A sea ice sample augmentation method, system, and apparatus

Abstract

The application provides a sea ice sample augmentation method, system and device, relates to the cross field of sea ice sample processing and computer vision, and the method comprises the steps of: labeling and preprocessing a ship-based sea ice image to obtain a standardized real sea ice scene graph and a corresponding sea ice label graph; constructing an improved conditional generative adversarial network comprising a generator and a double-head discriminator, wherein the double-head discriminator comprises a spatial domain discriminator and a wavelet domain discriminator; inputting the sea ice label graph into the generator to generate a sea ice scene graph; combining the generated graph and the label graph into a first input group, combining the real graph and the label graph into a second input group, inputting the two input groups into the two discriminators respectively to obtain spatial domain and wavelet domain discrimination results; generating a composite loss function according to the discrimination results, updating the parameters through back propagation; repeating the process until the loss converges to obtain a trained model; and inputting a sea ice label graph to be augmented into the model to output a sea ice augmented sample by the generator. The application can generate high-quality sea ice samples with global structure and detailed texture, and can alleviate the shortage of samples.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of sea ice sample processing and computer vision technology, and more specifically, to a method, system, and device for sea ice sample augmentation. Background Technology

[0002] Sea ice sample augmentation is a key technological step supporting research in polar sea ice monitoring, sea ice identification, and thickness inversion. Its core objective is to generate realistic, diverse, and usable sea ice samples under constraints such as sample scarcity, class imbalance, and complex detailed features, thereby alleviating the difficulty of obtaining measured data. Currently, the industry mainly relies on methods based on traditional image geometric transformations, such as rotation, translation, scaling, and flipping; or generative adversarial network augmentation methods based on a single discriminator, such as conditional generative adversarial networks.

[0003] In related technologies, traditional geometric transformation methods can only generate variants with low dissimilarity, resulting in limited augmentation effects and a negligible improvement in model generalization ability. Meanwhile, while conditional generation networks based on a single discriminator can learn sample distribution from labeled information, the single discriminator can only judge the overall realism of the generated samples, easily leading to class confusion and sample distortion. Furthermore, existing generation strategies often struggle to simultaneously capture the global structure and subtle textures of sea ice, resulting in limited scene representation of generated samples and a lack of explicit modeling of sample diversity, leading to insufficient generalization ability of augmentation strategies. Summary of the Invention

[0004] The problem addressed by this invention is how to generate high-quality sea ice augmentation samples that combine global structure with detailed texture.

[0005] To address the above problems, this invention provides a method, system, and device for sea ice sample augmentation.

[0006] In a first aspect, the present invention provides a method for augmenting sea ice samples, comprising: S1: The collected images of real sea ice scenes on the ship base are labeled and preprocessed to obtain standardized real sea ice scene images and corresponding standardized sea ice labeled images; S2: Construct an improved conditional generative adversarial network, which includes a generator and a dual-head discriminator; the dual-head discriminator includes a spatial domain discriminator and a wavelet domain discriminator; S3: Use the standardized sea ice annotation map as the input to the generator, which adopts a U-Net structure and introduces a CBSAF channel-spatial dual collaborative enhancement module at the jump connection of the U-Net structure. Under the collaborative enhancement of the CBSAF channel-spatial dual collaborative enhancement module, the generator outputs the generated sea ice scene map; S4: Combine the generated sea ice scene image with the corresponding standardized sea ice annotation image to form a first input group, combine the standardized real sea ice scene image with the corresponding standardized sea ice annotation image to form a second input group, and input the first input group and the second input group into the spatial domain discriminator and the wavelet domain discriminator respectively. The spatial domain discriminator is used to discriminate the spatial structure and texture details of the received first input group and the second input group; the wavelet domain discriminator is used to perform multi-scale frequency domain decomposition on the received first input group and the second input group through DB4 discrete wavelet transform, extract frequency domain features, and then discriminate to obtain spatial domain discrimination results and wavelet domain discrimination results. S5: Generate a composite loss function based on the spatial domain discrimination result and the wavelet domain discrimination result, and update the network parameters of the generator and the dual-head discriminator through backpropagation; S6: Repeat S3 to S5 until the composite loss function converges to obtain the trained sea ice sample augmentation model; S7: Input the sea ice annotation map to be augmented into the sea ice sample augmentation model, and the generator outputs the corresponding sea ice augmented sample.

[0007] Optionally, the step of classifying and preprocessing the collected ship-based sea ice scene images to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images includes: The sea ice categories in the collected real-world sea ice scene images of the ship base are labeled to obtain a sea ice label map; The real sea ice scene image and the sea ice annotation image are sequentially subjected to denoising, size unification and normalization processes to obtain the standardized real sea ice scene image and the corresponding standardized sea ice annotation image; and the standardized real sea ice scene image and the corresponding standardized sea ice annotation image are divided into training set, validation set and test set according to a preset ratio.

[0008] Optionally, the standardized sea ice map is used as input to the generator. The generator employs a U-Net structure and introduces a CBSAF channel-spatial dual collaborative enhancement module at the jump connection of the U-Net structure. Under the collaborative enhancement of the CBSAF channel-spatial dual collaborative enhancement module, the generator outputs a generated sea ice scene map, including: The standardized sea ice map is input into the encoder of the generator, and multi-scale features are extracted through downsampling; At the U-Net skip connection of the generator, the multi-scale features are enhanced in both channel and spatial dimensions through the CBSAF channel-space dual collaborative enhancement module to obtain the enhanced multi-scale features; The enhanced multi-scale features are input into the decoder of the generator, and after upsampling and reconstruction, the generated sea ice scene map is output.

[0009] Optionally, the step of combining the generated sea ice scene image with the corresponding standardized sea ice annotation image into a first input group, combining the standardized real sea ice scene image with the corresponding standardized sea ice annotation image into a second input group, and inputting the first input group and the second input group into the spatial domain discriminator and the wavelet domain discriminator respectively, wherein the spatial domain discriminator is used to discriminate the spatial structure and texture details of the received first input group and the second input group; the wavelet domain discriminator is used to perform multi-scale frequency domain decomposition on the received first input group and the second input group through DB4 discrete wavelet transform, extract frequency domain features, and then discriminate to obtain spatial domain discrimination results and wavelet domain discrimination results, including: The generated sea ice scene image and the corresponding standardized sea ice annotation image are stitched together along the channel dimension to obtain the first input group; the standardized real sea ice scene image and the corresponding standardized sea ice annotation image are stitched together along the channel dimension to obtain the second input group. The first input group and the second input group are respectively input into the spatial domain discriminator, and the spatial domain discriminator outputs the spatial domain discrimination result indicating the authenticity of the input; The first input group and the second input group are respectively input into the wavelet domain discriminator, and the wavelet domain discriminator outputs the wavelet domain discrimination result indicating the authenticity of the input.

[0010] Optionally, the step of generating a composite loss function based on the spatial domain discrimination result and the wavelet domain discrimination result, and updating the network parameters of the generator and the dual-head discriminator through backpropagation, includes: The adversarial loss is determined based on the spatial domain discrimination result and the wavelet domain discrimination result; L1 loss and perception loss are determined based on the generated sea ice scene map and the corresponding standardized real sea ice scene map. The adversarial loss, the L1 loss, and the perceptual loss are weighted and combined to obtain the composite loss function; Based on the composite loss function, the network parameters of the generator and the network parameters of the dual-head discriminator are updated respectively through the backpropagation algorithm.

[0011] Optionally, repeating steps S3 to S5 until the composite loss function converges to obtain the trained sea ice sample augmentation model includes: In each iteration, S3 to S5 are executed sequentially, and after each iteration, it is determined whether the composite loss function has converged. When the value of the composite loss function is less than or equal to a preset threshold, the iteration stops, and the generator and the dual-head discriminator under the current network parameters are used as the trained sea ice sample augmentation model.

[0012] Optionally, the step of inputting the sea ice annotation map to be augmented into the sea ice sample augmentation model, and having the generator output the corresponding augmented sea ice samples, includes: The sea ice annotation map to be augmented is input into the trained sea ice sample augmentation model, and the generator in the sea ice sample augmentation model is called. The generator outputs the corresponding sea ice augmentation sample based on the sea ice annotation map to be augmented.

[0013] Secondly, the present invention provides a sea ice sample augmentation system, comprising: The data processing unit is used to classify and preprocess the collected ship-based real sea ice scene images to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images. A network construction unit is used to construct an improved conditional generative adversarial network, wherein the improved conditional generative adversarial network includes a generator and a dual-head discriminator; the dual-head discriminator includes a spatial domain discriminator and a wavelet domain discriminator. The adversarial training unit is configured to use the standardized sea ice map as input to the generator, which employs a U-Net architecture and introduces a CBSAF channel-space dual collaborative enhancement module at the skip connections of the U-Net architecture. Under the collaborative enhancement of the CBSAF channel-space dual collaborative enhancement module, the generator outputs a generated sea ice scene map. It is also configured to combine the generated sea ice scene map with the corresponding standardized sea ice map as a first input group, and combine the standardized real sea ice scene map with the corresponding standardized sea ice map as a second input group, and input the first input group and the second input group respectively into the training unit. The method describes a spatial domain discriminator, which inputs the first input group and the second input group into the wavelet domain discriminator respectively. The spatial domain discriminator is used to discriminate the spatial structure and texture details of the received first input group and the second input group. The wavelet domain discriminator is used to perform multi-scale frequency domain decomposition on the received first input group and the second input group through DB4 discrete wavelet transform, extract frequency domain features, and then discriminate to obtain spatial domain discrimination results and wavelet domain discrimination results. It is also used to generate a composite loss function based on the spatial domain discrimination results and the wavelet domain discrimination results, and update the network parameters of the generator and the dual-head discriminator through backpropagation. An iterative training unit is used to repeatedly execute the processing of the adversarial training unit until the composite loss function converges, thereby obtaining a trained sea ice sample augmentation model. The sample generation unit is used to input the sea ice annotation map to be augmented into the sea ice sample augmentation model, and the generator outputs the corresponding sea ice augmented sample.

[0014] Thirdly, an electronic device according to the present invention includes a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the sea ice sample augmentation method as described above when executing the computer program.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described sea ice sample augmentation method.

[0016] The sea ice sample augmentation method, system, and device of this invention first perform category labeling and preprocessing on ship-based real sea ice scene images, such as denoising, size unification, and normalization, to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images. This normalization operation eliminates scale and grayscale differences between original images, enabling the subsequent network to stably learn the mapping relationship from annotation to scene. Simultaneously, using the annotation image as a conditional input to the generator provides explicit category space constraints for the generation process, avoiding the category confusion problem in unconditional generation. Unlike traditional geometric transformations that only produce low-discrepancy variants, the annotation condition information of this invention can guide the generator to produce new samples that conform to the sea ice category distribution, improving the category accuracy of augmented samples from the source.

[0017] Furthermore, a dual-head discriminator structure is adopted, comprising a spatial domain discriminator and a wavelet domain discriminator. The spatial domain discriminator directly discriminates the spatial structure and texture details of the input group (sea ice scene image + labeled image), making the sea ice scene image output by the generator approximate the real image in terms of overall layout, ice surface morphology, and other macroscopic features. The wavelet domain discriminator, through multi-scale frequency domain decomposition, focuses on capturing high-frequency detail features such as ice cracks and ice crevices, and uses this to determine authenticity. The two discriminators complement each other; the spatial domain discriminator ensures that the sample does not exhibit overall deformation, while the wavelet domain discriminator ensures that the sample does not lose subtle textures. Existing single discriminators can only judge overall authenticity, easily overlooking local details or leading to category confusion. This invention, through parallel discrimination by the dual-head discriminator, subjectes the generator to the dual pressure of global structure and local details simultaneously. As a result, the generated sea ice augmented samples not only have realistic macroscopic morphology but also retain key recognition features such as ice cracks and ice crevices, significantly improving the detail authenticity and usability of the samples.

[0018] Furthermore, the generator takes standardized sea ice annotation maps as input and outputs generated sea ice scene maps through an encoder-decoder structure. Combined with adversarial training using a dual-head discriminator, a stable feature learning loop is formed. In each iteration, the generator generates a scene map based on the annotation map, then combines it with the real scene map to form the first and second input groups, respectively, which are then input to two discriminators. The two discriminators output spatial domain discrimination results and wavelet domain discrimination results. Based on these, a composite loss function (including adversarial loss, L1 loss, and perceptual loss) is determined and simultaneously updates the network parameters of the generator and the dual-head discriminator through backpropagation. This process is repeated until the composite loss function converges. In this invention, the two discriminators provide two independent discrimination feedback paths, making the gradient update direction of the generator more accurate, the training process more stable, and effectively avoiding mode collapse or convergence oscillations. The generation-discrimination-update step is repeated until the loss converges, resulting in a trained sea ice sample augmentation model. During the inference stage, only the generator is needed to quickly output augmented samples. After training, the generator in the model has learned to map corresponding realistic sea ice scene maps from any sea ice annotation map. In actual augmentation, simply inputting the sea ice annotation map to be augmented into the generator will yield high-quality sea ice augmentation samples in one go, eliminating the need for a discriminator. This effectively reduces the computational overhead of the inference stage and enables the generation of diverse sea ice samples on a large scale, effectively alleviating the problems of scarce measured data and class imbalance. By generating samples with new texture features (such as ice patterns with different orientations and ice gaps with different widths), this invention significantly improves the diversity of the dataset and the generalization ability of subsequent recognition models.

[0019] Finally, this invention forms an end-to-end automated process from data preprocessing, network construction, adversarial training to sample generation. Throughout the process, a standardized sea ice annotation map serves as the sole input, and all discrimination and updates are performed based on a deterministic loss function. Different operators can obtain consistent augmented models using the same training data. Compared to traditional methods that rely on manual experience to adjust geometric transformation parameters or subjectively select generated samples, this invention provides an objective and reproducible augmentation scheme, facilitating deployment and application in practical scenarios such as polar sea ice monitoring and marine engineering.

[0020] In summary, this invention effectively solves the problems of sample distortion, missing details, category confusion, and limited scene in existing sea ice sample augmentation methods by using dual discrimination in the spatial and wavelet domains of a dual-head discriminator, conditional guidance of standardized annotation maps, multi-dimensional constraints of composite loss functions, and closed-loop adversarial training. It achieves the generation of high-quality sea ice samples with both global structure and detailed texture, providing reliable data support for deep learning-based sea ice identification and analysis tasks. Attached Figure Description

[0021] Figure 1This is a flowchart illustrating the sea ice sample augmentation method in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the sea ice sample augmentation system in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device in an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0023] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0024] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0025] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0026] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0027] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties. The collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0028] Combination Figure 1 As shown, this embodiment of the invention provides a method for augmenting sea ice samples, comprising: S1: The collected images of real sea ice scenes on the ship base are labeled and preprocessed to obtain standardized real sea ice scene images and corresponding standardized sea ice labeled images.

[0029] Specifically, the collected ship-based real-world sea ice scene images are categorized and preprocessed. The acquisition equipment can be a shipborne high-definition camera, capturing sea ice surface images during polar voyages. During annotation, professionals use image annotation tools (such as LabelMe) to perform pixel-level segmentation and annotation of sea ice categories in the images, distinguishing different sea ice types (such as flat ice, accumulated ice, melt ice, etc.), generating grayscale annotation maps with the same size as the original images, where different grayscale values ​​represent different categories. The preprocessing operations include: removing salt-and-pepper noise from the images using median filtering; uniformly scaling all real-world scene images and their annotation maps to a size of 256×256 pixels; and finally, normalizing the pixel values, linearly mapping the grayscale range from 0 to 255 to the [-1, 1] interval, obtaining standardized real-world sea ice scene images and corresponding standardized sea ice annotation maps. Finally, the standardized image pairs are divided into training, validation, and test sets according to a preset ratio for subsequent network training and performance verification, such as dividing the standardized image pairs into training, validation, and test sets in a 7:2:1 ratio.

[0030] S2: Construct an improved conditional generative adversarial network, which includes a generator and a dual-headed discriminator; the dual-headed discriminator includes a spatial domain discriminator and a wavelet domain discriminator.

[0031] Specifically, the improved conditional generative adversarial network in this embodiment is an improvement on the original Pix2Pix network framework. The generator adopts a U-Net structure and adds a CBSAF channel-spatial dual collaborative enhancement module at its skip connections. The CBSAF module adopts a dual-branch parallel processing architecture: the channel branch introduces the channel attention submodule of CBAM, performs global average pooling and global max pooling on the input features respectively, and feeds the resulting feature vectors into a shared multilayer perceptron for processing; the spatial branch adopts the spatial attention mechanism in CASAB, simultaneously aggregating four statistics of the input feature map in the channel dimension: mean, maximum, minimum, and sum. The descriptor is fed into a two-layer convolutional encoder. The first layer uses a 7×7 large-scale convolutional kernel and is activated by Swish. The second layer uses a 1×1 convolution to compress the features to a single channel. Finally, a spatial weight heatmap is generated by Sigmoid. The outputs of the two branches adopt a gated fusion strategy for nonlinear adaptive fusion, and the enhanced features are fused with the original features through residual connections. The dual-head discriminator comprises a spatial domain discriminator and a wavelet domain discriminator: the spatial domain discriminator adopts the original Pix2Pix spatial domain discriminator structure to compare and distinguish the spatial structure and texture details of generated and real sea ice images; the wavelet domain discriminator first performs multi-scale frequency domain decomposition on the generated and real sea ice images through DB4 discrete wavelet transform, extracts frequency domain features, and then performs discrimination to enhance the ability to capture subtle texture features of sea ice. Simultaneously, perceptual loss is introduced into the loss function.

[0032] In a preferred embodiment of the present invention, a CBSAF channel-space dual collaborative enhancement module is added at the skip connections of the U-shaped network: CBSAF adopts a dual-branch parallel processing architecture, which can simultaneously capture channel importance and spatial key region information. Specifically, the channel branch of the CBSAF channel-space dual collaborative enhancement module introduces the channel attention submodule from CBAM (Convolutional Block Attention Module). CBAM is a lightweight hybrid attention mechanism that infers attention maps sequentially along two independent dimensions, channel and space, to achieve adaptive feature refinement. The channel attention submodule used in this embodiment can perform global average pooling and global max pooling on the input features respectively, and feed the resulting feature vectors into a shared multilayer perceptron (MLP) for processing, thereby dynamically adjusting the response intensity of each channel. In this embodiment, this channel attention submodule is simply referred to as CBM (Channel Branch Module), and its calculation formula is as follows: ; in, The input feature map is typically the feature tensor output by a convolutional layer, with a shape of H×W×C, where H and W are the spatial dimensions, and C is the number of channels. In the channel branches of the CBSAF module... This refers to the feature map input to that branch; This represents the Sigmoid activation function; This refers to a multilayer perceptron. Perform global average pooling on the input feature map and output a vector of length C, representing the global average response of each channel; Perform global max pooling on the input feature map and output a vector of length C, representing the global maximum response of each channel.

[0033] In the spatial branch, CBSAF adopts the spatial attention mechanism from CASAB. This mechanism abandons the simplified strategy of using only mean and max pooling in the traditional approach, instead simultaneously aggregating four key statistics of the input feature map in the channel dimension: mean, maximum, minimum, and sum. Subsequently, this descriptor is fed into a two-layer convolutional encoder for depth processing. The first layer uses a large-scale 7×7 convolutional kernel, and the output of this convolution is passed through a Swish activation function to introduce smooth and non-saturating nonlinear characteristics. The second layer uses a 1×1 convolution as a projection layer to compress the features to a single channel, and finally generates a high-precision spatial weight heatmap using the Sigmoid function. The SAM formula is as follows: ; in, Input feature map; Use the Sigmoid activation function; It is a 1×1 convolutional layer; For activation functions; It is a 7×7 convolutional layer; Input feature map The multi-channel descriptor is obtained by aggregating four statistics (mean, maximum, minimum, and sum) along the channel dimension.

[0034] Furthermore, this module employs a gated fusion strategy to achieve more refined and adaptive nonlinear fusion of sea ice features. Finally, the enhanced features are fused with the original features through residual connections.

[0035] S3: Using the standardized sea ice annotation map as input to the generator, the generator outputs a generated sea ice scene map. Specifically, the standardized sea ice annotation images in the training set are input to the constructed U-Net generator. Based on this constraint, the generator learns the feature distribution of the real sea ice scene and outputs the corresponding generated sea ice scene map.

[0036] S4: Combine the generated sea ice scene map with the corresponding standardized sea ice annotation map into a first input group, combine the standardized real sea ice scene map with the corresponding standardized sea ice annotation map into a second input group, and input the first input group and the second input group into the spatial domain discriminator respectively, and input the first input group and the second input group into the wavelet domain discriminator respectively, to obtain the spatial domain discrimination result and the wavelet domain discrimination result.

[0037] Specifically, the generated sea ice scene image and its corresponding standardized sea ice annotation image are combined into a first input group, and the standardized real sea ice scene image and its corresponding standardized sea ice annotation image are combined into a second input group. Then, the first and second input groups are each input to a spatial domain discriminator, and simultaneously, they are each input to a wavelet domain discriminator. The spatial domain discriminator outputs a spatial domain discrimination result indicating the authenticity of the input, and the wavelet domain discriminator outputs a wavelet domain discrimination result indicating the authenticity of the input. For example, in each training iteration, the RGB sea ice scene image (256×256×3) generated in step S3 and its corresponding standardized sea ice annotation image (256×256×1) are concatenated along the channel dimension to form a 4-channel tensor, which serves as the first input group. Similarly, the standardized real sea ice scene image (256×256×3) and its corresponding standardized sea ice annotation image (256×256×1) in the training set are also concatenated into a 4-channel tensor, which serves as the second input group. Then, the first and second input groups are fed into the spatial domain discriminator. This spatial domain discriminator uses a PatchGAN structure, outputting a true / false probability for each 70×70 local region of the input image, and finally outputting a 30×30 feature matrix. The mean of the matrix is ​​taken as the spatial domain discrimination result. Simultaneously, the first and second input groups are fed into the wavelet domain discriminator. The wavelet domain discriminator first performs a DB4 discrete wavelet transform on each input group, decomposing it into low-frequency approximate components and high-frequency detail components in the horizontal, vertical, and diagonal directions, each component with a size of 128×128×1. These four components are then stacked in the channel dimension to form a 128×128×4 feature map, which is then passed through three convolutional layers (3×3 kernel, stride 2, output channels 32, 64, and 1 respectively). Finally, a scalar value is output after Sigmoid activation as the wavelet domain discrimination result. The outputs of the two discriminators are used together to construct the subsequent loss function.

[0038] S5: Generate a composite loss function based on the spatial domain discrimination result and the wavelet domain discrimination result, and update the network parameters of the generator and the dual-head discriminator through backpropagation.

[0039] Specifically, the composite loss function combines adversarial loss, L1 loss, and perceptual loss. The adversarial loss constrains the adversarial learning between the generator and the dual-head discriminator, enabling the generator to produce sea ice images closer to reality. The L1 loss reduces the pixel-level differences between the generated and real sea ice images. The perceptual loss measures the similarity between the generated and real sea ice images at the feature level, optimizing the structural rationality and visual quality of the generated images. After determining the composite loss function value based on the spatial domain and wavelet domain discrimination results, the network parameters of the generator and the dual-head discriminator are updated using the backpropagation algorithm.

[0040] In a preferred embodiment of the present invention, the loss function ; in, The total value of the composite loss function, , , They are respectively The weighting coefficients for loss, adversarial loss, and perceptual loss, where L1 represents L1 loss. Indicating resistance to loss, This indicates perceived loss.

[0041] S6: Repeat S3 to S5 until the composite loss function converges to obtain the trained sea ice sample augmentation model.

[0042] Specifically, in each iteration, the following steps are executed sequentially: generating a sea ice scene image, combining input groups and inputting them into a discriminator to obtain the discrimination result, calculating the composite loss, and updating the parameters. The iteration stops when the composite loss function value converges and the network performance stabilizes, the current network parameters are saved, and the trained sea ice sample augmentation model is obtained.

[0043] S7: Input the sea ice annotation map to be augmented into the sea ice sample augmentation model, and the generator outputs the corresponding sea ice augmented sample.

[0044] Specifically, the annotated sea ice images to be augmented are acquired and input into the U-Net generator in the trained sea ice sample augmentation model. The generator directly outputs the corresponding augmented sea ice samples. The quality of the generated sea ice images is validated using validation and test sets, and augmented samples that meet the requirements are selected to supplement the sea ice sample dataset, thereby alleviating the problems of sample scarcity and class imbalance.

[0045] The sea ice sample augmentation method in this embodiment first performs category labeling and preprocessing on ship-based real sea ice scene images, such as denoising, size unification, and normalization, to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images. This normalization operation eliminates scale and grayscale differences between original images, enabling the subsequent network to stably learn the mapping relationship from annotation to scene. Simultaneously, using the annotation image as a conditional input to the generator provides explicit category space constraints for the generation process, avoiding category confusion problems in unconditional generation. Unlike traditional geometric transformations that only produce low-dissimilarity variants, the annotation condition information of this invention can guide the generator to produce new samples that conform to the sea ice category distribution, improving the category accuracy of augmented samples from the source.

[0046] Furthermore, a dual-head discriminator structure is adopted, comprising a spatial domain discriminator and a wavelet domain discriminator. The spatial domain discriminator directly discriminates the spatial structure and texture details of the input group (sea ice scene image + labeled image), making the sea ice scene image output by the generator approximate the real image in terms of overall layout, ice surface morphology, and other macroscopic features. The wavelet domain discriminator, through multi-scale frequency domain decomposition, focuses on capturing high-frequency detail features such as ice cracks and ice crevices, and uses this to determine authenticity. The two discriminators complement each other; the spatial domain discriminator ensures that the sample does not exhibit overall deformation, while the wavelet domain discriminator ensures that the sample does not lose subtle textures. Existing single discriminators can only judge overall authenticity, easily ignoring local details or leading to category confusion. This embodiment, through parallel discrimination by the dual-head discriminator, subjectes the generator to the dual pressure of global structure and local details simultaneously. As a result, the generated sea ice augmented samples not only have realistic macroscopic morphology but also retain key recognition features such as ice cracks and ice crevices, significantly improving the detail authenticity and usability of the samples.

[0047] Furthermore, the generator takes standardized sea ice annotation maps as input and outputs generated sea ice scene maps through an encoder-decoder structure. Combined with adversarial training using a dual-head discriminator, a stable feature learning loop is formed. In each iteration, the generator generates a scene map based on the annotation map, then combines it with the real scene map to form the first and second input groups, respectively, which are then input to two discriminators. The two discriminators output spatial domain discrimination results and wavelet domain discrimination results. Based on these, a composite loss function (including adversarial loss, L1 loss, and perceptual loss) is determined and simultaneously updates the network parameters of the generator and the dual-head discriminator through backpropagation. This process is repeated until the composite loss function converges. In this embodiment, the two discriminators provide two independent discrimination feedback paths, making the gradient update direction of the generator more accurate, the training process more stable, and effectively avoiding mode collapse or convergence oscillations. The generation-discrimination-update step is repeated until the loss converges, resulting in a trained sea ice sample augmentation model. During the inference stage, only the generator is needed to quickly output augmented samples. After training, the generator in the model has learned to map corresponding realistic sea ice scene maps from any sea ice annotation map. In actual augmentation, simply input the sea ice annotation map to be augmented into the generator to obtain high-quality sea ice augmentation samples in one go, without the need for a discriminator. This effectively reduces the computational overhead of the inference stage and enables the generation of diverse sea ice samples on a large scale, effectively alleviating the problems of scarce measured data and class imbalance. In this embodiment, by generating samples with new texture features (such as ice patterns with different orientations and ice gaps with different widths), the diversity of the dataset and the generalization ability of the subsequent recognition model are significantly improved.

[0048] Finally, this embodiment presents an end-to-end automated process from data preprocessing, network construction, adversarial training to sample generation. Throughout the process, a standardized sea ice annotation map serves as the sole input, and all discrimination and updates are performed based on a deterministic loss function. Different operators can obtain consistent augmented models using the same training data. Compared to traditional methods that rely on manual experience to adjust geometric transformation parameters or subjectively select samples for generation, this embodiment provides an objective and reproducible augmentation scheme, facilitating deployment and application in practical scenarios such as polar sea ice monitoring and marine engineering.

[0049] In summary, this embodiment effectively solves the problems of sample distortion, missing details, category confusion, and single scene in existing sea ice sample augmentation methods by using dual discrimination in the spatial domain and wavelet domain of a dual-head discriminator, conditional guidance of standardized annotation maps, multi-dimensional constraints of composite loss functions, and closed-loop adversarial training. It achieves the generation of high-quality sea ice samples with both global structure and detailed texture, providing reliable data support for deep learning-based sea ice identification and analysis tasks.

[0050] Optionally, the step of classifying and preprocessing the collected ship-based sea ice scene images to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images includes: The sea ice categories in the collected real-world sea ice scene images of the ship base are labeled to obtain a sea ice label map; The real sea ice scene image and the sea ice annotation image are sequentially subjected to denoising, size unification and normalization processes to obtain the standardized real sea ice scene image and the corresponding standardized sea ice annotation image; and the standardized real sea ice scene image and the corresponding standardized sea ice annotation image are divided into training set, validation set and test set according to a preset ratio.

[0051] Specifically, the data acquisition equipment can use a shipborne high-definition camera to acquire sea ice surface images during polar voyages. During annotation, professional annotators use image annotation tools (such as LabelMe) to perform pixel-level segmentation and annotation of sea ice categories in the images, distinguishing different types such as flat ice, accumulated ice, and melting ice, generating grayscale annotation maps with the same size as the original images, where different grayscale values ​​represent different categories. Subsequently, the real sea ice scene images and their corresponding annotation maps are preprocessed: first, median filtering or Gaussian filtering is used to remove salt-and-pepper noise generated during image acquisition; then, all images are uniformly scaled to a fixed size, such as 256×256 pixels, to ensure consistency of subsequent network input; finally, the pixel values ​​are normalized, linearly mapping the grayscale range from 0 to 255 to the interval [-1, 1] or [0, 1], resulting in standardized real sea ice scene images and corresponding standardized sea ice annotation maps. After completing the above processing, the standardized image pairs are divided into training set, validation set and test set according to a preset ratio, for example, a ratio of 7:2:1 or 8:1:1, for network training, hyperparameter tuning and final performance evaluation, respectively.

[0052] In this optional embodiment, pixel-level category annotation is performed on the collected real-world sea ice scene images to obtain a sea ice annotation map precisely aligned with the original image. This annotation map serves as the conditional input for the subsequent generator, providing clear category space constraints for the generation process. This allows the generator to learn a deterministic mapping from annotation to scene, fundamentally avoiding the category confusion problem that easily occurs in unconditional generation. Based on this, the real-world scene image and its annotation map are sequentially denoised, resized, and normalized. Denoising eliminates interference such as salt-and-pepper noise introduced by image acquisition in harsh polar environments. Resizing ensures the consistency of input tensors during batch training. Normalization maps pixel values ​​to a fixed range, effectively avoiding gradient instability caused by grayscale differences between different images, thus providing a data foundation for the stable convergence of the deep network. Finally, the dataset is divided into training, validation, and test sets according to a preset ratio. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and overfitting monitoring, and the test set is used to independently evaluate the quality of generated samples. The independence of these three sets ensures the objectivity of model performance evaluation and the reliability of generalization ability. Through standardized preprocessing and supervised annotation, a high-quality and highly consistent data foundation was built for the subsequent adversarial training of the dual-head discriminator.

[0053] Optionally, the construction of the improved conditional generative adversarial network includes a generator and a dual-headed discriminator; the dual-headed discriminator includes a spatial domain discriminator and a wavelet domain discriminator, comprising: The generator is constructed using a U-Net structure, and a CBSAF channel-space dual collaborative enhancement module is introduced at the jump connection of the U-Net structure. Construct the spatial domain discriminator, which is used to discriminate the spatial structure and texture details of the received first input group and second input group; The wavelet domain discriminator is constructed, which is used to perform multi-scale frequency domain decomposition on the received first input group and second input group through DB4 discrete wavelet transform, extract frequency domain features, and then perform discrimination. The improved conditional generative adversarial network is constructed based on the generator, the spatial domain discriminator, and the wavelet domain discriminator.

[0054] Specifically, the generator adopts a U-Net structure, and a CBSAF channel-spatial dual collaborative enhancement module is added at the skip connections of the U-Net. This module adopts a dual-branch parallel processing architecture: the channel branch introduces the channel attention submodule of CBAM, which performs global average pooling and global max pooling on the input features respectively, and feeds the resulting feature vector into a shared multilayer perceptron for processing; the spatial branch adopts the spatial attention mechanism in CASAB, which simultaneously aggregates four statistics of the input feature map in the channel dimension: mean, maximum, minimum, and sum. This descriptor is fed into a two-layer convolutional encoder, where the first layer uses a 7×7 large-scale convolutional kernel, and the output of this convolution is passed through a Swish activation function. The second layer uses a 1×1 convolution as a projection layer to compress the features to a single channel, and finally generates a high-precision spatial weight heatmap through the Sigmoid function. This module uses a gated fusion strategy to achieve nonlinear adaptive fusion of the channel and spatial branches, and finally fuses the enhanced features with the original features through residual connections. The dual-head discriminator comprises a spatial domain discriminator and a wavelet domain discriminator: the spatial domain discriminator adopts the original Pix2Pix spatial domain discriminator structure to compare and distinguish the spatial structure and texture details of the generated sea ice image and the real sea ice image; the wavelet domain discriminator first performs multi-scale frequency domain decomposition on the generated sea ice image and the real sea ice image through DB4 discrete wavelet transform, extracts frequency domain features, and then performs discrimination to further enhance the ability to capture subtle texture features of sea ice. The generator, spatial domain discriminator, and wavelet domain discriminator together constitute the improved conditional generative adversarial network of this embodiment.

[0055] For example, when constructing the improved conditional generative adversarial network, all skip connections of the U-Net generator can be replaced with CBSAF modules. This module first performs global average pooling and max pooling on the input features through the channel attention branch, and then weights them through MLP. At the same time, it aggregates four statistical quantities—mean, maximum, minimum, and sum—through the spatial attention branch and generates spatial weights through 7×7 convolution + 1×1 convolution. The two are then gated and fused, and the residual connections are output. The spatial domain discriminator uses PatchGAN to perform local authenticity discrimination on the splicing result of the generated / real sea ice scene image and the labeled image. The wavelet domain discriminator first performs DB4 discrete wavelet transform on the same splicing result, stacks the decomposed low-frequency and high-frequency components, and then feeds them into the convolutional network for frequency domain authenticity discrimination. The three together constitute the network.

[0056] In this optional embodiment, by introducing a CBSAF module at the jump connection of the generator, the decoding stage can simultaneously preserve the spatial location of key channel responses and subtle textures, effectively overcoming the detail loss problem caused by the simple splicing of traditional jump connections. Simultaneously, the spatial domain discriminator and the wavelet domain discriminator exert adversarial constraints on the generated samples from two dimensions: spatial structure and high-frequency components in the frequency domain. This ensures the generator is subject to dual supervision of global realism and local texture consistency. The synergistic effect of both significantly improves the detail restoration capability and class discrimination of the augmented samples, avoiding the class confusion and sample distortion that easily occur with a single discriminator. The final generated sea ice sample possesses both macroscopic structural realism and local texture accuracy.

[0057] Optionally, the step of using the standardized sea ice map as input to the generator and outputting a generated sea ice scene map by the generator includes: The standardized sea ice map is input into the encoder of the generator, and multi-scale features are extracted through downsampling; At the U-Net skip connection of the generator, the multi-scale features are enhanced in both channel and spatial dimensions through the CBSAF channel-space dual collaborative enhancement module to obtain the enhanced multi-scale features; The enhanced multi-scale features are input into the decoder of the generator, and after upsampling and reconstruction, the generated sea ice scene map is output.

[0058] Specifically, a standardized sea ice annotation map (single-channel grayscale image) is fed into the encoder of the U-Net generator. The encoder extracts multi-scale features step by step through stacked convolutional layers and downsampling layers, such as convolutions with a stride of 2 or max pooling. Each downsampling operation halves the spatial size of the feature map and doubles the number of channels, resulting in several layers of feature maps at different scales. At each skip connection of the U-Net layer, the feature map output by the current layer of the encoder is concatenated with the upsampled feature map of the corresponding layer of the decoder along the channel dimension. The concatenated feature map is then input into the CBSAF channel-space dual collaborative enhancement module. As described in step S2, this module first performs global average pooling and max pooling on the concatenated features through the channel attention branch, followed by shared MLP weighting. Then, through the spatial attention branch, it simultaneously aggregates four statistical measures: mean, maximum, minimum, and sum, and generates a spatial weight heatmap through 7×7 and 1×1 convolutions. After gating and fusing the two branches, the enhanced multi-scale features are output through residual connections. The enhanced feature is then fed into the next layer of the decoder, which gradually restores the spatial resolution through upsampling layers (such as deconvolution or interpolated convolution) and fuses it with the enhanced feature from the encoder. Finally, it passes through the output layer, such as convolution plus Tanh activation, to generate a three-channel RGB sea ice scene image with the same size as the input annotation image.

[0059] In this optional embodiment, multi-scale features are extracted by encoder downsampling, enabling the network to simultaneously capture the macroscopic layout and local details of sea ice. The CBSAF module at the skip connection enhances the category-related response with channel attention and accurately locates the position of fine textures such as ice cracks and ice crevices with spatial attention. The residual connection after gating and fusion of the two significantly improves the feature representation of key areas while preserving the original context. The decoder reconstructs step by step based on the enhanced multi-scale features, so that the generated image recovers the high-frequency details that are easily lost in traditional U-Net under strict adherence to the annotation conditions. This effectively solves the contradiction in existing methods where the global structure is reasonable but the texture is blurry, or the texture is clear but the structure is distorted, and achieves simultaneous improvement of the macroscopic and microscopic quality of sea ice samples.

[0060] Optionally, the step of combining the generated sea ice scene image with the corresponding standardized sea ice annotation image into a first input group, combining the standardized real sea ice scene image with the corresponding standardized sea ice annotation image into a second input group, and inputting the first input group and the second input group into the spatial domain discriminator, and inputting the first input group and the second input group into the wavelet domain discriminator, respectively, to obtain spatial domain discrimination results and wavelet domain discrimination results, includes: The generated sea ice scene image and the corresponding standardized sea ice annotation image are stitched together along the channel dimension to obtain the first input group; the standardized real sea ice scene image and the corresponding standardized sea ice annotation image are stitched together along the channel dimension to obtain the second input group. The first input group and the second input group are respectively input into the spatial domain discriminator, and the spatial domain discriminator outputs the spatial domain discrimination result indicating the authenticity of the input; The first input group and the second input group are respectively input into the wavelet domain discriminator, and the wavelet domain discriminator outputs the wavelet domain discrimination result indicating the authenticity of the input.

[0061] Specifically, in each training iteration, the RGB sea ice scene image output by the generator is directly concatenated with the corresponding standardized sea ice annotation image along the channel dimension, that is, the two are merged into a multi-channel tensor as independent channel layers, forming the first input group. Similarly, the standardized real sea ice scene image and its corresponding standardized sea ice annotation image are also concatenated along the channel dimension to form the second input group. Subsequently, the first and second input groups are fed into the spatial domain discriminator, which extracts features from the spatial structure of the input groups layer by layer through convolutional layers, and finally outputs a scalar or feature matrix, representing the probability that the current input group is a real sample, i.e., the spatial domain discrimination result. Simultaneously, the first and second input groups are fed into a wavelet domain discriminator. This discriminator first performs a DB4 discrete wavelet transform on each input group, decomposing the input into low-frequency approximate components and high-frequency detail components in three directions (horizontal, vertical, and diagonal). Each component reflects the frequency domain features of the image at different scales. Then, these components are stacked along the channel dimension to form a multi-channel frequency domain feature map. Finally, a convolutional network is used to extract and classify features from this frequency domain feature map, outputting the wavelet domain discrimination result, representing the degree of realism of the input group in the frequency domain. The outputs of the two discriminators are used together to construct the subsequent loss function.

[0062] In this optional embodiment, by stitching the generated sea ice scene image and the labeled image into a first input group, and stitching the real scene image and the labeled image into a second input group, and inputting these into a spatial domain discriminator and a wavelet domain discriminator respectively, this invention constructs a dual adversarial discrimination mechanism: the spatial domain discriminator judges the spatial structure and texture authenticity of the stitched image, forcing the generator to approximate the real distribution in terms of macroscopic layout and local texture; the wavelet domain discriminator performs DB4 discrete wavelet transform on the same stitched result, capturing the differences in high-frequency details between the generated sample and the real sample from the frequency domain perspective, forcing the generator to be highly consistent with the real distribution in subtle features such as ice cracks and ice crevices in order to pass the frequency domain discrimination. The two output independent discrimination results, providing the generator with complementary gradient feedback in both the spatial and frequency domains, so that the generator is simultaneously constrained by both global structure and local high-frequency features during training, significantly improving the fidelity and detail restoration ability of the generated samples, and effectively avoiding the problem that a single discriminator may easily ignore high-frequency information or be insensitive to category features.

[0063] Optionally, the step of generating a composite loss function based on the spatial domain discrimination result and the wavelet domain discrimination result, and updating the network parameters of the generator and the dual-head discriminator through backpropagation, includes: The adversarial loss is determined based on the spatial domain discrimination result and the wavelet domain discrimination result; L1 loss and perception loss are determined based on the generated sea ice scene map and the corresponding standardized real sea ice scene map. The adversarial loss, the L1 loss, and the perceptual loss are weighted and combined to obtain the composite loss function; Based on the composite loss function, the network parameters of the generator and the network parameters of the dual-head discriminator are updated respectively through the backpropagation algorithm.

[0064] Specifically, based on the spatial domain discrimination results output by the spatial domain discriminator and the wavelet domain discrimination results output by the wavelet domain discriminator, the adversarial loss is determined according to the standard adversarial loss calculation method for conditional generative adversarial networks, such as least squares loss or cross-entropy loss. This loss is used to constrain the adversarial learning between the generator and the dual-head discriminator, forcing the generator to generate sea ice images that are closer to the real distribution. Simultaneously, the generated sea ice scene image is compared pixel-by-pixel with the corresponding standardized real sea ice scene image, and the L1 loss, i.e., the sum of absolute errors, is calculated to reduce the difference between the two at the pixel level. Then, a pre-trained feature extraction network, such as the VGG network (Visual Geometry Group), is used to extract multi-layer feature maps of the generated image and the real image respectively. The Euclidean distance between each layer of feature maps is calculated and summed to obtain the perceptual loss, which is used to measure the similarity between the two at the high-level semantic feature level, optimizing the structural rationality and visual quality of the generated image. The adversarial loss, L1 loss, and perceptual loss are weighted and summed according to preset weight coefficients to obtain the composite loss function. Finally, based on the value of the composite loss function, the gradients of each parameter in the generator, spatial domain discriminator, and wavelet domain discriminator are calculated using the backpropagation algorithm, and the parameters of each network are updated along the gradient descent direction using an optimizer such as Adam to complete one iteration of parameter optimization.

[0065] In this optional embodiment, a weighted combination of adversarial loss, L1 loss, and perceptual loss achieves multi-dimensional quality constraints on the generated samples: adversarial loss ensures that the samples are indistinguishable from real images in the global distribution, L1 loss suppresses pixel-level biases and blurring, and perceptual loss ensures the rationality of the high-level semantic structure. The synergistic effect of these three factors makes the training process more stable and converges faster. The resulting sea ice samples significantly outperform single-loss-driven methods in terms of pixel accuracy, structural integrity, and visual realism, effectively avoiding problems such as pattern collapse and edge distortion.

[0066] Optionally, repeating steps S3 to S5 until the composite loss function converges to obtain the trained sea ice sample augmentation model includes: In each iteration, S3 to S5 are executed sequentially, and after each iteration, it is determined whether the composite loss function has converged. When the value of the composite loss function is less than or equal to a preset threshold, the iteration stops, and the generator and the dual-head discriminator under the current network parameters are used as the trained sea ice sample augmentation model.

[0067] Specifically, in each training iteration, steps S3 (generating sea ice scene images), S4 (combining input groups and obtaining the discriminant results from the dual-head discriminator), and S5 (calculating the composite loss function and updating network parameters) are executed sequentially. After each iteration, the value of the current composite loss function is recorded. The model is continuously monitored for multiple iteration cycles, such as the change in loss value every 1000 iterations. When the value of the composite loss function continuously decreases and eventually stabilizes within a small range (e.g., the fluctuation range of the loss value is less than a preset tolerance threshold within 10 consecutive evaluation cycles), or the loss value no longer decreases significantly, the model is considered to have converged. At this point, iterative training is stopped, and the generator and dual-head discriminator under the current network parameters are saved together as the trained sea ice sample augmentation model. This model can then be used for subsequent sea ice sample augmentation tasks.

[0068] For example, the preset threshold can be set according to the range of loss values ​​at the beginning of training. For instance, the composite loss function value is approximately 5.0 at the start of training, and gradually decreases as iterations proceed. When the loss value is below 0.05 for 10 consecutive evaluation periods, for example, every 1000 iterations constitutes one evaluation period, it is considered convergent; or when the change in loss value between two adjacent evaluation periods is less than 0.5%, the loss is considered to have stabilized, i.e., the convergence condition is met. This threshold can be adjusted according to the specific dataset and network size, and this embodiment does not impose specific limitations.

[0069] In this optional embodiment, by continuously monitoring the convergence of the composite loss function and determining training termination based on a preset threshold or stability condition, this embodiment ensures that the model stops iterating when it reaches its optimal performance state, avoiding insufficient generation quality due to underfitting or decreased generalization ability due to overfitting. Simultaneously, the converged generator and dual-head discriminator are saved together as the final model, allowing the generator to be called independently for rapid inference in subsequent sample augmentation stages, achieving separation of training and inference responsibilities. This convergence determination mechanism, combining adversarial loss, L1 loss, and perceptual loss constraints, enables the training process to converge smoothly under dual supervision in the spatial and frequency domains, effectively shortening ineffective iteration time, improving the automation level of model training and the reproducibility of results, and ultimately obtaining an augmented model that can stably output high-quality sea ice samples with both global structure and detailed texture.

[0070] Optionally, the step of inputting the sea ice annotation map to be augmented into the sea ice sample augmentation model, and having the generator output the corresponding augmented sea ice samples, includes: The sea ice annotation map to be augmented is input into the trained sea ice sample augmentation model, and the generator in the sea ice sample augmentation model is called. The generator outputs the corresponding sea ice augmentation sample based on the sea ice annotation map to be augmented.

[0071] Specifically, one or more un-trained sea ice annotation images are obtained. These un-trained annotation images have the same size as those standardized during preprocessing, such as 256×256 pixels, and have been normalized. They are directly fed into the U-Net generator in the trained sea ice sample augmentation model. The generator performs one forward propagation, without passing through a dual-head discriminator, and outputs the corresponding RGB sea ice scene image, which is the generated augmented sample. Subsequently, the quality of the generated sea ice images is validated using pre-defined validation and test sets. For example, the FID (Fréchet Inception Distance) score or subjective evaluation metric is calculated between the generated image and the real images in the validation set. Augmented samples that meet quality standards are selected, such as those with an FID below a certain threshold or those that pass manual review. Finally, the selected augmented samples are added to the original sea ice sample dataset to expand the data size and alleviate the problems of sample scarcity and class imbalance.

[0072] In this optional embodiment, by separating the training and inference phases, only a single forward propagation is needed from the generator in the trained model during sample augmentation, eliminating the need for a dual-head discriminator. This significantly reduces the computational overhead and storage requirements of the inference phase, achieving efficient end-to-end mapping from the annotated image to the sea ice augmented sample. This design allows users to quickly generate corresponding realistic sea ice scene images in batches simply by providing any annotated sea ice image, overcoming the limitations of high cost and long cycle of actual data collection. Simultaneously, the generated augmented samples can be directly added to the original dataset, effectively alleviating the problems of sea ice sample scarcity and class imbalance, providing larger-scale and more diverse training data support for subsequent deep learning tasks such as sea ice identification and thickness inversion.

[0073] Combination Figure 2 As shown, this embodiment of the invention also provides a sea ice sample augmentation system, comprising: The data processing unit is used to classify and preprocess the collected ship-based real sea ice scene images to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images. A network construction unit is used to construct an improved conditional generative adversarial network, wherein the improved conditional generative adversarial network includes a generator and a dual-head discriminator; the dual-head discriminator includes a spatial domain discriminator and a wavelet domain discriminator. An adversarial training unit is configured to use the standardized sea ice map as input to the generator, and output a generated sea ice scene map from the generator; to combine the generated sea ice scene map with the corresponding standardized sea ice map as a first input group, and to combine the standardized real sea ice scene map with the corresponding standardized sea ice map as a second input group; to input the first input group and the second input group respectively into the spatial domain discriminator, and to input the first input group and the second input group respectively into the wavelet domain discriminator, to obtain spatial domain discrimination results and wavelet domain discrimination results; and to generate a composite loss function based on the spatial domain discrimination results and the wavelet domain discrimination results, and update the network parameters of the generator and the dual-head discriminator through backpropagation. An iterative training unit is used to repeatedly execute the processing of the adversarial training unit until the composite loss function converges, thereby obtaining a trained sea ice sample augmentation model. The sample generation unit is used to input the sea ice annotation map to be augmented into the sea ice sample augmentation model, and the generator outputs the corresponding sea ice augmented sample.

[0074] The sea ice sample augmentation system of the present invention has the same advantages over the prior art as the sea ice sample augmentation method described above, and will not be repeated here.

[0075] Combination Figure 3 As shown, embodiments of the present invention also provide an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the sea ice sample augmentation method as described above when executing the computer program.

[0076] The electronic device of the present invention has the same advantages over the prior art as the aforementioned sea ice sample augmentation method, and will not be repeated here.

[0077] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described sea ice sample augmentation method.

[0078] The computer-readable storage medium of the present invention has the same advantages over the prior art as the aforementioned sea ice sample augmentation method over the prior art, and will not be repeated here.

[0079] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for augmenting sea ice samples, characterized in that, include: S1: The collected images of real sea ice scenes on the ship base are labeled and preprocessed to obtain standardized real sea ice scene images and corresponding standardized sea ice labeled images; S2: Construct an improved conditional generative adversarial network, which includes a generator and a dual-headed discriminator; the dual-headed discriminator includes a spatial domain discriminator and a wavelet domain discriminator; S3: Using the standardized sea ice map as input to the generator, the generator employs a U-Net structure and introduces a CBSAF channel-spatial dual collaborative enhancement module at the jump connections of the U-Net structure. Under the collaborative enhancement of the CBSAF channel-spatial dual collaborative enhancement module, the generator outputs a generated sea ice scene map; including: S31: Input the standardized sea ice annotation map into the encoder of the generator, and extract multi-scale features through downsampling; S32: At the U-Net skip connection of the generator, the multi-scale features are enhanced in both channel and spatial dimensions through the CBSAF channel-space dual collaborative enhancement module to obtain the enhanced multi-scale features; S33: Input the enhanced multi-scale features into the decoder of the generator, reconstruct them by upsampling, and output the generated sea ice scene map; S4: Combine the generated sea ice scene image with the corresponding standardized sea ice annotation image to form a first input group, combine the standardized real sea ice scene image with the corresponding standardized sea ice annotation image to form a second input group, and input the first input group and the second input group into the spatial domain discriminator and the wavelet domain discriminator respectively. The spatial domain discriminator is used to discriminate the spatial structure and texture details of the received first input group and the second input group; the wavelet domain discriminator is used to perform multi-scale frequency domain decomposition on the received first input group and the second input group through DB4 discrete wavelet transform, extract frequency domain features, and then discriminate to obtain spatial domain discrimination results and wavelet domain discrimination results; including: S41: Concatenate the generated sea ice scene image and the corresponding standardized sea ice annotation image along the channel dimension to obtain the first input group; Concatenate the standardized real sea ice scene image and the corresponding standardized sea ice annotation image along the channel dimension to obtain the second input group; S42: Input the first input group and the second input group into the spatial domain discriminator respectively, and the spatial domain discriminator outputs the spatial domain discrimination result indicating the authenticity of the input; S43: Input the first input group and the second input group into the wavelet domain discriminator respectively, and the wavelet domain discriminator outputs the wavelet domain discrimination result indicating the authenticity of the input; S5: Generate a composite loss function based on the spatial domain discrimination result and the wavelet domain discrimination result, and update the network parameters of the generator and the dual-head discriminator through backpropagation; S6: Repeat S3 to S5 until the composite loss function converges to obtain the trained sea ice sample augmentation model; S7: Input the sea ice annotation map to be augmented into the sea ice sample augmentation model, and the generator outputs the corresponding sea ice augmented sample.

2. The sea ice sample augmentation method according to claim 1, characterized in that, The process of classifying and preprocessing the collected ship-based sea ice scene images to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images includes: The sea ice categories in the collected real-world sea ice scene images of the ship base are labeled to obtain a sea ice label map; The real sea ice scene image and the sea ice annotation image are sequentially subjected to denoising, size unification and normalization processes to obtain the standardized real sea ice scene image and the corresponding standardized sea ice annotation image; and the standardized real sea ice scene image and the corresponding standardized sea ice annotation image are divided into training set, validation set and test set according to a preset ratio.

3. The sea ice sample augmentation method according to claim 1, characterized in that, The step of generating a composite loss function based on the spatial domain discrimination result and the wavelet domain discrimination result, and updating the network parameters of the generator and the dual-head discriminator through backpropagation, includes: The adversarial loss is determined based on the spatial domain discrimination result and the wavelet domain discrimination result; L1 loss and perception loss are determined based on the generated sea ice scene map and the corresponding standardized real sea ice scene map. The adversarial loss, the L1 loss, and the perceptual loss are weighted and combined to obtain the composite loss function; Based on the composite loss function, the network parameters of the generator and the network parameters of the dual-head discriminator are updated respectively through the backpropagation algorithm.

4. The sea ice sample augmentation method according to claim 3, characterized in that, The process of repeating steps S3 to S5 until the composite loss function converges yields the trained sea ice sample augmentation model, including: In each iteration, S3 to S5 are executed sequentially, and after each iteration, it is determined whether the composite loss function has converged. When the value of the composite loss function is less than or equal to a preset threshold, the iteration stops, and the generator and the dual-head discriminator under the current network parameters are used as the trained sea ice sample augmentation model.

5. The sea ice sample augmentation method according to claim 4, characterized in that, The step of inputting the sea ice annotation map to be augmented into the sea ice sample augmentation model, and having the generator output the corresponding sea ice augmented samples, includes: The sea ice annotation map to be augmented is input into the trained sea ice sample augmentation model, and the generator in the sea ice sample augmentation model is called. The generator outputs the corresponding sea ice augmentation sample based on the sea ice annotation map to be augmented.

6. A sea ice sample augmentation system, characterized in that, include: The data processing unit is used to classify and preprocess the collected ship-based real sea ice scene images to obtain standardized real sea ice scene images and corresponding standardized sea ice annotation images. A network construction unit is used to construct an improved conditional generative adversarial network (GAN). The improved GAN includes a generator and a dual-head discriminator. The dual-head discriminator includes a spatial domain discriminator and a wavelet domain discriminator. The process includes: inputting the standardized sea ice annotation map into the encoder of the generator, and extracting multi-scale features through downsampling; at the U-Net skip connections of the generator, performing dual enhancement of the multi-scale features in both channel and spatial dimensions using a CBSAF channel-space dual collaborative enhancement module to obtain enhanced multi-scale features; inputting the enhanced multi-scale features into the decoder of the generator, and reconstructing them through upsampling to output the generated sea ice scene map. An adversarial training unit is used to take the standardized sea ice map as input to the generator. The generator adopts a U-Net structure and introduces a CBSAF channel-spatial dual collaborative enhancement module at the skip connections of the U-Net structure. Under the collaborative enhancement of the CBSAF channel-spatial dual collaborative enhancement module, the generator outputs a generated sea ice scene map. It is also used to combine the generated sea ice scene map with the corresponding standardized sea ice map as a first input group, and to combine the standardized real sea ice scene map with the corresponding standardized sea ice map as a second input group. The first input group and the second input group are each input to the spatial domain discriminator, and the first input group and the second input group are each input to the wavelet domain discriminator. The spatial domain discriminator is used to discriminate the spatial structure and texture details of the received first input group and second input group. The wavelet domain discriminator is used to discriminate the spatial structure and texture details of the received first input group and second input group. The system comprises two input groups: a first input group and a second input group. The first input group undergoes multi-scale frequency domain decomposition using DB4 discrete wavelet transform. Frequency domain features are extracted and then discriminated to obtain spatial domain and wavelet domain discrimination results. The system also generates a composite loss function based on the spatial domain and wavelet domain discrimination results and updates the network parameters of the generator and the dual-head discriminator through backpropagation. The process includes: concatenating the generated sea ice scene image with the corresponding standardized sea ice annotation image along the channel dimension to obtain a first input group; concatenating the standardized real sea ice scene image with the corresponding standardized sea ice annotation image along the channel dimension to obtain a second input group; inputting the first and second input groups into the spatial domain discriminator, which outputs a spatial domain discrimination result indicating the authenticity of the input; and inputting the first and second input groups into the wavelet domain discriminator, which outputs a wavelet domain discrimination result indicating the authenticity of the input. An iterative training unit is used to repeatedly execute the processing of the adversarial training unit until the composite loss function converges, thereby obtaining a trained sea ice sample augmentation model. The sample generation unit is used to input the sea ice annotation map to be augmented into the sea ice sample augmentation model, and the generator outputs the corresponding sea ice augmented sample.

7. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the sea ice sample augmentation method as described in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the sea ice sample augmentation method according to any one of claims 1 to 5.

Images