Sea ice image segmentation method, device, equipment, medium and product

Abstract

The invention discloses a sea ice image segmentation method, device and equipment, a medium and a product. A shipborne sea ice data set is constructed, the shipborne sea ice data set is composed of sea ice images and semantic tags, and the semantic tags comprise ocean open water areas, sea ice, ships and the sky; constructing a to-be-trained image segmentation model, wherein the to-be-trained image segmentation model at least comprises a lightweight backbone network, a dual-path attention mechanism and a pyramid pooling module; the lightweight backbone network adopts depth separable convolution and is embedded into cavity convolution; in the double-path attention mechanism, a channel attention branch performs channel dimension optimization on a feature map output by the lightweight backbone network, and a space attention branch performs space dimension optimization on the feature map after channel dimension optimization; and training and optimizing the to-be-trained image segmentation model based on the shipborne sea ice data set to obtain a target image segmentation model, and realizing sea ice image segmentation through the target image segmentation model. The model complexity is reduced, and the segmentation precision and the edge retention capability are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of computer technology, and in particular to a method, apparatus, device, medium and product for sea ice image segmentation. Background Technology

[0002] With the increasing demand for navigation in complex ice conditions, dredging vessels are operating more frequently in such environments. The shipborne vision system, as the core component for dredging vessels to perceive detailed local ice conditions, directly determines the safety and efficiency of operations. However, complex ice conditions generally present challenges such as variable lighting conditions and diverse sea ice morphologies (including thin ice, broken ice, and floating ice), posing serious challenges to sea ice image segmentation technology.

[0003] General semantic segmentation models may suffer from the following problems in the aforementioned scenarios: the backbone network has a large number of model parameters and high computational complexity, making it difficult to achieve real-time inference on shipborne terminals with limited computing and storage resources, thus failing to meet the real-time requirements of dynamic operations on engineering vessels; the segmentation accuracy of sea ice images is insufficient in complex ice conditions, such as thin ice being easily misidentified as waves, seawater in broken ice areas being easily missed, and small target hull features being easily lost; for complex boundaries such as broken ice edges, the segmentation results are blurry, and edge localization is inaccurate, making it difficult to support accurate ice risk assessment and operational decisions. Therefore, there is an urgent need for a sea ice image segmentation method that is adapted to shipborne scenarios and can cope with complex ice conditions. Summary of the Invention

[0004] This invention provides a sea ice image segmentation method, apparatus, device, medium, and product. Through the collaborative design of lightweight structure, feature enhancement, and scene adaptation, it significantly reduces model complexity while greatly improving the accuracy and edge preservation capability of sea ice image segmentation in complex ice conditions.

[0005] In a first aspect, embodiments of the present invention provide a sea ice image segmentation method, comprising:

[0006] Construct a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels, the semantic labels including at least open ocean waters, sea ice, ships, and sky;

[0007] A training image segmentation model is constructed, which includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network employs depthwise separable convolutions and embeds dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the feature map after channel dimension optimization through a spatial attention branch.

[0008] The image segmentation model to be trained is trained and optimized based on the shipborne sea ice dataset to obtain the target image segmentation model, and the sea ice image segmentation is achieved through the target image segmentation model.

[0009] Furthermore, the lightweight backbone network includes sequentially connected feature extraction layers 1 to 5, with dilated convolutions introduced in the 4th and 5th feature extraction layers.

[0010] Furthermore, the execution steps of the channel attention branch include:

[0011] Global average pooling and global max pooling are performed on the feature maps output by the lightweight backbone network to obtain two channel description vectors.

[0012] Two channel description vectors are input into a two-layer neural network, and the channel weight coefficients are output through a non-linear activation function.

[0013] The channel weight coefficients are multiplied channel by channel by the feature map output by the lightweight backbone network to obtain the feature map with optimized channel dimensions.

[0014] Furthermore, the execution steps of the spatial attention branch include:

[0015] The channel-dimensional optimized feature maps are subjected to channel-dimensional average pooling and max pooling respectively to obtain two spatial description maps;

[0016] Two spatial description maps are concatenated along the channel dimension, and spatial weight coefficients are output through convolutional layers and nonlinear activation functions.

[0017] The spatial weight coefficients are multiplied pixel-by-pixel by the channel dimension-optimized feature map to obtain the spatial dimension-optimized feature map.

[0018] Furthermore, the execution steps of the pyramid pooling module include:

[0019] The feature map output by the dual-path attention mechanism is subjected to four-level pooling. After each pooling branch is reduced in dimensionality by convolution and upsampled, it is concatenated with the feature map output by the dual-path attention mechanism in the channel dimension to output a segmentation feature map.

[0020] Furthermore, the model parameters of the image segmentation model to be trained are optimized using the average intersection-union ratio and edge detection accuracy as evaluation indicators.

[0021] Secondly, embodiments of the present invention provide a sea ice image segmentation apparatus, comprising:

[0022] The dataset construction module is used to construct a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels. The semantic labels include at least open ocean waters, sea ice, ships, and sky.

[0023] The model building module is used to build a training image segmentation model, which includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network uses depthwise separable convolutions and embeds dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the feature map after channel dimension optimization through a spatial attention branch.

[0024] The model training and application module is used to train and optimize the image segmentation model to be trained based on the shipborne sea ice dataset, obtain the target image segmentation model, and realize sea ice image segmentation through the target image segmentation model.

[0025] Thirdly, embodiments of the present invention provide an electronic device, including:

[0026] At least one processor; and

[0027] A memory communicatively connected to the at least one processor; wherein,

[0028] The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0029] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions that cause a processor to execute the method described in the first aspect.

[0030] Fifthly, embodiments of the present invention provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0031] The technical solution of this invention involves constructing a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels. The semantic labels include at least open ocean water, sea ice, ships, and sky. A training image segmentation model is constructed, comprising at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network employs depthwise separable convolutions with embedded dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the channel dimension-optimized feature map through a spatial attention branch. The training image segmentation model is trained and optimized based on the shipborne sea ice dataset to obtain a target image segmentation model, which is then used to segment sea ice images. This scheme employs a lightweight backbone network combining depthwise separable convolutions and dilated convolutions, enabling the model to be applied in shipboard scenarios with limited computing and storage resources. Through a channel-spatial dual-path attention mechanism, the model adaptively focuses on key regions such as sea ice and thin ice edges, suppressing background interference. Using shipboard sea ice datasets, the model learns domain-specific feature representations. In short, through the synergistic design of lightweight structure, feature enhancement, and scene adaptation, the scheme significantly reduces model complexity while substantially improving sea ice image segmentation accuracy and edge preservation capabilities in complex ice conditions.

[0032] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0034] Figure 1 This is a flowchart of a sea ice image segmentation method provided in Embodiment 1 of the present invention;

[0035] Figure 2 This is a schematic diagram of the model structure of an image segmentation model to be trained according to Embodiment 1 of the present invention;

[0036] Figure 3 This is a schematic diagram of the structure of a sea ice image segmentation device according to Embodiment 2 of the present invention;

[0037] Figure 4This is a schematic diagram of the structure of an electronic device that implements an embodiment of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0039] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0040] Example 1

[0041] Figure 1 This is a flowchart of a sea ice image segmentation method according to Embodiment 1 of the present invention. This embodiment is applicable to the implementation of sea ice image segmentation. The method can be executed by a sea ice image segmentation device, which can be implemented in software and / or hardware and integrated into an electronic device. Furthermore, the electronic device may include, but is not limited to, computers, laptops, servers, shipborne terminals, etc.

[0042] like Figure 1 As shown, the method includes:

[0043] S110. Construct a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels. The semantic labels include at least open ocean waters, sea ice, ships, and sky.

[0044] In this step, sea ice images taken by icebreakers in complex ice conditions can be acquired. Here, sea ice images can be understood as images containing sea ice. The acquired sea ice images are preprocessed, such as adjusting them to a preset uniform pixel size. Using annotation tools, pixel-level semantic labels are added to the sea ice images to obtain a shipborne sea ice dataset. The semantic labels are of four categories: open ocean waters, sea ice, ships, and sky. Open ocean waters are waters without ice bodies, ships, or other interference covering the sea surface. Sea ice refers to all visible ice bodies. Ships refer to the hull area. Sky refers to the visible sky area. Multiple shipborne sea ice datasets are generated in the same way to form a shipborne sea ice dataset.

[0045] Optionally, generate 5000 labeled images (i.e., shipborne sea ice data) and divide them into training, validation and test sets in an 8:1:1 ratio.

[0046] S120. Construct an image segmentation model to be trained. The image segmentation model to be trained includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network uses depthwise separable convolutions and embeds dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the feature map after channel dimension optimization through a spatial attention branch.

[0047] The image segmentation model to be trained can be a model for sea ice image segmentation, and in this embodiment of the invention, it can be an improved Pyramid Scene Parsing Network (PSPNet) model. The model structure of the image segmentation model to be trained includes the following:

[0048] Input layer: Receives the input image from the model, which is the sea ice image to be segmented.

[0049] Lightweight Backbone Network: Replacing the traditional PSPNet backbone network, this network achieves a balance between lightweight design and feature extraction capabilities through depthwise separable convolutions and dilated convolutions. Specifically, the lightweight backbone network includes multiple feature extraction layers, each processed according to the execution logic of depthwise separable convolutions, with dilated convolutions embedded in deeper feature extraction layers. The execution logic of depthwise separable convolutions replaces standard convolutions, decomposing them into depthwise convolutions (channel-wise convolutions, reducing spatial dimension computation) and pointwise convolutions (1×1 convolutions, integrating channel features), thus reducing model computational complexity while maintaining feature extraction performance. The execution logic of dilated convolutions, while maintaining the output feature map size of the current feature extraction layers, increases the receptive field compared to traditional backbone networks, enhancing the ability to capture features from large-scale sea ice.

[0050] Dual-path attention mechanism: Before fusing the pyramid pooling module with the backbone features, a dual-path attention mechanism is introduced to enhance the network's ability to focus on key sea ice regions. The dual paths are the channel attention branch and the spatial attention branch. First, the channel attention branch performs channel dimension optimization on the feature map output by the lightweight backbone network. The purpose is to learn the importance weights of each channel, strengthen key channels (such as ship outlines, sea ice textures, etc.), and suppress redundant channels (such as image noise). Then, the spatial attention branch optimizes the spatial dimension of the feature map processed by the channel attention branch, resulting in the feature map output by the dual-path attention mechanism. The purpose of spatial dimension optimization is to learn the spatial position weights of the feature map, allowing the network to automatically focus on meaningful spatial regions (such as ships, ice floes, etc.) and suppress background interference (such as sky noise, water reflections, etc.).

[0051] Pyramid pooling module: Performs multi-level pooling on the feature map output by the dual-path attention mechanism. After channel dimensionality reduction and upsampling of each pooling result, it is concatenated with the feature map output by the dual-path attention mechanism to obtain the feature map output by the pyramid pooling module.

[0052] Output layer: This layer integrates the features from the pyramid pooling module and restores the feature maps to the input image size, outputting the image segmentation result. The image segmentation result is a pixel-level class probability tensor, which can be understood as a pixel-level sea ice segmentation map that matches the input image size. Each pixel in the sea ice segmentation map is labeled with a specific segmentation category, namely open ocean, sea ice, ships, and sky.

[0053] S130. Train and optimize the image segmentation model to be trained based on the shipborne sea ice dataset to obtain the target image segmentation model, and realize sea ice image segmentation through the target image segmentation model.

[0054] In this step, the image segmentation model to be trained is trained and optimized using the shipborne sea ice dataset, as detailed below:

[0055] Data loading and augmentation: Read the labeled images and perform data augmentation using random horizontal flipping, brightness jitter (±10%), and Gaussian blur (kernel size 3×3) to improve the model's generalization ability;

[0056] Parameter initialization: The convolutional layer parameters are initialized using a Gaussian distribution with a standard deviation of 0.01, and the batch size is set to 8, and the number of training epochs is set to 100.

[0057] Core Training: The training set data is processed in batches, with the current batch of data input into the model. After passing through a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module, the predicted image segmentation result is output. The segmentation loss for the current batch is calculated by comparing the predicted image segmentation result with the true semantic labels using the cross-entropy loss function. Based on the segmentation loss, the gradients of all trainable parameters are calculated backward from the model's output layer to the input layer. The model parameters are then updated using the calculated gradients through a Stochastic Gradient Descent (SGD) optimizer, completing one parameter iteration. The initial learning rate of this optimizer is 10. -7 The weight decay is 5×10 -4 The momentum is 0.95, and the optimizer's learning rate is dynamically adjusted by decaying to 1 / 10 of the current value every 20 rounds.

[0058] Validation set evaluation and model convergence determination: Every 5 rounds, the evaluation metric of model training is calculated on the validation set, the current evaluation metric is compared with the historical evaluation metric, and the optimal model is saved; when the evaluation metric of the validation set does not improve for 10 consecutive rounds, training is stopped and the final model is output.

[0059] Test set evaluation: Use the test set to calculate evaluation metrics on the final model to evaluate whether the model training has achieved the expected results. If so, the final model is used as the target image segmentation model.

[0060] In one embodiment, the model parameters of the image segmentation model to be trained are optimized using the average intersection-union ratio (AUR) and edge detection accuracy as evaluation metrics. The AUR is calculated by averaging the intersection-union ratios (intersection area / union area) for each of the four semantic labels; a higher AUR indicates better overall segmentation performance. Edge detection accuracy is a quantitative indicator of the degree of overlap between the model's predicted category boundaries and the ground truth labeled boundaries; higher accuracy indicates more precise boundary segmentation.

[0061] Sea ice image segmentation is achieved by using a target image segmentation model. The sea ice image to be segmented is input into the target image segmentation model, and the model can output the image segmentation result, which labels each pixel of the sea ice image with a specific segmentation category, namely open ocean water, sea ice, ships, and sky.

[0062] The technical solution of this invention involves constructing a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels. The semantic labels include at least open ocean water, sea ice, ships, and sky. A training image segmentation model is constructed, comprising at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network employs depthwise separable convolutions with embedded dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the channel dimension-optimized feature map through a spatial attention branch. The training image segmentation model is trained and optimized based on the shipborne sea ice dataset to obtain a target image segmentation model, which is then used to segment sea ice images. This scheme employs a lightweight backbone network combining depthwise separable convolutions and dilated convolutions, enabling the model to be applied in shipboard scenarios with limited computing and storage resources. Through a channel-spatial dual-path attention mechanism, the model adaptively focuses on key regions such as sea ice and thin ice edges, suppressing background interference. Using shipboard sea ice datasets, the model learns domain-specific feature representations. In short, through the synergistic design of lightweight structure, feature enhancement, and scene adaptation, the scheme significantly reduces model complexity while substantially improving sea ice image segmentation accuracy and edge preservation capabilities in complex ice conditions.

[0063] Figure 2 This is a schematic diagram of the model structure of an image segmentation model to be trained according to Embodiment 1 of the present invention. The following is in conjunction with... Figure 2 A detailed explanation of the components included in the model structure is provided below:

[0064] Input layer: The size of the input image is 512×512×3.

[0065] Lightweight backbone network:

[0066] In one embodiment, the lightweight backbone network includes sequentially connected feature extraction layers 1 to 5, with dilated convolutions introduced in the 4th and 5th feature extraction layers.

[0067] Layer 1: Input size 512×512×3, number of repetitions 1, output channels 32;

[0068] Layer 2: Input size 256×256×32, number of repetitions 3, output channels 64;

[0069] Layer 3: Input size 128×128×64, number of repetitions 4, output channels 128;

[0070] Layer 4: Input size 64×64×128, number of repetitions 6, output channels 256;

[0071] Layer 5: Input size 64×64×256, number of repetitions 3, output channels 512.

[0072] Each of the five layers is implemented according to the execution logic of depth-separable convolution, and dilated convolution is embedded in the four and five layers. While maintaining the output feature map size as 1 / 8 of the input image, the receptive field is increased from the traditional 235×235 to 403×403.

[0073] The lightweight backbone network outputs a 64×64×512 feature map F, which can also be represented as a size H×W×C, where H is the height, W is the width, and C is the number of channels.

[0074] Dual-path attention mechanism:

[0075] In one embodiment, the execution steps of the channel attention branch include:

[0076] Global average pooling and global max pooling are performed on the feature maps output by the lightweight backbone network to obtain two channel description vectors.

[0077] Two channel description vectors are input into a two-layer neural network, and the channel weight coefficients are output through a non-linear activation function.

[0078] The channel weight coefficients are multiplied channel by channel by the feature map output by the lightweight backbone network to obtain the feature map with optimized channel dimensions.

[0079] Global average pooling and global max pooling are performed on the feature map F (size H×W×C) output by the lightweight backbone network to obtain two channel description vectors (both with dimensions 1×1×C). These two channel description vectors are then input into a two-layer neural network consisting of a fully connected layer, a Rectified Linear Unit (ReLU) activation function, and another fully connected layer. The channel weight coefficients M are output through a sigmoid activation function. C (Dimensions are 1×1×C); M C Multiplying the original feature map F channel by channel gives the channel attention-optimized feature map F. C .

[0080] In one embodiment, the execution steps of the spatial attention branch include:

[0081] The channel-dimensional optimized feature maps are subjected to channel-dimensional average pooling and max pooling respectively to obtain two spatial description maps;

[0082] Two spatial description maps are concatenated along the channel dimension, and spatial weight coefficients are output through convolutional layers and nonlinear activation functions.

[0083] The spatial weight coefficients are multiplied pixel-by-pixel by the channel dimension-optimized feature map to obtain the spatial dimension-optimized feature map.

[0084] For F C Perform average pooling and max pooling along the channel dimension to obtain two spatial description maps (each with a size of H×W×1). Concatenate the two spatial description maps along the channel dimension to obtain a feature map of size H×W×2. Pass this feature map through a convolutional layer with a kernel size of 7×7 (reducing the number of channels to 1) and a sigmoid activation function to output the spatial weight coefficients M. S (Dimensions are H×W×1); M S With F C Pixel-by-pixel multiplication yields a spatially optimized feature map, i.e., the feature map F optimized by the dual-path attention mechanism. CS .

[0085] Pyramid pooling module:

[0086] In one embodiment, the execution steps of the pyramid pooling module include:

[0087] The feature map output by the dual-path attention mechanism is subjected to four-level pooling. After each pooling branch is reduced in dimensionality by convolution and upsampled, it is concatenated with the feature map output by the dual-path attention mechanism in the channel dimension to output a segmentation feature map.

[0088] For F CS Four levels of pooling are performed (global average pooling, 2×2 grid average pooling, 3×3 grid average pooling, and 6×6 grid average pooling). Each pooling branch is reduced in dimensionality by 1×1 convolution, then upsampled to the original feature map size through bilinear interpolation, and finally compared with F. CS It is spliced ​​and merged at the channel dimension.

[0089] In relation to F CS After channel-level stitching and fusion, the feature map is first integrated through a convolutional layer, and then upsampling is performed to restore the size of the feature map to the size of the original input image. The final output is a segmentation feature map (also known as a sea ice segmentation map), which is the image segmentation result.

[0090] The technical solutions of the embodiments of the present invention can have the following beneficial effects:

[0091] Significantly improved lightweight deployment capability: Through depthwise separable convolution and hierarchical optimization, the number of model parameters has been reduced from 23.5M in the traditional PSPNet to 2.1M, reducing the computational cost by 68%, and it can be directly deployed on shipboard terminals;

[0092] Segmentation accuracy is significantly improved: After incorporating the dual-path attention mechanism, the model's average intersection-union ratio reaches 97.8%, which is 5.8 percentage points higher than the traditional PSPNet. The accuracy of each key category is significantly optimized, with sea ice recognition accuracy improved by 14.7% (solving the problem of confusion between thin ice and waves), open ocean area recognition accuracy improved by 36.7% (reducing missed detections in seawater in broken ice areas), sky recognition accuracy improved by 21% (suppressing cloud reflection interference), and ship hull recognition accuracy improved by 8.9% (preserving small target features).

[0093] Excellent edge detection performance: The edge detection accuracy reaches 90.1%, which is 20.3% higher than the traditional PSPNet. It can accurately identify the boundaries of broken ice and thin ice, and support the risk assessment of dredging operations.

[0094] Real-time inference: The model's inference speed reaches 25 frames per second, meeting the requirements for real-time monitoring and can be directly applied to ice condition monitoring of dredging vessels;

[0095] Highly robust: It exhibits stable performance in complex ice conditions (broken ice, thin ice, cloud interference, etc.), and can be adapted to the dynamic operating environment of dredging engineering vessels, providing core technical support for navigation safety.

[0096] Example 2

[0097] Figure 3 This is a schematic diagram of a sea ice image segmentation device according to Embodiment 2 of the present invention. This embodiment is applicable to situations involving sea ice image segmentation, such as... Figure 3 As shown, the specific structure of the device includes:

[0098] Dataset construction module 31 is used to construct a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels. The semantic labels include at least open ocean waters, sea ice, ships, and sky.

[0099] The model building module 32 is used to build a training image segmentation model, which includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network uses depthwise separable convolutions and embeds dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the feature map after channel dimension optimization through a spatial attention branch.

[0100] The model training and application module 33 is used to train and optimize the image segmentation model to be trained based on the shipborne sea ice dataset to obtain the target image segmentation model, and to achieve sea ice image segmentation through the target image segmentation model.

[0101] The sea ice image segmentation device provided in this embodiment constructs a shipborne sea ice dataset through a dataset construction module. This dataset consists of sea ice images and pixel-level semantic labels, which include at least open ocean water, sea ice, ships, and sky. A model construction module constructs a training image segmentation model, which includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network employs depthwise separable convolutions with embedded dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the channel dimension-optimized feature map through a spatial attention branch. The model training and application module trains and optimizes the training image segmentation model based on the shipborne sea ice dataset to obtain a target image segmentation model, which is then used to segment sea ice images. This scheme employs a lightweight backbone network combining depthwise separable convolutions and dilated convolutions, enabling the model to be applied in shipboard scenarios with limited computing and storage resources. Through a channel-spatial dual-path attention mechanism, the model adaptively focuses on key regions such as sea ice and thin ice edges, suppressing background interference. Using shipboard sea ice datasets, the model learns domain-specific feature representations. In short, through the synergistic design of lightweight structure, feature enhancement, and scene adaptation, the scheme significantly reduces model complexity while substantially improving sea ice image segmentation accuracy and edge preservation capabilities in complex ice conditions.

[0102] Furthermore, the lightweight backbone network includes sequentially connected feature extraction layers 1 to 5, with dilated convolutions introduced in the 4th and 5th feature extraction layers.

[0103] Furthermore, the execution steps of the channel attention branch include:

[0104] Global average pooling and global max pooling are performed on the feature maps output by the lightweight backbone network to obtain two channel description vectors.

[0105] Two channel description vectors are input into a two-layer neural network, and the channel weight coefficients are output through a non-linear activation function.

[0106] The channel weight coefficients are multiplied channel by channel by the feature map output by the lightweight backbone network to obtain the feature map with optimized channel dimensions.

[0107] Furthermore, the execution steps of the spatial attention branch include:

[0108] The channel-dimensional optimized feature maps are subjected to channel-dimensional average pooling and max pooling respectively to obtain two spatial description maps;

[0109] Two spatial description maps are concatenated along the channel dimension, and spatial weight coefficients are output through convolutional layers and nonlinear activation functions.

[0110] The spatial weight coefficients are multiplied pixel-by-pixel by the channel dimension-optimized feature map to obtain the spatial dimension-optimized feature map.

[0111] Furthermore, the execution steps of the pyramid pooling module include:

[0112] The feature map output by the dual-path attention mechanism is subjected to four-level pooling. After each pooling branch is reduced in dimensionality by convolution and upsampled, it is concatenated with the feature map output by the dual-path attention mechanism in the channel dimension to output a segmentation feature map.

[0113] Furthermore, the model parameters of the image segmentation model to be trained are optimized using the average intersection-union ratio and edge detection accuracy as evaluation indicators.

[0114] The sea ice image segmentation apparatus provided in this embodiment of the invention can execute the sea ice image segmentation method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0115] Example 3

[0116] Figure 4 This is a schematic diagram of the structure of an electronic device implementing embodiments of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0117] like Figure 4 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 performs various appropriate actions and processes based on the computer programs stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0118] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0119] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as sea ice image segmentation methods.

[0120] In some embodiments, the sea ice image segmentation method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or mounted on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the sea ice image segmentation method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the sea ice image segmentation method by any other suitable means (e.g., by means of firmware).

[0121] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0122] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0123] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0124] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0125] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0126] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0127] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0128] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A sea ice image segmentation method, characterized in that, include: Construct a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels, the semantic labels including at least open ocean waters, sea ice, ships, and sky; A training image segmentation model is constructed, which includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network employs depthwise separable convolutions and embeds dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the feature map after channel dimension optimization through a spatial attention branch. The image segmentation model to be trained is trained and optimized based on the shipborne sea ice dataset to obtain the target image segmentation model, and the sea ice image segmentation is achieved through the target image segmentation model.

2. The method according to claim 1, characterized in that, The lightweight backbone network includes 1 to 5 feature extraction layers connected in sequence, with dilated convolutions introduced in the 4th and 5th feature extraction layers.

3. The method according to claim 1, characterized in that, The execution steps of the channel attention branch include: Global average pooling and global max pooling are performed on the feature maps output by the lightweight backbone network to obtain two channel description vectors. Two channel description vectors are input into a two-layer neural network, and the channel weight coefficients are output through a non-linear activation function. The channel weight coefficients are multiplied channel by channel by the feature map output by the lightweight backbone network to obtain the feature map with optimized channel dimensions.

4. The method according to claim 1, characterized in that, The execution steps of the spatial attention branch include: The channel-dimensional optimized feature maps are subjected to channel-dimensional average pooling and max pooling respectively to obtain two spatial description maps; Two spatial description maps are concatenated along the channel dimension, and spatial weight coefficients are output through convolutional layers and nonlinear activation functions. The spatial weight coefficients are multiplied pixel-by-pixel by the channel dimension-optimized feature map to obtain the spatial dimension-optimized feature map.

5. The method according to claim 1, characterized in that, The execution steps of the pyramid pooling module include: The feature map output by the dual-path attention mechanism is subjected to four-level pooling. After each pooling branch is reduced in dimensionality by convolution and upsampled, it is concatenated with the feature map output by the dual-path attention mechanism in the channel dimension to output a segmentation feature map.

6. The method according to claim 1, characterized in that, The model parameters of the image segmentation model to be trained are optimized using the average intersection-union ratio and edge detection accuracy as evaluation indicators.

7. A sea ice image segmentation device, characterized in that, include: The dataset construction module is used to construct a shipborne sea ice dataset, which consists of sea ice images and pixel-level semantic labels. The semantic labels include at least open ocean waters, sea ice, ships, and sky. The model building module is used to build a training image segmentation model, which includes at least a lightweight backbone network, a dual-path attention mechanism, and a pyramid pooling module. The lightweight backbone network uses depthwise separable convolutions and embeds dilated convolutions. In the dual-path attention mechanism, channel dimension optimization is performed on the feature map output by the lightweight backbone network through a channel attention branch, and spatial dimension optimization is performed on the feature map after channel dimension optimization through a spatial attention branch. The model training and application module is used to train and optimize the image segmentation model to be trained based on the shipborne sea ice dataset, obtain the target image segmentation model, and realize sea ice image segmentation through the target image segmentation model.

8. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.

Images