Self-adaptive phase field sea ice image segmentation method
By reconstructing the sea ice segmentation task as a phase field evolution process and introducing edge and texture features, the adaptability problem of sea ice segmentation to irregular shapes and complex environments is solved, achieving high-precision and low-dependency sea ice segmentation results.
Patent Information
- Application Number
- CN202510967850.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-14
- Publication Date
- 2025-11-21
AI Technical Summary
Existing sea ice segmentation methods lack accuracy in handling irregular shapes, multi-scale features, and complex environmental conditions, and are highly dependent on a large amount of labeled data, exhibiting poor robustness.
The sea ice segmentation task is reconstructed as a phase field evolution process. Image edge and texture features are introduced as adaptive driving forces. Segmentation is performed through adaptive parameter adjustment and phase field evolution equations, thus constructing a segmentation mechanism with clear physical meaning.
It improves the segmentation accuracy and robustness for complex sea ice scenes, reduces the dependence on labeled data, and has good generalization ability and parameter interpretability.
Smart Images
Figure CN120997244A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of image processing and remote sensing, and in particular, relates to a self-adaptive phase-field sea ice image segmentation method. BACKGROUND
[0002] Sea ice is an important component of the polar ocean system, and accurate monitoring of its extent, concentration, and type is of great significance for climate change research, polar shipping safety, and ecological environment assessment. Remote sensing technology is the main means of obtaining large-scale, long-term sea ice information. However, sea ice in remote sensing images often exhibits complex characteristics such as irregular shape, fuzzy boundary, diverse scale, and low contrast with open water or thin ice. In addition, environmental factors such as clouds, fog, uneven illumination, and sensor noise also pose great challenges to accurate sea ice segmentation.
[0003] Existing sea ice segmentation methods mainly include threshold-based methods, edge detection-based methods, region growing-based methods, and machine learning-based methods (especially deep learning) that have developed rapidly in recent years. Threshold-based methods (such as Otsu method) are simple and fast, but sensitive to illumination and noise, and a single threshold cannot adapt to complex scene changes. Edge detection-based methods (such as Sobel and Canny operators) focus on local gradient changes in images, but are difficult to handle complex textures or fuzzy, discontinuous regions, and are prone to false edges or edge loss. Region growing-based methods are sensitive to seed point selection and growth criteria, and are prone to over-segmentation or under-segmentation. Deep learning-based methods, such as convolutional neural networks (CNN), perform well on specific data sets, but usually require a large amount of high-quality pixel-level annotation data for end-to-end training, which is costly and time-consuming to obtain. At the same time, the generalization ability of the model and its adaptability to new ice conditions not covered by the training data still need to be improved, and the internal decision-making process of the model lacks transparency and has poor interpretability.
[0004] Phase-field method is a powerful mathematical tool initially used to simulate interface dynamics problems in materials science, such as dendrite growth, phase separation, and crack propagation. Its core idea is to use a continuous auxiliary field variable (phase-field variable, typically ranging from [0,1] to [-1,1]) to describe the distribution of different phases (or regions), and to drive the evolution of the phase-field variable by minimizing the total free energy of the system (usually including interface energy and volume energy), thereby capturing the dynamic behavior of the interface in the form of a diffuse interface. Introducing the phase-field method into image segmentation transforms the segmentation problem into a minimization problem of an energy functional, naturally integrating region and boundary information, and possessing inherent advantages in handling complex boundaries and topological changes. However, when traditional phase-field segmentation models (such as the Chan-Vese model or its variants) are directly applied to sea ice remote sensing images, they may suffer from insufficient segmentation accuracy or parameter sensitivity due to insufficient consideration of the local adaptability of sea ice's unique multi-scale texture and strong edge features. Summary of the Invention
[0005] In view of this, the purpose of this invention is to propose an adaptive phase-field sea ice image segmentation method to address the problems of insufficient ability to handle irregular shapes, difficulty in multi-scale feature coupling, weak robustness to complex environmental conditions, and high dependence on large amounts of labeled data in existing sea ice segmentation techniques. This invention reconstructs the sea ice segmentation task as an equivalent physical process of sea ice "phase separation" or "breakup," and innovatively introduces image edge and texture features as spatial adaptive driving forces and adjustment factors for phase-field evolution, constructing a segmentation mechanism with clear physical meaning and strong data adaptability.
[0006] The technical means employed in this invention are as follows: An adaptive phase-field sea ice image segmentation method includes the following steps: S1. Initialization: Obtain the input sea ice image. The phase field function is initialized based on the intensity information of the input sea ice image. ; S2. Feature Extraction: Calculate the input sea ice image. Edge feature map And extract the input sea ice image. Multi-scale texture features to form texture feature terms ; S3. Adaptive parameter determination: Based on edge feature map Determine the spatially adaptive evolution rate parameters Coupling parameters that are spatially adaptive ; S4. Phase field evolution: Based on spatially adaptive evolution rate parameters Coupling parameters that are spatially adaptive The phase field function is iteratively updated based on the following evolution equation. :
[0007] in, This is the interface width parameter. The Laplace operator for the phase field function; terms This is a double-well potential term; Based on the elm image Image-driven items; S5. Segmentation result generation: In the phase field function After the convergence condition is met, the final phase field function is obtained. The final phase field function is thresholded to generate a binarized sea ice segmentation image.
[0008] Furthermore, in S1, the phase field function is initialized. The method involves using the Otsu thresholding method on the input image. The process involves setting the phase field function value of pixels with values higher than the Otsu threshold to a value close to 1, and setting the phase field function value of pixels with values lower than or equal to the Otsu threshold to a value close to 0.
[0009] Furthermore, in S2, the edge feature map By calculating the input image The Sobel gradient magnitude is obtained by normalizing it, and its calculation formula is as follows:
[0010] in It is a pixel. The magnitude of the Sobel gradient. It is the maximum magnitude of the gradient in the image. Use a small positive number to prevent the denominator from being zero.
[0011] Furthermore, in S2, texture feature terms By analyzing the input image The result is obtained by performing multi-scale Gaussian filtering and calculating the weighted sum of the absolute differences between the images before and after filtering. The calculation formula is as follows:
[0012] in The standard deviation is Gaussian filter, Representing an image With Gaussian filter convolution, Indicates multiple different scales Summing the results This represents the weighting coefficient for the texture item.
[0013] Furthermore, the weighting coefficients of the texture item Coupling coefficient with the basic Related, specifically:
[0014] in It is a preset constant.
[0015] Furthermore, in S3, the adaptive coupling parameters The calculation formula is:
[0016] in Based on the coupling coefficient, This represents the value at the corresponding location in the edge feature map. It is a positive number.
[0017] Furthermore, in S3, the adaptive evolution rate parameter The calculation formula is:
[0018] in Based on the fundamental evolutionary rate parameter, This represents the value at the corresponding location in the edge feature map. It is a positive number, and Use positive constants to ensure that the denominator is not zero and It is positive.
[0019] Furthermore, in S4, the image-driven item The calculation formula is:
[0020] in The normalized image intensity value. This is the reference grayscale constant.
[0021] Furthermore, in S4, the iterative update of the phase field function is solved numerically using the finite difference method, and a time step dt is set to ensure numerical stability and convergence efficiency.
[0022] Furthermore, in S5, the final phase field function... When performing threshold processing, the threshold is set to 0.5, that is, when When the value is greater than 0.5, the corresponding pixel is segmented as sea ice; otherwise, it is segmented as non-sea ice.
[0023] Compared with the prior art, the present invention has the following advantages: The present application abstracts the sea ice segmentation problem as a phase evolution process in a physical system, and gives the segmentation model a clear physical connotation and inherent regularization characteristics. At the same time, by introducing the low-level visual features such as edges and textures of the image as a spatially adaptive external driving force, the model can closely fit the actual content of the image, and effectively combine physical constraints and data information.
[0024] By introducing an adaptive evolution speed and coupling strength adjustment mechanism based on the edge feature map, the model can slow down evolution in strong edge regions and enhance the influence of the data fidelity term, thereby effectively maintaining the sharpness and accuracy of the sea ice boundary; in homogeneous regions, it accelerates convergence and strengthens the smoothing constraint to suppress noise effects.
[0025] The introduction of the multi-scale texture feature term in the present application enables the model to perceive image structure information at different scales, which helps to distinguish complex ice types and water areas with different texture characteristics, and improves the recognition ability of complex sea ice scenes.
[0026] The phase field method of the present application can naturally handle complex geometric shapes and topological changes (such as ice block merging and separation) without additional complex logic processing, and is very suitable for describing irregular shapes of sea ice.
[0027] Due to the physical basis of the model and its dependence on the intrinsic features of the image, the present application method has lower dependence on training samples compared to learning methods that rely solely on a large amount of labeled data, and has certain robustness to interference such as illumination changes and noise, and is expected to have better generalization potential on sea ice images of different sensors and different regions.
[0028] The main parameters (such as interface width , basic evolution speed , basic coupling coefficient ) in the model of the present application mostly have a clear physical or geometric meaning, which is convenient for users to understand, adjust and optimize according to prior knowledge or experiments. BRIEF DESCRIPTION OF DRAWINGS
[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.
[0030] Figure 1 The flowchart of the method of the present application.
[0031] Figure 2 The pre-processing example diagram of the sea ice remote sensing input image of the present application.
[0032] Figure 3 An example of the final segmented sea ice image obtained after applying the method of the present application is shown, wherein the black regions represent the segmented sea ice. DETAILED DESCRIPTION
[0033] In order to make the persons skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by the persons skilled in the art without creative labor should belong to the protection scope of the present application.
[0034] It should be noted that the terms “first”, “second”, and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms “include” and “have” and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.
[0035] As shown in Figure 1 , the present application provides an adaptive phase-field sea ice image segmentation method, comprising the following steps: S1, initialization: obtaining an input sea ice image , which is usually pre-processed (such as radiation correction, geometric correction, and normalization of pixel values to the range [0, 1]). The phase field function is initialized based on the intensity information of the input image . The Otsu threshold method can be used to initialize the region with intensity greater than the threshold value calculated by the Otsu method as sea ice phase, and the other regions as non-sea ice phase. This initialization provides a reasonable starting point for subsequent evolution.
[0036]
[0037] wherein is the threshold value calculated by the Otsu method.
[0038] S2, feature extraction: a. Edge feature map Calculation: the Sobel gradient operator is used to calculate the input image The gradient magnitude is calculated and normalized to obtain an edge intensity map ranging from [0,1]. This feature map characterizes the salience of edges in the image.
[0039] b. Multi-scale texture features Construction: By analyzing the input image Apply a set of different standard deviations A Gaussian filter is applied to obtain images with different degrees of smoothness. The absolute difference between the original image and the smoothed images at each scale is calculated, and these differences are weighted and summed to form texture feature terms. This feature can capture information about texture changes in an image at different scales.
[0040] Calculate the input image Edge feature map and extract the input image. Multi-scale texture features to form texture feature terms ; Edge feature map By calculating the input image The Sobel gradient magnitude is obtained by normalizing it, and its calculation formula is as follows:
[0041] in It is a pixel. The magnitude of the Sobel gradient. It is the maximum magnitude of the gradient in the image. Use a small positive number to prevent the denominator from being zero.
[0042] Texture features By analyzing the input image The result is obtained by performing multi-scale Gaussian filtering and calculating the weighted sum of the absolute differences between the images before and after filtering. The calculation formula is as follows:
[0043] in The standard deviation is Gaussian filter, Representing an image With Gaussian filter convolution, Indicates multiple different scales Summing the results This represents the weighting coefficient for the texture item.
[0044] Weight coefficient of texture item Coupling coefficient with the basic Related, specifically
[0045] wherein is a preset constant, .
[0046] S3, adaptive parameter determination: According to the edge feature map obtained in S2a , the key parameters in the evolution equation of the phase field are adaptively adjusted in space. Specifically, the spatially adaptive evolution speed parameter and the spatially adaptive coupling parameter are determined. This adaptive mechanism makes the coupling parameter increase in the edge region (e ) is larger), enhancing the image driving force effect, while the evolution speed decreases, making the phase field evolution slower and more fine, which helps to maintain the clarity of the edge; in the smooth region (e ) is smaller), the opposite is true, which helps to quickly stabilize the region.
[0047] According to the edge feature map , the spatially adaptive evolution speed parameter and the spatially adaptive coupling parameter are determined; The calculation formula of the adaptive coupling parameter is:
[0048] wherein is a basic coupling coefficient, is the value of the edge feature map at the corresponding position, is a normal number. The constant .
[0049] The calculation formula of the adaptive evolution speed parameter is:
[0050] wherein is a basic evolution speed parameter, is the value of the edge feature map at the corresponding position, is a normal number, and is a small normal number to ensure that the denominator is not zero and is positive. The constant .
[0051] S4, phase field evolution: the phase field function is iteratively updated at discrete time steps dt by numerical method (finite difference method) according to the following phase field evolution equation which contains interface regularization term, adaptive image driven term and texture feature term
[0052]
[0053] where, is the derivative of double well potential function, usually or , which is used to drive the phase field function to evolve towards stable two phases (e.g. 0 and 1). is a parameter to control the interface diffuse width, smaller corresponds to sharper interface. is the Laplacian operator of phase field function , which plays a smoothing role. is the image driven term based on the intensity information of input image , which is expressed as , where is the reference gray value (e.g. 0.5 for normalized image), which tries to push the phase field to the state consistent with the image intensity.
[0054] The calculation formula of image driven term is:
[0055] where is the normalized image intensity value, is the reference gray constant, usually taking 0.5.
[0056] The iterative update of phase field function is solved by finite difference method, and appropriate time step dt is set to ensure numerical stability and convergence efficiency.
[0057] S5, segmentation result generation: after the phase field function evolves over time and reaches the preset convergence condition (the maximum absolute value of the change of phase field function between two consecutive iterations is less than a small threshold value, or the preset maximum iteration number is reached), the final steady state phase field function is obtained. By threshold processing (with 0.5 as the threshold) on the final phase field function, the image is segmented into sea ice region and non-sea ice region , generating a binary sea ice segmentation result image.
[0058] Embodiment An adaptive phase field sea ice segmentation method, comprising the following steps: Reference Figure 1 The adaptive phase-field sea ice segmentation method described in this embodiment aims to accurately segment the sea ice region from the input remote sensing image .
[0059] S1, first image preprocessing such as Figure 2 : input sea ice remote sensing image , the following preprocessing steps are performed to reduce noise, enhance contrast and improve segmentation quality: gray scale conversion for ice thickness and density analysis based on intensity change, convert the original RGB image to a gray scale image. Homomorphic filtering: eliminate uneven illumination by frequency domain processing, retain texture details. Adaptive histogram equalization method to enhance local contrast. Then, linearly stretch or normalize the image pixel values to the standard range [0, 1] to obtain . Secondly, the initial phase field function generation adopts Otsu adaptive threshold method to calculate the global threshold . The best segmentation threshold is automatically determined by maximizing the inter-class variance. The phase field function is initialized according to the threshold . Select values close to 1 (such as 0.95) and close to 0 (such as 0.05) to avoid numerical calculation problems and provide a smooth evolution starting point.
[0060] S2, feature extraction, first edge feature extraction. For the preprocessed image , apply the Sobel operator to calculate the gradient approximation values and in the horizontal (x direction) and vertical (y direction) directions, respectively . Then, by calculating the gradient amplitude of each pixel point, the edge intensity of the image is obtained. Next, normalize these gradient amplitudes to the range to form the normalized edge feature map . Secondly, the extraction of texture features is completed by selecting a set of Gaussian filter standard deviations of different analysis scales. For each scale, use a Gaussian filter with a standard deviation of to smooth the input image , obtaining . Calculate the texture feature term of each pixel point , where is the number of selected scales (3 in this example), used to average the contribution; is the basic coupling coefficient, is the weight constant, set to 0.3, used to adjust the overall influence strength of the texture feature relative to other feature terms. This texture feature term can capture the residual error of the image under different smoothing levels, thus reflecting the complexity of local texture.
[0061] S3, determine adaptive parameters, first set basic parameters, basic evolution speed parameter , control the overall rate of phase field evolution. Typical values can be taken between 0.01 to 0.5. Basic coupling parameter , control the overall influence strength of image driving term on phase field evolution. Typical values can be taken between 1 to 100. These basic parameters are global constants, which can be set by a small amount of experimental data or experience. Secondly, calculate the spatial adaptive parameters: according to the edge feature map obtained in S2a , calculate adaptive parameters, the first is adaptive coupling parameter
[0062] wherein, is the edge feature map. Adaptive evolution speed parameter:
[0063] wherein, . Condition: to ensure , the condition needs to be met:
[0064] Since has been normalized to , the condition is simplified to:
[0065] S4, perform phase field evolution first set other model parameters: interface width parameter control the diffuse width of phase field from 0 to 1 transition region. In theory the smaller the interface is sharper, but too small may require smaller spatial step and time step. Typical values can be taken between 1.0 to 5.0 times the pixel spacing. Time step dt: time discrete step length for numerical integration. To ensure numerical stability (especially for explicit format), the selection of dt is usually subject to the CFL (Courant-Friedrichs-Lewy) condition, which needs to be small enough, dt can be taken between 0.001 to 0.1, depending on , and spatial discrete step length (usually 1 pixel). Secondly, define the image driving term, and finally solve the evolution equation. The explicit finite difference method is used to solve the phase field evolution equation. Laplace operator can be approximated by standard five-point or nine-point difference format. Five-point format is:
[0066] wherein, Space step.
[0067] Dual well potential definition:
[0068] Iterative update formula:
[0069] Final convergence judgment: the iteration process continues until any of the following convergence conditions is met: Maximum change criterion:
[0070] Wherein is a preset small tolerance .
[0071] Total energy change criterion:
[0072] The preset maximum iteration number .
[0073] The final steady-state phase field function is:
[0074] S5, generating a segmentation result on the final steady-state phase field function Simple threshold processing is performed to obtain a binary (hard) segmentation image As shown in Figure 3 excellent results are obtained.
[0075]
[0076] Wherein, is a segmentation threshold, usually and are pixel values representing sea ice and non-sea ice (water area) respectively.
[0077] The method of the present application can effectively combine the low-level features (edge, texture, intensity) of the image and the physical constraints of the phase field model through the above detailed steps, realize adaptive and robust segmentation of complex sea ice scenes. This method does not rely on large-scale labeled data sets, and the parameter adjustment has certain physical interpretation, providing an effective technical approach for automatic and high-precision interpretation of sea ice remote sensing images.
[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.
Claims
1. An adaptive phase-field sea ice image segmentation method, characterized in that, Includes the following steps: S1. Initialization: Obtain the input sea ice image. The phase field function is initialized based on the intensity information of the input sea ice image. ; S2. Feature Extraction: Calculate the input sea ice image. Edge feature map And extract the input sea ice image. Multi-scale texture features to form texture feature terms ; S3. Adaptive parameter determination: Based on edge feature map Determine the spatially adaptive evolution rate parameters Coupling parameters that are spatially adaptive ; S4. Phase field evolution: Based on spatially adaptive evolution rate parameters Coupling parameters that are spatially adaptive The phase field function is iteratively updated based on the following evolution equation. : in, This is the interface width parameter. The Laplace operator for the phase field function; terms This is a double-well potential term; Based on the elm image Image-driven items; S5. Segmentation result generation: In the phase field function After the convergence condition is met, the final phase field function is obtained. The final phase field function is thresholded to generate a binarized sea ice segmentation image.
2. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S1, the phase field function is initialized. The method involves using the Otsu thresholding method on the input image. The process involves setting the phase field function value of pixels with values higher than the Otsu threshold to a value close to 1, and setting the phase field function value of pixels with values lower than or equal to the Otsu threshold to a value close to 0.
3. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S2, the edge feature map By calculating the input image The Sobel gradient magnitude is obtained by normalizing it, and its calculation formula is as follows: in It is a pixel. The magnitude of the Sobel gradient. It is the maximum magnitude of the gradient in the image. Use a small positive number to prevent the denominator from being zero.
4. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S2, texture features By analyzing the input image The result is obtained by performing multi-scale Gaussian filtering and calculating the weighted sum of the absolute differences between the images before and after filtering. The calculation formula is as follows: in The standard deviation is Gaussian filter, Representing an image With Gaussian filter convolution, Indicates multiple different scales Summing the results This represents the weighting coefficient for the texture item.
5. The adaptive phase field sea ice segmentation method according to claim 4, characterized in that: The weight coefficient of the texture item Coupling coefficient with the foundation Related, specifically: in It is a preset constant.
6. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S3, the adaptive coupling parameters The calculation formula is: in Based on the coupling coefficient, This represents the value at the corresponding location in the edge feature map. It is a positive number.
7. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S3, the adaptive evolution rate parameter The calculation formula is: in Based on the fundamental evolutionary rate parameter, The value at the corresponding position in the edge feature map. It is a positive number, and Use positive constants to ensure that the denominator is not zero and It is positive.
8. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S4, the image-driven item The calculation formula is: in The normalized image intensity value. This is the reference grayscale constant.
9. The adaptive phase-field sea ice segmentation method according to claim 1, characterized in that: In S4, the iterative update of the phase field function is solved numerically using the finite difference method, and a time step dt is set to ensure numerical stability and convergence efficiency.
10. The adaptive phase field sea ice segmentation method according to claim 1, characterized in that: In S5, the final phase field function When performing threshold processing, the threshold is set to 0.5, that is, when When the value is greater than 0.5, the corresponding pixel is segmented as sea ice; otherwise, it is segmented as non-sea ice.