A weakly supervised pixel-level annotation method for shipboard sea ice images

Abstract

This invention belongs to the field of computer vision technology and provides a weakly supervised pixel-level annotation method for shipborne sea ice images: The method involves acquiring the original image of Arctic shipborne sea ice, cropping the sky and visually blurred areas, removing ship pixels using a ship region mask, and obtaining an effective observation image containing only sea ice and seawater. The image is then partitioned from far to near, converted to the LAB color space, and pixel feature vectors are extracted. These feature vectors are input into an SPGMM model, and pixel categories are determined based on the posterior probability of sea ice and seawater. The partitioned annotation results are then stitched together and post-processed to obtain a pixel-level binary mask for sea ice. This invention addresses the problems of high manual annotation costs for shipborne sea ice images and the poor adaptation of traditional global segmentation models to image spatial heterogeneity, leading to low segmentation accuracy, blurred boundaries, and misclassification.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and provides a weakly supervised pixel-level annotation method for shipborne sea ice images. Background Technology

[0002] With the increasing activity in polar shipping and engineering, the acquisition and analysis of sea ice environmental information is of great significance in fields such as polar navigation safety assessment, ship ice maneuvering research, and sea ice dynamics analysis. In recent years, shipborne visual observation systems have gradually become one of the important means of sea ice observation. Compared with satellite remote sensing, shipborne camera systems can acquire high-resolution, real-time sea ice image data from the ship's perspective, thus having high application value in engineering applications such as ice channel identification, sea ice concentration estimation, and ship ice resistance research.

[0003] In image-based sea ice analysis tasks, pixel-level segmentation of sea ice and seawater regions is a crucial foundation for many subsequent studies, such as sea ice concentration calculation, ice floe scale statistics, and sea ice structural feature extraction. Current mainstream image segmentation methods primarily rely on deep learning models for training, which typically require a large amount of high-quality pixel-level labeled data as training samples. However, shipborne sea ice images often possess complex visual features, such as irregular sea ice morphology, blurred ice-water boundaries, significant variations in lighting conditions, and interference factors like fog or melt pools, all of which increase the difficulty of manual annotation. Therefore, performing complete pixel-level manual annotation for each image often requires a significant investment of time and manpower, becoming a major factor limiting the construction of sea ice datasets.

[0004] To reduce annotation costs, some studies have attempted to use unsupervised or weakly supervised methods for automatic segmentation of sea ice images. For example, unsupervised methods based on K-means clustering classify sea ice and seawater by clustering pixel features. However, these methods typically assume a relatively stable global feature distribution, which can easily lead to misclassification in complex sea ice scenarios. Furthermore, segmentation methods based on Gaussian Mixture Models (GMMs) can describe the statistical features of images using probabilistic models and combine this with limited annotation information for pixel classification. When there are significant spatial variations in the image, such as scale differences due to changes in observation distance, illumination variations, and local ice condition differences, the global model often struggles to adapt to the feature distributions of different regions simultaneously, thus reducing the stability and accuracy of the segmentation results.

[0005] Furthermore, in actual Arctic shipborne observation scenarios, sea ice images typically exhibit significant spatial heterogeneity. For example, ice structures are usually clearly visible in the foreground, while the texture features in distant areas are often more blurred due to perspective effects and atmospheric influences. Simultaneously, there may be significant differences in light intensity, sea ice coverage, and melt pool distribution across different regions. In such cases, traditional global segmentation models struggle to effectively characterize the local variations in the image, leading to problems such as blurred segmentation boundaries, loss of local structures, or misidentification of sea ice regions.

[0006] Therefore, how to make full use of limited annotation information while reducing the cost of manual annotation and improving the adaptability of segmentation methods to complex shipborne sea ice images has become a pressing technical problem in the field of automatic annotation and segmentation of sea ice images. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a weakly supervised pixel-level annotation method for shipborne sea ice images. This method solves the technical problems of high cost of manual pixel-level annotation for shipborne sea ice images, low segmentation accuracy caused by the inability of traditional global segmentation models to adapt to the spatial heterogeneity of images, and blurred boundaries and misclassification.

[0008] The technical solution of the present invention includes: The original image of Arctic ship-borne sea ice is acquired, the sky area and visually blurred areas in the original image are cropped, the location of the ship in the original image is located, and the pixel area of ​​the ship's location is covered and removed using a ship area mask with the same size and scale as the original image, so as to obtain an effective observation image containing only sea ice and seawater information.

[0009] The effective observation image content is divided into multiple partitions from far to near. The image of each partition is converted from RGB color space to LAB color space, and the three LAB channels of each pixel are used as pixel feature vectors.

[0010] The pixel feature vector is input into the corresponding SPGMM model. The SPGMM model calculates the posterior probability of each pixel belonging to the sea ice component and the posterior probability of it belonging to the seawater component. The two posterior probabilities are compared, and the pixel category is determined by the component corresponding to the larger posterior probability, thus completing the pixel labeling in each partition.

[0011] The various regions with completed pixel annotations are stitched together to obtain a complete image mask. The complete image mask is then post-processed to obtain a pixel-level binary mask for sea ice, thus completing the annotation. In the binary mask, a pixel value of 0 represents seawater and a pixel value of 255 represents sea ice.

[0012] Furthermore, an overlapping area is set between two adjacent partitions. The overlapping area is used to reduce the discontinuity problem at the partition boundary and realize the sharing of context information between adjacent areas.

[0013] Furthermore, the SPGMM model includes Gaussian mixture models constructed independently for each partition.

[0014] Furthermore, the training process for the Gaussian mixture model includes: Manually annotate the valid observation images and establish an annotation matrix with the same size as the valid observation images. Unannotated pixels are marked as 0, sea ice pixels as 1, and seawater pixels as 2. The number of pixels in the graffiti annotation samples accounts for at least 5% of the number of pixels in the target area, and the number of graffiti annotation samples of any category is not less than a preset threshold.

[0015] Extract graffiti-annotated sea ice and seawater samples from each partition, and merge the two types of samples to form the training sample set for each partition.

[0016] Based on the binary classification task of sea ice and seawater, the number of Gaussian mixture models for each partition is determined to be 2, with the two Gaussian components corresponding to the probability distribution characteristics of sea ice and seawater within the partition, respectively.

[0017] The covariance type of each partition Gaussian mixture model is set to the complete covariance form. The complete covariance matrix is ​​used to characterize the fluctuation range, channel linkage, distribution directionality and ellipsoid shape of the Gaussian components in the feature space of the three channels of LAB, and output the Gaussian mixture model with the defined distribution shape.

[0018] The training sample sets of each partition are input into a Gaussian mixture model with a defined distribution shape. The K-means algorithm is used to perform clustering operations on the training sample sets. The cluster centers are used as the initial mean of the Gaussian components, the distribution of samples within the clusters are used as the initial covariance of the Gaussian components, and the proportion of samples in each cluster is used as the initial mixing weight of the Gaussian components. The Gaussian mixture model with initialized parameters is then output.

[0019] Based on the training sample set, the parameters of the Gaussian mixture model after parameter initialization are iteratively optimized using the EM algorithm until the parameters tend to stabilize, and the trained Gaussian mixture models for each partition are output.

[0020] Furthermore, the iterative optimization process of the EM algorithm involves repeatedly executing the E-step and M-step until the model parameters stabilize, including: Step E: Based on the Gaussian mixture model parameters initialized with the current parameters, calculate the posterior probability of each training sample belonging to each Gaussian component, and estimate the probability of each training sample being classified as sea ice or seawater.

[0021] M-step: Based on the posterior probability results obtained in the E-step, update the mean vector, covariance matrix, and mixture weights of each Gaussian component to make the parameters of the Gaussian mixture model after parameter initialization match the actual distribution of the training samples.

[0022] Furthermore, the semantic category mapping method for the sea ice component, seawater component, and Gaussian component in the Gaussian mixture model is as follows: input the sea ice graffiti annotation sample into the trained SPGMM model, count the Gaussian component number to which the sea ice graffiti annotation sample belongs, define the Gaussian component that appears most frequently as the sea ice component, and the other Gaussian component corresponds to the seawater component.

[0023] Further post-processing of the complete image mask includes: Based on the 8-neighborhood, a connected component analysis is performed on the complete image mask. The area information of each foreground connected region is statistically analyzed, and discrete noise regions with an area smaller than the minimum connected component area threshold are removed.

[0024] The denoised mask is inverted, and the hole regions inside the mask are detected again based on the connected component analysis of the 8-neighborhood. The hole regions with an area smaller than the maximum hole area threshold are filled.

[0025] The filled mask is smoothed by using median filtering with a preset window size to suppress boundary spikes and isolated noise points.

[0026] Furthermore, the minimum connected component area threshold is 600 pixels, the maximum hole area threshold is 300 pixels, and the median filter window size is 5×5.

[0027] Furthermore, when stitching together the annotation results of each partition, the pixels of the ship image masking area corresponding to the ship area mask are uniformly set to 0.

[0028] The technical solution provided by this invention has the following advantages compared with the prior art: First, invalid regions such as the sky, ships, and visual blur are removed through preprocessing to avoid irrelevant information interfering with the model's distribution learning. Then, the image is partitioned according to near and far viewing distances and converted to the LAB color space, effectively adapting to the spatial heterogeneity features of near-field and far-field textures and large differences in illumination in shipborne sea ice images. Next, the partitioned SPGMM model is used to perform posterior probability discrimination of sea ice / seawater for pixels in each region, avoiding the defect of global models being unable to fit local features. Finally, a regularized binary mask is obtained through partitioning, stitching, and post-processing. This approach not only significantly reduces the time and labor costs of manual annotation through a weak supervision mechanism, but also significantly improves the accuracy, continuity, and stability of pixel-level annotation in complex sea ice scenes. It solves the technical problems of high cost of manual pixel-level annotation in shipborne sea ice images and low segmentation accuracy, blurred boundaries, and misclassification caused by the inability of traditional global segmentation models to adapt to the spatial heterogeneity of images. Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0029] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 The image is the original image obtained in an embodiment of the present invention.

[0031] Figure 2 The image is a result of cropping the sky area and the visually blurred area in an embodiment of the present invention.

[0032] Figure 3 The image shown is an example of a mask covering and removing the location of the ship in an embodiment of the present invention.

[0033] Figure 4 This is an embodiment of the present invention. Figure 2 and Figure 3 The combined image.

[0034] Figure 5 The image is manually annotated according to an embodiment of the present invention.

[0035] Figure 6 This is the final binarized mask image generated in the embodiment of the present invention. Detailed Implementation

[0036] The following detailed description of a specific embodiment of the present invention is provided in conjunction with the accompanying drawings. However, it should be understood that the scope of protection of the present invention is not limited to the specific embodiment.

[0037] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0038] In the description of the embodiments of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0039] like Figures 1 to 6 As shown, this invention provides a weakly supervised pixel-level annotation method for shipborne sea ice images, comprising: Shipborne images of Arctic sea ice during navigation are acquired using cameras mounted on the bow of the vessel. These are the original images of Arctic shipborne sea ice. Since the original images usually contain sky areas, ship structures, and areas that are difficult to identify from a distance, it is necessary to first crop the sky areas and visually blurred areas from the original images. Then, the location of the ship in the original images is located, and the pixel area of ​​the ship's location is covered and removed using a mask with the same size and scale as the original image. This yields an effective observation image containing only sea ice and seawater information, reducing the impact of irrelevant information on subsequent processing.

[0040] Considering the significant spatial differences in shipborne sea ice images—such as substantial variations in texture, brightness, and sea ice structure between near and far areas—the images are divided vertically into multiple horizontal regions to represent different observation distances. This spatial partitioning allows for the establishment of local statistical models in different regions, thereby improving the model's adaptability to changes in local features. Then, the image for each partition is converted from the RGB color space to the LAB color space, and the three LAB channels of each pixel are used as the pixel feature vector.

[0041] The pixel feature vector is input into the corresponding SPGMM model. The SPGMM model calculates the posterior probability of each pixel belonging to the sea ice component and the posterior probability of it belonging to the seawater component. The two posterior probabilities are compared, and the pixel category is determined by the component corresponding to the larger posterior probability, thus completing the pixel labeling in each partition.

[0042] The various regions with completed pixel annotations are stitched together to obtain a complete image mask. The complete image mask is then post-processed to obtain a binary mask image of sea ice at the pixel level, thus completing the annotation. In the binary mask, a pixel value of 0 represents seawater and a pixel value of 255 represents sea ice.

[0043] In the embodiments provided by the present invention, in order to reduce the discontinuity problem that may occur at the partition boundary, an overlapping area of ​​a certain pixel range is set between adjacent partitions, so that adjacent areas can share some context information, thereby improving the continuity and stability of the segmentation results.

[0044] In the embodiments provided by this invention, the SPGMM model includes Gaussian mixture models constructed independently for each partition. The training process of the Gaussian mixture model includes:

[0045] Manual annotations were added to the valid observation images, such as... Figure 5 For the triangular and pentagram stars in the image, a labeling matrix consistent with the size of the effective observed image is established. Unlabeled pixels are marked as 0, sea ice pixels as 1, and seawater pixels as 2. The number of pixels in the graffiti-labeled samples accounts for at least 5% of the total number of pixels in the target area, and the number of graffiti-labeled samples of any category is not less than a preset threshold, thus significantly reducing the cost of manual labeling. The graffiti-labeled pixels, as known category samples, are used to initialize and guide the estimation process of model parameters, thereby assigning corresponding category semantic information to each Gaussian distribution component. In this way, the Gaussian mixture model can probabilistically model pixels based on the statistical characteristics of local image regions, thus avoiding the problem that a single global model cannot adapt to complex visual changes.

[0046] Extract graffiti-annotated sea ice and seawater samples from each partition, and merge the two types of samples to form the training sample set for each partition.

[0047] Based on the binary classification task of sea ice and seawater, the number of Gaussian mixture models for each partition is determined to be 2, with the two Gaussian components corresponding to the probability distribution characteristics of sea ice and seawater within the partition, respectively.

[0048] The covariance type of each partition Gaussian mixture model is set to the complete covariance form. The complete covariance matrix is ​​used to characterize the fluctuation range, channel linkage, distribution directionality and ellipsoid shape of the Gaussian components in the feature space of the three channels of LAB, and output the Gaussian mixture model with the defined distribution shape.

[0049] The training sample sets of each partition are input into a Gaussian mixture model with a defined distribution shape. The K-means algorithm is used to perform clustering operations on the training sample sets. The cluster centers are used as the initial mean of the Gaussian components, the distribution of samples within the clusters are used as the initial covariance of the Gaussian components, and the proportion of samples in each cluster is used as the initial mixing weight of the Gaussian components. The Gaussian mixture model with initialized parameters is then output.

[0050] Based on the training sample set, the parameters of the Gaussian mixture model after parameter initialization are iteratively optimized using the EM algorithm until the parameters tend to stabilize, and the trained Gaussian mixture models for each partition are output.

[0051] In the embodiments provided by this invention, the iterative optimization process of the EM algorithm involves repeatedly executing the E-step and M-step until the model parameters stabilize, including: Step E: Based on the Gaussian mixture model parameters initialized with the current parameters, calculate the posterior probability of each training sample belonging to each Gaussian component, and estimate the probability of each training sample being classified as sea ice or seawater.

[0052] M-step: Based on the posterior probability results obtained in the E-step, update the mean vector, covariance matrix, and mixture weights of each Gaussian component to make the parameters of the Gaussian mixture model after parameter initialization match the actual distribution of the training samples.

[0053] In the embodiments provided by the present invention, the semantic category mapping method of the sea ice component, the seawater component and the Gaussian component in the Gaussian mixture model is as follows: input the sea ice graffiti annotation sample into the trained SPGMM model, count the Gaussian component number to which the sea ice graffiti annotation sample is predicted, define the Gaussian component that appears most frequently as the sea ice component, and the other Gaussian component corresponds to the seawater component.

[0054] In the embodiments provided by this invention, the post-processing of the complete image mask includes: Based on the 8-neighborhood, a connected component analysis is performed on the complete image mask. The area information of each foreground connected region is statistically analyzed, and discrete noise regions with an area smaller than the minimum connected component area threshold are removed.

[0055] The denoised mask is inverted, and the hole regions inside the mask are detected again based on the connected component analysis of the 8-neighborhood. The hole regions with an area smaller than the maximum hole area threshold are filled.

[0056] The filled mask is smoothed by using median filtering with a preset window size to suppress boundary spikes and isolated noise points.

[0057] In the embodiments provided by the present invention, the minimum connected region area threshold is 600 pixels, the maximum hole area threshold is 300 pixels, and the median filter window size is 5×5.

[0058] In the embodiments provided by the present invention, when stitching together the labeling results of each partition, the pixels of the ship image shielding area corresponding to the ship area mask are uniformly set to 0.

[0059] This invention can use a small amount of graffiti annotations as weak supervision information, and achieve automatic pixel-level annotation of sea ice and seawater regions by combining image spatial partitioning with local probability models, thereby reducing the cost of manual annotation while improving the accuracy and stability of segmentation results.

[0060] Example: This invention proposes a pixel-level annotation method for shipborne sea ice images based on doodle-guided annotation and a partitioned Gaussian mixture model. This method uses a small amount of hand-drawn doodles to provide category priors, constructs a partitioned Gaussian mixture probability model within the effective image region, and automatically classifies unannotated pixels, thereby generating a pixel-level segmentation mask for sea ice and seawater. The overall technical process includes image preprocessing, hand-drawn annotation, partitioned Gaussian mixture model modeling, unannotated pixel classification, and segmentation result optimization.

[0061] First, the input shipborne sea ice images are preprocessed. Since shipborne sea ice images typically contain not only the sea ice and seawater regions to be classified, but also sky and ship hull areas, these invalid observation areas can interfere with the statistical distribution if directly used in modeling. Therefore, effective region filtering is necessary before modeling. The system first removes the sky region from the shipborne images, then locates the ship's position in the original image and uses a mask of the same size and proportion as the ship's region to cover and exclude this area, forming a ship image masked region. This results in an effective observation image containing only sea ice and seawater regions. This processing ensures that subsequent model parameters are determined solely by the pixels of the real sea ice scene, avoiding interference from ship structure images or sky background noise on the statistical distribution.

[0062] After preprocessing, weakly supervised samples are generated through doodle annotation. For each image, an annotation matrix with the same size as the image is created, where unannotated pixels are marked as 0, sea ice pixels as 1, and seawater pixels as 2. Annotators only need to select representative sea ice and seawater areas in the image and draw a small number of lines; the pixel positions corresponding to these lines are written into the annotation matrix. Compared to traditional pixel-by-pixel annotation, this method only requires a small number of class seeds to complete subsequent automatic expansion, thus significantly reducing the workload of manual annotation. To ensure the stability of model estimation, the system constrains the number of samples; when the number of samples for any class is less than a preset threshold, model expansion is not performed on the image.

[0063] In the model building phase, this invention employs a partitioned Gaussian mixture model for probabilistic modeling. Considering the significant color distribution differences between foreground and background areas in shipborne images due to perspective effects, using a single global model for fitting could easily lead to aliasing. Therefore, this invention divides the image vertically into four spatial partitions, establishing an independent Gaussian mixture model for each partition. Let the image height be (H), then the height of each partition is (H / 4), and the pixel range corresponding to the (i)th partition is ([(i-1)H / 4, iH / 4]). Within each partition, the color distribution of sea ice and seawater is modeled based on manually annotated samples, and their probability distribution in the feature space is learned through the Gaussian mixture model. Unannotated pixels are then probabilistically classified. This partitioning approach allows the model to learn the color distribution features of different viewing distances, thereby improving classification accuracy.

[0064] Within each partition, the image is first converted from the RGB color space to the LAB color space, and the three LAB channels are used as pixel feature vectors. Compared to the RGB color space, the LAB color space can better separate luminance and chromaticity information, making it more suitable for representing the high brightness of sea ice, the low brightness of seawater, and the color differences between ice and water. Subsequently, combined with a manually labeled matrix, sea ice smear samples and seawater smear samples are extracted within the current partition, and the two types of samples are merged to form a training sample set. A Gaussian mixture model is then used to probabilistically model the color distribution of these samples.

[0065] In the process of establishing the Gaussian mixture model, the model is first determined to be composed of two Gaussian components according to the binary classification task. Then, the distribution shape of each component is defined by the complete covariance matrix. On this basis, the k-means algorithm is used to provide stable initial values ​​for the model parameters. Finally, the parameters are iteratively optimized by the EM algorithm to complete the probabilistic modeling of the color distribution of sea ice and seawater. (1) Since the ice and water in the image need to be divided, which is a binary classification task, two probability distribution components are used to represent the main distribution pattern of the samples in the current partition. All subsequent parameter estimations are based on these two components. The number of components in the model structure is determined to be 2. (2) Subsequently, since a Gaussian component does not have only one center point, it is also necessary to describe its diffusion range and direction in the feature space. Therefore, the mean vector and covariance matrix are used to represent the sample distribution shape inside each Gaussian component. The complete covariance matrix is ​​used here because the three channels LAB in the ship-based sea ice image are not completely independent. The brightness change is often accompanied by the color change, and the sea ice and seawater may change simultaneously in terms of brightness difference and color difference. The complete covariance matrix can simultaneously describe the fluctuation range of each channel, the linkage between channels, the directionality and ellipsoidal shape of the distribution in the feature space. The complete covariance matrix solves the problem of what mathematical form to use to characterize the distribution shape of each Gaussian component. (3) After determining the model structure and prediction form, it is necessary to input the initial parameters for iterative optimization calculation. Here, the K-means initialization method is used to calculate the initial value before the formal iterative optimization. The K-means initialization method first performs preliminary clustering of the samples, roughly dividing the samples into two clusters. The cluster center is used as the initial mean of the Gaussian component, the distribution of samples within the cluster is used to estimate the initial covariance, and the sample ratio of each cluster is used to estimate the initial mixing weight. The two clusters are used to give the initial parameters for subsequent iterative calculation. (4) Based on the initialization, the EM algorithm is used to correct the model. The training sample pixels of sea ice and seawater obtained by graffiti annotation are probabilistically modeled to estimate their distribution parameters in the feature space, so that the constructed Gaussian mixture model can accurately characterize the color distribution features of sea ice and seawater. The E-step calculates the posterior probability of each sample belonging to each Gaussian component based on the current parameters, estimating the probability of each sample being classified as ice or water under the current model. The M-step updates the mean vector, covariance matrix, and mixture weights of each component based on the probability results obtained in the E-step, making the model parameters more consistent with the actual distribution of the current samples. Finally, the E-step and M-step are repeated until the parameters are basically stable.

[0066] After model training is complete, probability prediction and category determination are performed on unlabeled pixels. Specifically, the feature vector of a pixel in the LAB color space is input into the Gaussian mixture model, and the posterior probability of it belonging to each Gaussian component is calculated. Since the components output by the Gaussian mixture model do not have explicit semantic categories, it is necessary to determine the correspondence between each component and the semantic category based on doodle annotation samples. The specific method is to re-input the sea ice annotation samples into the trained model, count the component numbers to which they are predicted, and define the component that appears most frequently as the sea ice component, with the other component corresponding to the seawater category. After completing the category mapping, for any unlabeled pixel within a partition, its posterior probability under each Gaussian component is calculated, and it is classified into the corresponding category according to the maximum posterior probability principle, thereby realizing the propagation of labels from a small number of doodle annotations to all pixels within the partition.

[0067] After obtaining the classification results for each partition, the predicted masks for each partition are stitched together to restore the segmentation results at the full image scale. For the ship mask shielding area in the preprocessing stage, the corresponding pixels are uniformly set to 0 to avoid interference from invalid areas on the classification results, thus obtaining the final pixel-level segmentation mask for sea ice and seawater.

[0068] To further improve the quality of the binary mask predicted by the model, this invention introduces a mask optimization process after classification. First, an 8-neighborhood connected component analysis is performed on the initial segmentation mask. Here, 8 neighbors refer to the eight adjacent pixels in the horizontal, vertical, and diagonal directions centered on the current pixel, which together form the connectivity determination range. The area information of each foreground connected region is statistically analyzed, and a minimum connected region area threshold of 600 pixels is set. Only connected regions with an area not less than this threshold are retained, while smaller discrete regions are identified as noise and removed. Subsequently, the denoised binary mask is inverted, and another 8-neighborhood connected component analysis is performed to detect internal hole regions. A maximum hole area threshold of 300 pixels is set. For hole regions with an area less than this threshold, a filling operation is performed in the original mask to restore the internal continuous structure of the sea ice region and reduce voids caused by local misclassification. After small-area removal and hole filling, a median filter with a window size of 5×5 is further applied to smooth the binary mask to suppress boundary burrs and isolated noise points, improving the continuity and regularity of the sea ice region edges. Through the above processing, a sea ice pixel-level segmentation mask with a more complete structure, smoother boundaries, and less noise can be obtained.

[0069] The generated masks correspond one-to-one with the original images, forming a standardized dataset structure. The system stores the original images and binary masks in the image directory and mask directory, respectively, where the mask pixel values ​​use 0 and 255 to represent seawater and sea ice categories. This dataset can be directly used to train semantic segmentation models, thereby enabling automatic recognition and analysis of sea ice images.

[0070] It should be noted that any parts not disclosed or specifically described in this invention are existing technology or conventional configurations, and their specific structures and working principles will not be elaborated further. In this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, 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, article, 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, article, or apparatus that includes that element.

[0071] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A weakly supervised pixel-level annotation method for shipborne sea ice images, characterized in that, include: The original image of Arctic ship-borne sea ice is obtained, the sky area and visually blurred areas in the original image are cropped, the location of the ship in the original image is located, and the pixel area of ​​the ship location is covered and removed using a ship area mask with the same size and scale as the original image, so as to obtain an effective observation image containing only sea ice and seawater information. The effective observed image content is divided into multiple partitions from far to near. The image of each partition is converted from RGB color space to LAB color space, and the three LAB channels of each pixel are used as pixel feature vectors. The pixel feature vector is input into the corresponding SPGMM model. The SPGMM model calculates the posterior probability of each pixel belonging to the sea ice component and the posterior probability of each pixel belonging to the seawater component. The two posterior probabilities are compared, and the pixel category is determined by the component corresponding to the larger posterior probability, thus completing the pixel labeling in each partition. The various partitions with completed pixel annotations are stitched together to obtain a complete image mask. The complete image mask is then post-processed to obtain a binary mask image of sea ice at the pixel level, thus completing the annotation. In the binary mask, a pixel value of 0 represents seawater and a pixel value of 255 represents sea ice.

2. The weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 1, characterized in that, An overlapping region is set between two adjacent partitions. The overlapping region is used to reduce the discontinuity problem at the partition boundary and realize the sharing of context information between adjacent regions.

3. The weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 1, characterized in that, The SPGMM model includes Gaussian mixture models constructed independently for each partition.

4. The weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 3, characterized in that, The training process of the Gaussian mixture model includes: The effective observation images are manually annotated to establish an annotation matrix with the same size as the effective observation images. Unannotated pixels are marked as 0, sea ice pixels as 1, and seawater pixels as 2. The number of pixels in the graffiti annotation samples accounts for at least 5% of the number of pixels in the target area, and the number of graffiti annotation samples of any category is not less than a preset threshold. Extract graffiti-annotated sea ice and seawater samples from each partition, and merge the two types of samples to form the training sample set for each partition. Based on the binary classification task of sea ice and seawater, the number of Gaussian mixture models for each partition is determined to be 2, and the two Gaussian components correspond to the probability distribution characteristics of sea ice and seawater in the corresponding partition, respectively. The covariance type of each partition Gaussian mixture model is set to the complete covariance form. The complete covariance matrix is ​​used to characterize the fluctuation range, channel linkage relationship, distribution directionality and ellipsoid shape of the Gaussian components in the feature space of the three channels of LAB, and output the Gaussian mixture model with the defined distribution shape. The training sample set of each partition is input into the Gaussian mixture model with a defined distribution shape. The K-means algorithm is used to perform clustering operation on the training sample set. The cluster center is used as the initial mean of the Gaussian component, the distribution of samples within the cluster is used as the initial covariance of the Gaussian component, and the proportion of samples in the cluster is used as the initial mixing weight of the Gaussian component. The Gaussian mixture model with initialized parameters is output. Based on the training sample set, the parameters of the Gaussian mixture model after parameter initialization are iteratively optimized using the EM algorithm until the parameters tend to stabilize, and the trained Gaussian mixture models for each partition are output.

5. A weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 4, characterized in that, The iterative optimization process of the EM algorithm involves repeatedly executing the E-step and M-step until the model parameters stabilize, including: Step E: Based on the Gaussian mixture model parameters after initialization with the current parameters, calculate the posterior probability of the training sample belonging to each Gaussian component, and estimate the probability of each training sample being classified as sea ice or seawater. M-step: Based on the posterior probability results obtained in the E-step, update the mean vector, covariance matrix, and mixture weights of each Gaussian component to make the parameters of the Gaussian mixture model after parameter initialization match the actual distribution of the training samples.

6. The weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 1, characterized in that, The semantic category mapping method for the sea ice component, seawater component, and Gaussian component in the Gaussian mixture model is as follows: Input the sea ice graffiti annotation samples into the trained SPGMM model, count the Gaussian component numbers to which the sea ice graffiti annotation samples belong, define the Gaussian component that appears most frequently as the sea ice component, and the other Gaussian component as the seawater component.

7. A weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 1, characterized in that, Post-processing of the complete image mask includes: Perform 8-neighborhood-based connected component analysis on the complete image mask, statistically analyze the area information of each foreground connected region, and remove discrete noise regions with an area smaller than the minimum connected component area threshold; The denoised mask is inverted, and the hole regions inside the mask are detected again based on the connected component analysis of the 8-neighborhood. The hole regions with an area smaller than the maximum hole area threshold are filled. The filled mask is smoothed by using median filtering with a preset window size to suppress boundary spikes and isolated noise points.

8. A weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 7, characterized in that, The minimum connected component area threshold is 600 pixels, the maximum hole area threshold is 300 pixels, and the median filter window size is 5×5.

9. A weakly supervised pixel-level annotation method for shipborne sea ice images according to claim 1, characterized in that, When stitching together the annotation results of each partition, the pixels of the ship image masking area corresponding to the ship area mask are uniformly set to 0.

Images