SAM models. Background Technology

[0002] With global warming, the frequency of use of polar shipping routes is increasing, and their strategic value is gradually rising. When ordinary ships navigate icebreaking channels or ice floes, the size, shape, and density of ice fragments are important factors affecting ice resistance. Therefore, efficient identification and perception of ice fragment characteristics are crucial for shipping and route planning in polar regions.

[0003] In recent years, sea ice image recognition technology has made significant progress, mainly forming two major technical routes: traditional methods and deep learning-based methods.

[0004] Traditional methods mainly include thresholding, clustering, and region growing methods: the thresholding method was proposed by Weszka JS et al. [1] One approach divides an image into multiple slices using different grayscale values, suitable for images with significantly different grayscale values ​​or high background differentiation. Zhang Qin points out that threshold segmentation based on the Otsu algorithm distinguishes ice and water pixels by analyzing the grayscale histogram features of the image; however, this method has performance limitations in low-contrast scenes and is not suitable for ice and water mixing scenes with low color differentiation. [2] Furthermore, the threshold method has not been the mainstream research method since 2000. [3] Clustering methods project images into other spaces and select several image features for clustering. [4] M.-ANMoen et al. used clustering methods for automatic sea ice identification and compared the results of manual and automatic identification using synthetic aperture radar. They found that the automatic segmentation effect was poor due to the lack of clear distinction between sea ice segments. [5] The region growing algorithm searches outwards from a single pixel. Adams [6] It was pointed out that this was proposed by Davis LS et al. 7] Compared to the region growing algorithm, the watershed algorithm is a more prominent traditional method. [4] This method, proposed by Beucher S et al., uses the pixel grayscale values ​​of the entire region as the elevation to define the watershed as the boundary. [8] Steven Leigh et al. developed an automatic ice-water discrimination system using the watershed algorithm, achieving efficient and consistent classification of ice and water in dual-polarization synthetic aperture radar images. [9] However, gradient operators often produce many local minima due to noise or quantization errors. [6]Regarding the application performance of sea ice identification, region growing and watershed algorithms can avoid the continuity problem of boundary extraction by gradient calculation methods through region similarity criteria, but they are prone to oversegmentation in densely packed ice areas. Active contour models (ACMs) are an improved type of region growing method, first proposed by Michael Kass et al.

[10] Gradient Vector Flow (GVF) models are an improved active contouring method proposed by Chenyang Xu et al.

[11] The GVF algorithm achieves sub-pixel-level edge detection by establishing an energy functional minimization model, solving the convergence problem of active contours towards narrow boundary concavities in computer vision and image processing.

[11] Qin Zhang et al. applied the GVF algorithm to identify polar sea ice captured by drones.

[12] However, GVF is sensitive to the initial contour position and has high computational complexity. Márquez Neila et al. developed geodesic active contour (GAC)...

[12] Based on this, the problem of solving partial differential equations is approximately transformed into a morphological operator problem, and Morphological GAC (MGAC) is proposed.

[14] MGAC is a contour extraction method that considers morphology and is faster than GAC and GVF. Currently, traditional methods have been applied in the field of sea ice image recognition; for example, Alberto Alberello used MATLAB development tools to complete polar ice fragment segmentation.

[15] .

[0005] In summary, traditional methods have been applied in the field of sea ice image recognition, but still face many technical challenges. Traditional 2D image processing algorithms such as thresholding, region growing, and watershed algorithms can accurately track contour edges, but they heavily rely on the accuracy of the initial contour, thus only applicable to global scenes and unsuitable for segmenting real-world ice fragments. When dealing with highly dense ice fragments of varying sizes, with severe boundary adhesion, and complex lighting conditions, boundary fragmentation, over-segmentation, or under-segmentation are prone to occur. MGAC has the potential to address these issues.

[0006] For deep learning-based methods, with the development of deep learning technology, convolutional neural networks have demonstrated superior performance in sea ice segmentation tasks through end-to-end training. Chengqian Zhang et al. proposed the Ice-Deeplab improved network architecture, introducing an attention mechanism into the convolutional network and combining it with shallow features to obtain feature target boundaries of multi-scale sea ice.

[16] Ozgun et al. introduced an end-to-end learning method, U-Net, to semi-automatically and fully automatically segment 3D volume data from sparse annotations, achieving efficient data augmentation.

[17] Liang-Chieh Chen et al. improved the segmentation results by adding a decoder module and applied depthwise separable convolution to the pooling and decoder modules, thus obtaining the DeepLabv network.

[18] Salman Khaleghian et al. employed existing or novel convolutional neural network (CNN) architectures for the classification of various sea ice and open water types.

[19] Dong Wenbo applied neural networks to identify ice fragments on the hull side during icebreaking operations in a real-world environment.

[19] The image samples were few, and the shooting distance was relatively close; Zeng Dinghan applied neural networks to identify ice shards floating in the waterway after icebreaking in a virtual environment.

[20] Visual foundational models such as the Segment Anything Model (SAM)

[22] Leveraging its powerful feature extraction capabilities and cue-driven mechanism, it achieved zero-shot transfer in polar iceberg segmentation tasks.

[23] Passerotti et al. proposed the SAM-GVF method for ice fragment recognition, which performs basic segmentation based on SAM and then uses GVF to segment the remaining region once. However, because only one GVF segmentation is performed, it cannot perform fine segmentation of the remaining region, resulting in a limited segmentation rate. Furthermore, GVF is not a state-of-the-art contour model.

[24] .

[0007] In summary, architectures such as U-Net and DeepLabv have significantly improved segmentation accuracy under complex ice conditions through multi-scale feature fusion. However, these methods still face challenges such as scarce ice flake annotation data and missed detection of small ice floes. While intelligent segmentation methods like SAM have achieved some success in natural image segmentation, deep learning requires high-quality base samples, self-trained high-quality weight packs, and random numbers that match reality. Therefore, they lack the ability to accurately perceive weak edges, contiguous regions, and small-scale targets, and still suffer from incomplete boundary recognition and missing structural information in polar ice flake images.

[0008] In summary, there is an urgent need to provide a sea ice image segmentation method that can effectively improve the structural extraction accuracy and boundary integrity of dense ice fragment regions. Summary of the Invention

[0009] This invention provides a highly dense ice fragmentation image segmentation method based on iterative MGAC and SAM models to overcome the aforementioned technical problems.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] A highly dense ice fragment image segmentation method based on iterative MGAC and SAM models includes the following steps:

[0012] S1: Acquire video images of ice fragment distribution, and perform orthophoto correction on the video images of ice fragment distribution to obtain corrected image data; and the video images of ice fragment distribution are images captured by cameras installed on the bow, main mast, or side of the bridge of the ship.

[0013] S2: Preprocess the corrected image data to obtain a preprocessed image;

[0014] Binarize the preprocessed image to obtain a binary image;

[0015] S3: Perform sea ice instance segmentation on the preprocessed image based on the preset SAM model to obtain the sea ice mask image of the corresponding preprocessed image;

[0016] S4: Based on the constructed residual masking mechanism, obtain the initial sea ice residual region image that was not recognized by the SAM model based on the binarized image and the sea ice mask image;

[0017] S5: Based on the multi-threshold gray-level slicing method, sea ice residual region mask images under different gray-level thresholds are obtained from the initial sea ice residual region image, and these are used as candidate seed mask images.

[0018] Obtain the strength of each ice block in each candidate seed mask image, filter and obtain candidate seed mask images that meet the preset strength threshold, and stack and merge the filtered candidate seed mask images into a binary mask image to serve as the initial seed mask image for the preset MGAC contour model.

[0019] S6: The initial sea ice residual region image and the initial seed mask image are used as inputs to the pre-set MGAC contour model, and the sea ice recognition results of multiple rounds are obtained based on the multi-round iterative partitioning mechanism.

[0020] The sea ice identification results from each round are stacked and merged to obtain the MGAC sea ice identification results;

[0021] S7: Perform a union operation on the MGAC sea ice recognition result and the sea ice mask image obtained in S3 to obtain the ice fragmentation mask result, and the outer contour of the ice fragmentation mask result is the outer contour of the sea ice; thereby realizing the segmentation of the highly dense ice fragment image.

[0022] Furthermore, S2 specifically includes the following steps:

[0023] S21: Preprocess the corrected image data to obtain a preprocessed image;

[0024] The preprocessing includes: converting the image data to grayscale to obtain a grayscale image; performing median filtering denoising on the grayscale image to obtain a filtered and denoised image; performing contrast-limited adaptive histogram equalization on the filtered and denoised image to obtain an equalized image; and performing adaptive sharpening on the equalized image to obtain a preprocessed image.

[0025] S22: Binarization segmentation of the preprocessed image is performed based on the Phansalkar local thresholding algorithm to obtain the binary image M. bin .

[0026] Furthermore, S3 specifically includes the following steps:

[0027] S31: Use the preprocessed image as input to the preset SAM model, and obtain a set of initial mask images containing region masks through an automatic mask generator;

[0028] S32: Obtain the target mask image based on the constructed mask filtering rules;

[0029] And the mask filtering rules include:

[0030] Set a threshold for the area range of the mask region, and retain the initial mask image that meets the area range of the mask region; discard the initial mask images that do not meet the area range of the mask region, and obtain the optimized mask image;

[0031] Set a contour firmness threshold, obtain the contour firmness corresponding to each optimized mask image; retain the optimized mask images whose contour firmness meets the contour firmness threshold, discard the optimized mask images that do not meet the contour firmness threshold, and use the optimized mask images that meet the contour firmness threshold as the target mask images.

[0032] S33: Integrate all target mask images based on logical OR operations to obtain the sea ice mask image M, which represents the ice breakage area perceived by the SAM model. SAM .

[0033] Furthermore, the residual masking mechanism constructed in S4 is expressed as follows:

[0034]

[0035] Where: M residual Represents the initial sea ice residual region image; M bin Represents a binarized image; M SAM Represents a sea ice mask image; ∧ represents the bitwise AND operation symbol; The symbol representing the logical inverse operation of the mask.

[0036] Furthermore, S6 specifically includes the following steps:

[0037] S61: The initial sea ice residual region image and the initial seed mask image are used as inputs to the pre-set MGAC contour model, and the sea ice recognition results of multiple rounds are obtained based on the multi-round iterative partitioning mechanism.

[0038] Specifically, it includes:

[0039] S611: Mask the sea ice residual region image M residual As the current residual region of the pre-set MGAC contour model, the initial seed mask image M seed , which serves as the current seed mask for the pre-defined MGAC contour model;

[0040] S612: Perform MGAC segmentation on the current seed mask image and extract the sea ice identification results within the region corresponding to the current seed mask image.

[0041] S613: Confirm and remove sea ice identification results from the current residual region. The corresponding part is to update the current residual region to obtain a new residual region; the new residual region is defined as the initial sea ice residual region image, and the current seed mask image is updated based on step S5 to obtain a new seed mask image;

[0042] S614: Use the new residual region as the current residual region, use the new seed mask as the current seed mask, and repeat step S612 until the termination iteration condition is met, and obtain the sea ice identification results of multiple rounds; and the termination iteration condition: the difference between the residual region area corresponding to the current iteration round and the residual region area corresponding to the previous round does not exceed the preset difference threshold.

[0043] S62: Stack and merge the sea ice identification results from each round to obtain the MGAC sea ice identification result.

[0044] Furthermore, the formula for obtaining the ice fragmentation mask result in S7 is as follows:

[0045]

[0046] Where: M final This represents the result of ice fragmentation masking; M SAM Represents a sea ice mask image; This represents the sea ice recognition result output by the MGAC contour model in round t; T represents the total number of iterations; ∨ represents the union operator.

[0047] Beneficial Effects: This invention provides a highly dense ice fragment image segmentation method based on iterative MGAC and SAM models. By combining the global perception capability of the SAM model to segment sea ice instances in preprocessed images to obtain sea ice mask images, it can adapt to complex polar sea ice scenarios. Simultaneously, by constructing a residual mask mechanism, it obtains images of initial sea ice residual regions not recognized by the SAM model, and by introducing robustness as a screening index for the mask structure and combining it with a multi-threshold grayscale slicing method, it obtains an initial seed mask image, which can effectively improve the initialization stability of the active contour model, i.e., the MGAC contour model. This invention, through the constructed multi-round iterative partitioning mechanism, performs multiple rounds of sea ice recognition based on the MGAC contour model, which significantly improves the decoupling and boundary continuity of adhered targets, and does not require a large number of labeled samples. It has good transferability and practical engineering feasibility. This invention not only has an accuracy advantage in sea ice single-unit boundary extraction, but also reduces the omission, misjudgment, and fictitious sea ice contour phenomena of sea ice blocks, and can effectively reproduce the real distribution pattern of sea ice in complex environments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the 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 based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the channel density ice fragmentation image segmentation method based on iterative MGAC and SAM models of the present invention.

[0050] Figure 2 This is a diagram showing the layout of the image acquisition equipment in this embodiment;

[0051] Figure 3 This is a comparison diagram of orthophoto transformation in this embodiment;

[0052] Figure 4 This is the image preprocessing and binarization diagram in this embodiment;

[0053] Figure 5 This is the SAM segmentation and residual mask diagram in this embodiment;

[0054] Figure 6 This is a schematic diagram illustrating the automatic acquisition of the seed mask image in this embodiment;

[0055] Figure 7 This is a robustness histogram of the contour data of the original sea ice dataset in this embodiment.

[0056] Figure 8This is a SAM-MGAC segmentation result image of a typical sea ice image in this embodiment;

[0057] Figure 9 This is a comparison of a single model and SAM-MGAC segmentation of a typical sea ice image in this embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] This embodiment provides a highly dense ice fragmentation image segmentation method based on iterative MGAC and SAM models to overcome the above-mentioned technical problems.

[0060] To achieve the above objectives, the technical solution of the present invention is as follows:

[0061] A highly dense ice fragment image segmentation method based on iterative MGAC and SAM models, such as Figure 1 As shown, the specific steps include:

[0062] S1: Acquire video images of ice fragment distribution, and perform orthophoto correction on the video images of ice fragment distribution to obtain corrected image data; and the video images of ice fragment distribution are images captured by cameras installed on the bow, main mast, or side of the bridge of a ship;

[0063] Specifically, cameras are installed on the bow pillar or bridge side of the ship to capture images of ice in the ice-breaking zone or the ice-breaking channel being followed. Existing orthophoto correction methods are then used to project the images onto a plane perpendicular to the optical axis, avoiding perspective distortion and inconsistent scale between distant and near ice blocks caused by camera tilt. The orthophoto parameters involved in the orthophoto correction method include the camera's focal length, aperture, resolution, and optical center, as well as the camera's tilt angle and height. Figure 2 The diagram shown is a layout of shipboard equipment. Figure 3 The image shown is a sample image of the original image and the image after orthorectification correction. Figure 3 (a) is the original image; Figure 3 (b) is a sample image after orthophoto transformation correction, which verifies the effectiveness of the orthophoto correction method in improving image processing stability.

[0064] S2: Preprocess the corrected image data to obtain a preprocessed image, and binarize the preprocessed image to obtain a binary image. This includes the following steps:

[0065] S21: Preprocess the corrected image data to obtain a preprocessed image;

[0066] The preprocessing includes: converting the image data to grayscale to obtain a grayscale image; and performing median filtering denoising on the grayscale image to obtain a denoised image. The main function of median filtering denoising is to remove noise from the image while preserving its details and edge features. In this embodiment, median filtering is a non-linear digital filtering technique that reduces random noise by sorting the pixels in the image and selecting the median value as the smoothed pixel value. Median filtering is widely used in digital image processing because it preserves edges while removing noise, and is particularly suitable for removing non-linear noise such as salt-and-pepper noise and impulse noise. Compared with traditional average filtering methods, median filtering can better maintain image clarity and detail.

[0067] The filtered and denoised image is subjected to contrast-limited adaptive histogram equalization (CLAHE) to obtain an equalized image. In this embodiment, the purpose of contrast-limited adaptive histogram equalization is to improve the local contrast of the image while limiting the enhancement of contrast, thereby reducing the problem of noise amplification. In this embodiment, the CLAHE algorithm (Contrast Limited Adaptive Histogram Equalization) is used to improve the image distortion problem that may be caused by traditional histogram equalization by limiting the enhancement of contrast. When processing the image, CLAHE divides the image into several sub-blocks and performs histogram equalization on each sub-block, thereby improving the local contrast. At the same time, CLAHE limits the contrast amplification of each sub-block, avoiding image distortion and noise amplification caused by excessive local contrast.

[0068] An adaptive sharpening process is applied to the equalized image to obtain a preprocessed image. In this embodiment, the adaptive sharpening process aims to more accurately enhance image details while reducing noise and overshoot, and improving contrast distortion. Adaptive sharpening is an image enhancement technique that dynamically adjusts the enhancement factor to highlight image details and edges. Compared to traditional desharpening mask algorithms, adaptive sharpening adjusts the enhancement based on local image information to effectively suppress noise and overshoot while maintaining detail clarity, thereby improving the overall image quality and readability. Figure 4 As shown, where Figure 4 (a) is a sample image after orthophoto transformation correction; Figure 4(b) is the preprocessed image;

[0069] S22: Binarization segmentation of the preprocessed image is performed based on the Phansalkar local thresholding algorithm to obtain the binary image M. bin .like Figure 4 (c) is a binary image. In this embodiment, the Phansalkar local thresholding algorithm can adaptively determine the segmentation threshold based on the gray-scale statistical parameters of the local area of ​​the image. Compared with the traditional global thresholding method (such as Otsu), it shows higher foreground response capability in environments with uneven background and blurred boundaries. Its output binary image not only covers most of the broken ice area, but also retains small targets with weak boundaries.

[0070] S3: Based on the pre-set SAM model, perform sea ice instance segmentation on the pre-processed image to obtain the corresponding sea ice mask image of the pre-processed image, such as... Figure 5 As shown, the specific steps include:

[0071] S31: Use the preprocessed image as input to the preset SAM model, and obtain a set of initial mask images containing region masks through an automatic mask generator;

[0072] In this embodiment, after image preprocessing, to achieve rapid instance segmentation of sea ice targets, the Segment Anything Model (SAM) is introduced for master mask generation. The SAM model is a general segmentation architecture based on a visual fundamental model proposed by Meta, which has the ability to perform unsupervised mask inference for any foreground target in the image. In this embodiment, it runs through the "Automatic Mask Generator" interface, using the preprocessed image as the input image of the SAM model to automatically generate a set of region masks, which is especially suitable for remote sensing scenes that lack category labels or have strong structural heterogeneity. In polar sea ice images, the SAM model can be used to efficiently extract the main sea ice regions, reduce manual intervention, and improve model deployment efficiency.

[0073] S32: Obtain the target mask image based on the constructed mask filtering rules;

[0074] Considering the large differences in target particle size and the abundance of interference components in ice fragment images, this embodiment introduces mask filtering rules to improve mask quality;

[0075] And the mask filtering rules include:

[0076] Set a threshold for the area range of the mask region, and retain the initial mask image that meets the area range of the mask region; discard the initial mask images that do not meet the area range of the mask region, and obtain the optimized mask image;

[0077] In this embodiment, by removing false targets (such as noise) whose area is smaller than the set minimum area threshold, and targets whose area exceeds the maximum area threshold, it is ensured that only ice fragments with actual structural significance are retained.

[0078] Set a contour firmness threshold and obtain the contour firmness corresponding to each optimized mask image; retain optimized mask images whose contour firmness meets the contour firmness threshold, discard optimized mask images that do not meet the contour firmness threshold, and use the optimized mask images corresponding to the contour firmness threshold as the target mask image; wherein, in this embodiment, the contour firmness is defined as the ratio between the area of ​​the target contour region and its minimum bounding convex hull area, that is, by using the findContours function of OpenCV to obtain the edge contour of each ice block in each optimized mask image, using convexHull to obtain the minimum bounding convex hull contour, and then using contourArea to obtain the area of ​​the ice block and the minimum bounding convex hull, and the formula for obtaining the firmness is:

[0079]

[0080] S33: Integrate all target mask images based on logical OR operations to obtain the sea ice mask image M, which represents the ice breakage area perceived by the SAM model. SAM In this embodiment, after filtering according to the above mask filtering rules, all obtained target masks are integrated through a logical OR operation. Figure 5 (a) is the preprocessed image, resulting in a sea ice mask image as shown in the figure. Figure 5 (b) shows the resulting sea ice mask image, used to represent the main ice debris areas that the SAM model can perceive, i.e., the sea ice mask image M. SAM The sea ice mask image M SAM With its broad coverage and focused objectives, it provides a good initial structural support for subsequent boundary optimization.

[0081] S4: Based on the constructed residual masking mechanism, obtain the initial sea ice residual region image M that was not recognized by the SAM model, using the binarized image and the sea ice mask image. residual ;

[0082] In this embodiment, the SAM model is prone to missing detections or contour adhesion when processing areas of adhered debris, small target clusters, and weak boundary regions in highly dense polar sea ice images, which affects the overall boundary restoration accuracy. Therefore, this embodiment designs a residual mask mechanism based on difference construction to explicitly extract potential target regions that the SAM model fails to identify, as candidate regions for active contour evolution.

[0083] Specifically, the constructed residual masking mechanism is expressed as follows:

[0084]

[0085] Where: M residual Represents the initial sea ice residual region image; M bin Represents a binarized image; M SAM Represents a sea ice mask image; ∧ represents the bitwise AND operation symbol; The logical inverse operation symbol for the mask; in this embodiment, the initial sea ice residual region image M residual This characterizes sea ice regions in the current image that have not yet been captured by the SAM model but possess foreground features in structure, particularly concentrated in areas with blurred boundaries, small areas, or localized strong reflective interference. For example... Figure 5 (c) shows the initial sea ice residual region image. It can be seen that the initial sea ice residual region image effectively fills in the missed areas, enhances the integrity and perception range of the subsequent fine segmentation stage, and thus realizes coarse-grained perception and differential refinement separation of sea ice targets. It provides a clear structural starting point and target boundary range for subsequent seed region construction and multi-round contour evolution of the MGAC algorithm.

[0086] S5: Based on the multi-threshold gray-level slicing method, the initial sea ice residual region image is sliced ​​at equal intervals with different gray-level thresholds to obtain mask images of the sea ice residual region under different gray-level thresholds, and these are used as candidate seed mask images for binarization. In this embodiment, multi-threshold gray-level slicing is an image processing technique used to change the pixel values ​​in the image to form binarized slices, so as to highlight or suppress specific gray-level ranges, thereby improving the visual effect of the image, such as... Figure 6 As shown, where Figure 6 (a) is an image of the initial sea ice residual region; Figure 6 (b) is an image of the initial sea ice residual region of equally spaced slices with different grayscale thresholds;

[0087] The robustness of each ice block in each candidate seed mask image is obtained, and candidate seed masks that meet a preset robustness threshold are selected. In this embodiment, the robustness of each candidate seed mask image is obtained, such as... Figure 7 As shown, by analyzing the contour data of all images in the original sea ice dataset, it was found that 90% of the sea ice has a robustness ≥ 0.85; therefore, a robustness of 0.85 was selected as the screening threshold, and only areas that meet the closure requirement were retained.

[0088] The selected candidate seed mask images are stacked and merged into a binary mask image, which is used as the initial seed mask image of the preset MGAC contour model. This embodiment embeds this mechanism into the MGAC segmentation process of the MGAC contour model, enabling the model to achieve more targeted and convergent contour evolution in complex ice fragment images, laying the foundation for accurate boundary construction of adhered targets.

[0089] For example: Select and obtain candidate seed masks that meet a preset robustness threshold as a first candidate seed mask and a second candidate seed mask of the same size, and stack the first candidate seed mask and the second candidate seed mask together; if it is confirmed that there is no ice at a certain position in both images, then it is defined that there is no sea ice at the corresponding position in the merged binary mask; otherwise, it is defined that there is sea ice at the same position in the merged binary mask, and the final binary mask is used as the initial seed mask for the preset MGAC contour model, such as... Figure 6 (c) is the initial seed mask image;

[0090] In this embodiment, to ensure the effective initialization of the active contour model, i.e., the MGAC contour model, within the initial residual region, it is necessary to provide morphologically stable and geometrically reliable candidate mask regions as initial seed mask maps. Traditional Snake or level set models typically rely on manual interaction or heuristic strategies to select contour seeds, which are easily affected by factors such as adhesion and blurred boundaries in sea ice images, leading to deviations in the initial position of contour evolution and causing false convergence, structural splitting, or severe oversegmentation. This embodiment employs a multi-threshold grayscale slicing method, combined with an automatic seed region generation mechanism based on the robustness of contour geometric properties, to accurately extract structurally stable seed mask regions from the initial sea ice residual region image, which are then used to guide the adaptive convergence of the MGAC contour model.

[0091] S6: Using the initial sea ice residual region image and the initial seed mask image as input to the pre-set MGAC contour model, sea ice recognition results from multiple rounds are obtained based on a multi-round iterative partitioning mechanism; and the sea ice recognition results from each round are stacked and merged to obtain the MGAC sea ice recognition result. Specifically, this includes the following steps:

[0092] S61: The initial sea ice residual region image and the initial seed mask image are used as inputs to the pre-set MGAC contour model, and the sea ice recognition results of multiple rounds are obtained based on the multi-round iterative partitioning mechanism.

[0093] This embodiment completes the sea ice masking image M SAMAfter initializing the seed mask image, a Morphological Geodesic Active Contour (MGAC) model is introduced. MGAC is a level set-based, parameter-free contour model to further improve the segmentation accuracy of small-scale ice blocks, weak-edge targets, and adherent regions. Simultaneously, considering the large number of adherent structures, small ice blocks, and complex boundaries in polar ice fragment images, a single round of MGAC evolution is insufficient to achieve complete coverage. Therefore, a multi-round Refinement Strategy is designed, which uses a closed-loop execution process of "seed → evolution → fusion → residual update" to progressively clear residual regions, thereby improving the recognition accuracy of sea ice.

[0094] The specific steps include:

[0095] S611: Mask the sea ice residual region image M residual As the current residual region of the pre-set MGAC contour model, the initial seed mask image M seed , which serves as the current seed mask for the pre-defined MGAC contour model;

[0096] S612: Perform MGAC segmentation on the current seed mask image and extract the sea ice identification results within the region corresponding to the current seed mask image.

[0097] S613: Confirm and remove sea ice identification results from the current residual region. The corresponding part is to update the current residual region to obtain a new residual region; the new residual region is defined as the initial sea ice residual region image, and the current seed mask image is updated based on step S5 to obtain a new seed mask image;

[0098] S614: Use the new residual region as the current residual region, use the new seed mask as the current seed mask, and repeat steps S612 to S613 until the termination iteration condition is met, and obtain the sea ice recognition results of multiple rounds; and the termination iteration condition is: the difference between the residual region area corresponding to the current iteration round and the residual region area corresponding to the previous round does not exceed the preset difference threshold.

[0099] S62: Stack and merge the sea ice identification results from each round to obtain the MGAC sea ice identification result.

[0100] S7: Perform a union operation on the MGAC sea ice recognition result and the sea ice mask image obtained in S3 to obtain the ice fragmentation mask result, and the outer contour of the ice fragmentation mask result is the outer contour of the sea ice; thereby realizing the segmentation of the highly dense ice fragment image;

[0101] Specifically, the formula for obtaining the ice fragmentation mask result is as follows:

[0102]

[0103] Where: M final This represents the result of ice fragmentation masking; M SAM Represents a sea ice mask image; This represents the sea ice recognition result output by the MGAC contour model in round t; T represents the total number of iterations; ∨ represents the union operator.

[0104] like Figure 8 The image shows the SAM-MGAC segmentation results for a typical sea ice image; as shown... Figure 9 The image shows a comparison of the segmentation performance of typical sea ice images using the single-model SAM, MGAC, and the SAM-MGAC model described in this embodiment. The comparison demonstrates that, compared to the single-model MGAC, the SAM-MGAC model significantly reduces the omission and misclassification of ice blocks. Compared to the single-model SAM, the SAM-MGAC model significantly reduces the omission of ice blocks and the fabrication of sea ice outlines, effectively reproducing the true distribution pattern of sea ice in complex environments.

[0105] The beneficial effects of the method described in this embodiment are as follows: By combining the SAM model with the MGAC model, it not only has an advantage in the accuracy of sea ice block boundary extraction, but also reduces the omission, misjudgment, and fictitious sea ice contour phenomena of ice blocks. It can effectively reproduce the real distribution pattern of sea ice in complex environments. At the same time, its compatibility with multi-scale structures, its ability to identify adhesion boundaries, and its overall statistical consistency are significantly better than the single model SAM or MGAC. This method has certain practical application value in polar ice resistance modeling scenarios. In this embodiment, by combining the SAM model and the MGAC model, the SAM model is first used to perform preliminary global perception identification of sea ice. Then, considering the parametric characteristics of broken ice, the MGAC model is used for multiple iterative updates to achieve sea ice segmentation of the remaining sea ice residual region image, thus forming a new method suitable for broken ice segmentation. This method can avoid the drawbacks of fixed parameters in geometric evaluation of traditional methods, greatly reduce the necessity of high-quality datasets for intelligent segmentation methods, and effectively improve the sea ice structure extraction accuracy and boundary integrity in dense broken ice areas.

[0106] The following documents are involved in this embodiment:

[0107] [1]Weszka JS,Nagel et al.A Threshold Selection Technique.IEEETransactions on Computers,C-23(12):1322-1326,1974.

[0108] [2]Zhang Q,Skjetne R.Image Techniques for Identifying Sea-IceParameters.Modeling Identification and Control(MIC),2014,35(4):293-301.

[0109] [3] Zhang Jing, Zhang Ce, Zhang Ru, et al. Review of image segmentation: basic overview, typical algorithms and comparative analysis. Computer Technology and Development, 2024, 34(1):1-8.

[0110] [4] Li Hongyan, Cao Jianrong, Tan Wenting, et al. A review of image segmentation techniques. Journal of Shandong Jianzhu University, 2010, 25(1):85-89,93.

[0111] [5] Moen, M.-A., Doulgeris, AP, Anfinsen, SN, Renner, AH, Hughes, N., Gerland, S., & Eltoft, T. (2013). Comparison of feature based segmentation of full polarimetric SAR satellite sea ice images with manually drawn ice charts. TheCryosphere, 7(6), 1693–1705.

[0112] [6]Adams R.; Bischof L., Seeded region growing, IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(6):641-647,1994.

[0113] [7]Davis LS, A survey of edge detection techniques, Computer Graphics and Image Processing, 4(3):248-270,1975.

[0114] [8]Beucher,S.;Lantuéjoul,C.Use of Watersheds in Contour De tection.InProceedings of the International Workshop on Image Processing:Real-time Edgeand Motion Detection / Estimation,Rennes,France,Ja nuary 1979;Volume 132.

[0115] [9]Leigh,S.,Wang,Z.,&Clausi,D.A.(2013).Automated ice–waterclassification using dual polarization SAR satellite imagery.IEEETransactions on Geoscience and Remote Sensing,52(9),5529–5539.

[0116]

[10] Michael Kass,Andrew Witkin,and Demetri Terzopoulos.Snakes:Activecontour models.International Journal of Computer Vision,1(4):321–331,1988.

[0117]

[11] Xu,C.,&Prince,J.L.(1998).Generalized gradient vector flowexternal forces for active contours.Signal Processing,71(2),131–139.

[0118]

[12] Qin Zhang.,&Roger Skjetne.Image processing for identification ofsea-ice floes and the floe size distributions.IEEE Transactions on Geoscienceand Remote Sensing,53(5):2913–2924,2015.

[0119]

[13] Caselles,V.,Kimmel,R.&Sapiro,G.Geodesic ActiveContours.International Journal of Computer Vision,22:61–79,1997.

[0120]

[14] P.Márquez-Neila,L.Baumela and L.Alvarez,A Morphological Approachto Curvature-Based Evolution of Curves and Surfaces,IEEE Transactions onPattern Analysis and Machine Intelligence,36(1):2-17,2014.

[0121]

[15] Alberello,A.,Onorato,M.,Bennetts,L.,Vichi,M.,Eayrs,C.,Machutchon,K.,&Toffoli,A.(2019).Brief communication:Pancake ice floe size distributionduring the winter expansion of the Antarctic marginal ice zone.Cryosphere,13(1),41–48.

[0122]

[16] Chengqian Zhang,Xiaodong Chen,and Shunying Ji.Semantic imagesegmentation for sea ice parameters recognition using deep convolutionalneural networks.International Journal of Applied Earth Observation andGeoinformation,112:102885,2022.

[0123]

[17] Abdulkadir,A.,Lienkamp,SS,Brox,T.,&Ronneberger,O.(2016).3D U-Net:Learning dense volumetric segmentation from sparseannotation.Medical Image Computing and Computer-Assisted Intervention–MICCAI2016:19th International Conference,Athens,Greece,October 17-21,2016,Proceedings,Part II 19,424–432.

[0124]

[18] Chen,L.-C.,Zhu,Y.,Papandreou,G.,Schroff,F.,&Adam,H.(2018).Encoder-decoder with atrous separable convolution for semantic imagesegmentation.Proceedings of the European Conference on Computer Vision(ECCV),801–818.

[0125]

[19] Dong Wenbo. Research on ice channel identification and ship target tracking based on computer vision. Jiangsu: Jiangsu University of Science and Technology, 2023.

[0126]

[20] Zeng Dinghan. Research on polar sea ice recognition and application based on computer vision. Jiangsu: Jiangsu University of Science and Technology, 2024.

[0127]

[21] Khaleghian, S.,Ullah, H., T., Hughes, N., Eltoft, T., & Marinoni, A. (2021). Sea ice classification of SAR imagery based on convolution neural networks. Remote Sensing, 13(9), 1734.

[0128]

[22] Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, AC, Lo, W.-Y., & others. (2023). Segmentanything. Proceedings of the IEEE / CVF International Conference on ComputerVision, 4015–4026.

[0129]

[23] Siddharth Shankar,Leigh A.Stearns,and CJvan derVeen.Semantic.Segmentation of glaciological features across multiple remotesensing platforms with the segment anything model(SAM).Journal of Glaciology,pages 1–10,November 2023.ISSN 1727-5652.

[0130]

[24] Passerotti, Giulio&Alberello,Alberto&Vichi,Marcello&Bennetts,Luke&Toffoli,Alessandro.(2024). Segmenting sea ice floes inclose-range opticalimagery with active contour and foundation models.10.48550 / arXiv.2409.06641.

[0131] 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 scope of the technical solutions of the embodiments of the present invention.

Claims

1. A highly dense ice fragment image segmentation method based on iterative MGAC and SAM models, characterized in that, Specifically, the following steps are included: S1: Acquire video images of ice fragment distribution, and perform orthophoto correction on the video images of ice fragment distribution to obtain corrected image data; and the video images of ice fragment distribution are images captured by cameras installed on the bow, main mast, or side of the bridge of the ship. S2: Preprocess the corrected image data to obtain a preprocessed image; Binarize the preprocessed image to obtain a binary image; S3: Perform sea ice instance segmentation on the preprocessed image based on the preset SAM model to obtain the sea ice mask image of the corresponding preprocessed image; S4: Based on the constructed residual masking mechanism, obtain the initial sea ice residual region image that was not recognized by the SAM model based on the binarized image and the sea ice mask image; S5: Based on the multi-threshold grayscale slicing method, the sea ice residual region mask image under different grayscale thresholds is obtained according to the initial sea ice residual region image, and it is used as a candidate seed mask image. Obtain the strength of each ice block in each candidate seed mask image, filter and obtain candidate seed mask images that meet the preset strength threshold, and stack and merge the filtered candidate seed mask images into a binary mask image to serve as the initial seed mask image for the preset MGAC contour model. S6: The initial sea ice residual region image and the initial seed mask image are used as inputs to the pre-set MGAC contour model, and the sea ice recognition results of multiple rounds are obtained based on the multi-round iterative partitioning mechanism. The sea ice identification results from each round are stacked and merged to obtain the MGAC sea ice identification results; S7: Perform a union operation on the MGAC sea ice recognition result and the sea ice mask image obtained in S3 to obtain the ice fragmentation mask result, and the outer contour of the ice fragmentation mask result is the outer contour of the sea ice; thereby realizing the segmentation of the highly dense ice fragment image.

2. The highly dense ice fragment image segmentation method based on iterative MGAC and SAM models according to claim 1, characterized in that, S2 specifically includes the following steps: S21: Preprocess the corrected image data to obtain a preprocessed image; The preprocessing includes: converting the image data to grayscale to obtain a grayscale image; performing median filtering denoising on the grayscale image to obtain a filtered and denoised image; performing contrast-limited adaptive histogram equalization on the filtered and denoised image to obtain an equalized image; and performing adaptive sharpening on the equalized image to obtain a preprocessed image. S22: Binarization segmentation of the preprocessed image is performed based on the Phansalkar local thresholding algorithm to obtain the binary image M. bin .

3. The highly dense ice fragment image segmentation method based on iterative MGAC and SAM models according to claim 2, characterized in that, S3 specifically includes the following steps: S31: Use the preprocessed image as input to the preset SAM model, and obtain a set of initial mask images containing region masks through an automatic mask generator; S32: Obtain the target mask image based on the constructed mask filtering rules; And the mask filtering rules include: Set a threshold for the area range of the mask region, and retain the initial mask image that meets the area range of the mask region; discard the initial mask images that do not meet the area range of the mask region, and obtain the optimized mask image; Set a contour firmness threshold, obtain the contour firmness corresponding to each optimized mask image; retain the optimized mask images whose contour firmness meets the contour firmness threshold, discard the optimized mask images that do not meet the contour firmness threshold, and use the optimized mask images that meet the contour firmness threshold as the target mask images. S33: Integrate all target mask images based on logical OR operations to obtain the sea ice mask image M, which represents the ice breakage area perceived by the SAM model. SAM .

4. The highly dense ice fragment image segmentation method based on iterative MGAC and SAM models according to claim 3, characterized in that, The residual masking mechanism constructed in S4 is expressed as follows: Where: M residual Represents the initial sea ice residual region image; M bin Represents a binarized image; M SAM Represents a sea ice mask image; ∧ represents the bitwise AND operation symbol; The symbol representing the logical inverse operation of the mask.

5. The highly dense ice fragment image segmentation method based on iterative MGAC and SAM models according to claim 4, characterized in that, S6 specifically includes the following steps: S61: The initial sea ice residual region image and the initial seed mask image are used as inputs to the pre-set MGAC contour model, and the sea ice recognition results of multiple rounds are obtained based on the multi-round iterative partitioning mechanism. Specifically, it includes: S611: Mask the sea ice residual region image M residual As the current residual region of the pre-set MGAC contour model, the initial seed mask image M seed , which serves as the current seed mask for the pre-defined MGAC contour model; S612: Perform MGAC segmentation on the current seed mask image and extract the sea ice identification results within the region corresponding to the current seed mask image. S613: Confirm and remove sea ice identification results from the current residual region. The corresponding part is to update the current residual region to obtain a new residual region; the new residual region is defined as the initial sea ice residual region image, and the current seed mask image is updated based on step S5 to obtain a new seed mask image; S614: Use the new residual region as the current residual region, use the new seed mask as the current seed mask, and repeat step S612 until the termination iteration condition is met, and obtain the sea ice identification results of multiple rounds; and the termination iteration condition: the difference between the residual region area corresponding to the current iteration round and the residual region area corresponding to the previous round does not exceed the preset difference threshold. S62: Stack and merge the sea ice identification results from each round to obtain the MGAC sea ice identification result.

6. The highly dense ice fragmentation image segmentation method based on iterative MGAC and SAM models according to claim 5, characterized in that, The formula for obtaining the ice fragmentation mask result in S7 is as follows: Where: M final This represents the result of ice fragmentation masking; M SAM Represents a sea ice mask image; This represents the sea ice recognition result output by the MGAC contour model in round t; T represents the total number of iterations; ∨ represents the union operator.

Images