Adaptive fine-grained ocean ship rescue positioning method based on severe weather

Abstract

The invention discloses an adaptive fine-grained ocean ship rescue positioning method based on severe weather, and belongs to the technical field of ocean ship rescue positioning, and the method comprises the steps: collecting photo data, carrying out the feature extraction of pollution elements, and carrying out the fitting with the photo data to form a data set; based on a multi-scale linear attention decoupling head, the fine-grained classification features and the fine-grained regression features are decoupled; using wavelet transform features to capture bottom texture information, and aligning the bottom texture information with the features of the decoupling head; a loss function and a training model are provided, a total loss function value is calculated, back propagation is carried out, optimization is carried out through a selected optimizer and corresponding parameters, and a final ship positioning model is obtained. According to the method, a multi-scale linear attention decoupling head, a wavelet transform feature decomposition space semantic alignment module and a progressive reinforcement learning loss function fine-grained module are introduced, the fine-grained learning ability of the method is improved, the detection ability and robustness of the model are improved, and finally the ship positioning accuracy and efficiency in severe weather are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of marine ship rescue positioning, and in particular relates to a method for marine ship rescue positioning based on adaptive fine-grained severe weather. Background Art

[0002] As a vast body of water on the surface of the earth, the ocean has a huge impact on human activities. With the acceleration of the process of world economic integration and information globalization, the global maritime shipping industry has developed rapidly. At the same time, the probability of maritime accidents is also increasing, posing a huge threat to people's life safety and property. In order to reduce casualties and property losses after maritime accidents, the maritime rescue system came into being.

[0003] However, faced with complex ocean backgrounds, tiny targets and crowded scenes, maritime rescue has always been plagued by problems such as long distances, long times and difficult search and rescue. In particular, the search and rescue process will be affected by extreme conditions, resulting in data distribution shifts in the detected remote sensing images. Therefore, traditional ship positioning algorithms are easily affected by factors such as heavy rain, sea fog, high exposure and cloud occlusion, resulting in missed targets, false detections or a significant decrease in classification accuracy, making it difficult to accurately locate ocean ships, prolonging the search and rescue time, and preventing ships in distress from being rescued in a timely manner. Summary of the invention

[0004] In view of the shortcomings existing in the background technology, the purpose of the present invention is to propose a method for adaptive fine-grained marine ship rescue positioning based on severe weather, and adopt a fine-grained method to redesign the ship positioning model so that it has fine-grained feature level processing capabilities, thereby improving the detection capability and robustness of the model in severe weather, separating the classification features from the boundary regression features, realizing targeted fine-grained feature learning, enhancing the model's ability to capture the region of interest (ROI), and improving the matching of positive and negative samples, and adopting a progressive reinforcement learning loss function. Through iterative reinforcement learning, the detection capability of difficult-to-detect categories is enhanced, and the ship target can be positioned more accurately, solving the current problem of difficulty in positioning rescued objects at sea.

[0005] The technical solution adopted by the present invention is:

[0006] A method for adaptive fine-grained ocean ship rescue positioning based on severe weather comprises the following steps: S1) Collect ocean remote sensing ship positioning and ocean severe weather remote sensing satellite photo data to build an ocean ship positioning dataset under severe weather conditions; S2) With the help of the pollution element generation module, feature extraction is performed on the pollution elements, and they are fitted with the marine remote sensing ship positioning photos to generate pollution image data sets with different data distributions; S3) A multi-scale linear attention decoupling head based on a region proposal network takes features related to classification features and boundary regression features as input features, and decouples the output into fine-grained classification features Xcls and fine-grained regression features Xreg; S4) Based on the wavelet transform feature decomposition spatial semantic alignment module, the wavelet transform feature is used to capture the underlying texture information, and the feature is aligned with the decoupling head. The N-th layer feature of the feature X that has been spatially aligned by FPN is aligned with the N-1-th layer feature of Xcls and Xreg for spatial semantic information. S5) Based on the progressive reinforcement loss function, the localization ability of difficult-to-localize categories is improved through iterative reinforcement learning; S6) The polluted image dataset is trained and input into step S3) and step S4), and then the total loss function value is calculated and back propagated. The connection weights are optimized through the selected optimizer and corresponding parameters to obtain a fine-grained marine ship rescue positioning model based on severe weather adaptation.

[0007] Preferably, in step S2), the specific process of generating the contaminated image data set is as follows: Sa1) Embedding and initialization of pollution conditions: Based on the CLIP text encoder, the pollution elements are converted into conditional embedding vectors E, and the remote sensing ship positioning photos are encoded by the VAE of the Stable Diffusion framework to obtain the initial potential expression , and then inject Gaussian noise to generate the noise potential expression , concatenate the noise potential expressions to generate multi-pollution joint features PF; Sa2) Multi-scale pollution feature extraction: Input multi-pollution joint features, build pollution feature pyramids at three scales of 16×16, 32×32 and 64×64 based on U-Net skip connection, align the conditional embedding vector E with the multi-pollution joint features through hierarchical cross attention, and use the feature pyramid network to fuse multi-scale features to simultaneously retain the local texture of the rainstorm and the global morphology of the cloud layer; Sa3) Dynamic pollution intensity control: Based on the dual-path discrimination module, the spatial attention path is used to pass the relative position matrix Control the intensity attenuation of pollution elements above the hull and learn RGB channel weights using channel attention path , and through the learnable parameters through Function fusion to generate dynamic masks ; Sa4) Reverse diffusion and data fusion: Based on the dynamic mask, the noise latent vector is iteratively updated in the latent space through the conditional reverse diffusion process, combining the conditional embedding vector E and the noise prediction network , using VAE decoding to generate polluted images , build a pollution image dataset.

[0008] Preferably, in step S3), the input features are decoupled in the multi-scale linear attention head The specific steps of the processing are as follows: Sb1) Multi-scale feature extraction and fusion: Input features Perform average pooling along the width and height dimensions respectively to obtain the feature after average pooling along the height. And the average pooling feature along the width ;Will and After transposition, concatenate along the channel dimension to obtain the concatenated features ;right The application kernel size is The depth convolution obtains multi-scale features in the height direction and multi-scale features in width direction ,in The values ​​are 1, 3, and 5. Determine the receptive field, gradually add the features of the large receptive field to the features of the small receptive field, and the assignment operation formula is as follows:

[0009] ,

[0010] in, is a height feature with a receptive field of 1×1, is a height feature with a receptive field of 3×3, is a height feature with a receptive field of 5×5, is the width feature with a receptive field of 1×1, is a width feature with a receptive field of 3×3, is a width feature with a receptive field of 5×5; Sb2) Multi-scale attention mechanism calculation: by extending , , and Features are multiplied and then passed through the activation function Calculate the attention weight and assign the following formula:

[0011]

[0012] in, is an extension function, is the activation function, is the feature after high directional attention weight, It is the feature after attention weight in width direction; Will and It is divided into G groups in the channel dimension, and the total number of channels is , so that the number of channels in each group is , average pooling is performed on each group of channels, and the assignment operation formula is as follows:

[0013] in, is the average pooling function, After grouping The pooling features, After grouping Pooling features; Sb3) Cross-matrix similarity and feature update: Calculation and , and The cross matrix similarity is calculated and the result is obtained through softmax. The assignment operation formula is as follows:

[0014]

[0015] in, is the matrix multiplication function, It is a normalized exponential function that converts the input into a probability distribution; After normalization and and application After the function, with the input features Multiply and update the final result to obtain fine-grained classification features X cls and fine-grained regression features X reg , the formula is as follows:

[0016]

[0017] Preferably, in step S4), the specific steps of the wavelet transform feature decomposition space semantic alignment module are as follows: Sc1) Deep convolution feature enhancement: input features Perform deep convolution operations to enhance the overall correlation and nonlinearity of features while keeping the input dimension unchanged. The assignment operation formula is as follows:

[0018] In the formula is the depth convolution operation function; is the output feature after deep convolution feature enhancement; Sc2) Haar wavelet transform feature decomposition: output features for each channel Perform a first-order Haar wavelet transform and decompose it into a low-frequency component Ac and three high-frequency components. The assignment operation formula is as follows:

[0019]

[0020]

[0021]

[0022] In the formula and is the position index of the feature matrix, It is a low-frequency component that captures the overall trend of the feature; It is the horizontal high-frequency component, capturing the high-frequency detail information of the feature in the horizontal direction; It is the vertical high-frequency component, capturing the high-frequency detail information of the feature in the vertical direction; It is the diagonal high-frequency component, which captures the high-frequency detail information of the feature in the diagonal direction; Sc3) High- and low-frequency feature fusion: The low-frequency component and high-frequency component are processed separately through point convolution, and convolution processing and feature fusion are performed on them, and the current layer output is output through WTFD With the features of the previous layer Align and accumulate to obtain the final fused low-frequency features and fused high-frequency features. The formula is as follows:

[0023]

[0024]

[0025] in, is the low-frequency feature after fusion, is the high-frequency feature after fusion, To fuse high and low frequency feature operations; WTFD repeats the above steps, that is, each output is spatially semantically aligned with the N-1 layer.

[0026] Preferably, in step S5), the specific steps are as follows: Progressive reinforcement loss function, including reinforcement numerator factor function Fl and iteration factor , the formula of the total loss function is:

[0027] in, is the loss function, is the model’s predicted probability for the target category, is the balance factor, is the regulating factor; According to the length of the data set and the set number of reinforcement learning rounds, refine is iterated to obtain the iteration factor. The assignment operation formula is as follows:

[0028] in, For each classification Ap value, it is a hyperparameter. is the number of samples in the dataset, is the set reinforcement learning round, is the current iteration number.

[0029] Preferably, the present invention further proposes a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the method for adaptive fine-grained ocean ship rescue positioning based on severe weather is implemented.

[0030] Compared with the prior art, the present invention proposes a method for adaptive fine-grained marine ship rescue positioning based on severe weather, and the advantages of the method are: (1) The present invention adopts a fine-grained method to redesign the ship positioning model so that it has the ability to process fine-grained feature levels, thereby improving the detection ability and robustness of the model in severe weather. At the same time, the features of heavy rain, sea fog, high exposure and cloud occlusion pollution elements are extracted and then fused with remote sensing ship images. By taking advantage of the relatively stable characteristics of the ocean background and data distribution, the photos after the data distribution is offset can be made more natural. (2) The present invention introduces a multi-scale linear attention decoupling head of the region proposal network. Through the multi-scale linear attention mechanism, the input feature X containing the separation of classification features and boundary regression features is analyzed, and the decoupling output is Xcls fine-grained classification features and Xreg fine-grained regression features, thereby realizing targeted fine-grained feature learning. This improvement enhances the model's ability to capture the region of interest (ROI) and improves the matching of positive and negative samples. Fine-grained attention learning replaces ordinary convolution learning, which can more accurately locate ship targets. (3) The present invention introduces a wavelet transform feature decomposition spatial semantic alignment module, uses wavelet transform features to capture underlying texture information, and performs feature alignment with the previous stage decoupling head to solve the problem of insufficient semantic information in the feature layer extracted by the feature pyramid network (FPN). In the specific operation, the Nth layer feature of the feature X that has been spatially aligned by FPN is aligned with the N-1th layer feature of Xcls and Xreg for spatial semantic information. (4) The present invention introduces an iterative reinforcement learning training process for difficult-to-detect classification and proposes a new loss function to improve the problems of uneven sample proportion distribution and insufficient texture information features of difficult-to-detect classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of the method for adaptive fine-grained ocean ship rescue positioning based on severe weather of the present invention; Figure 2 This is a framework diagram of the severe weather adaptive fine-grained ocean remote sensing ship rescue positioning model proposed by the present invention; Figure 3 This is a block diagram of the multi-scale linear attention decoupling head structure; Figure 4 Processing flow chart for wavelet transform feature decomposer and spatial semantic alignment module; Figure 5 Comparison chart of enhanced rainstorm, sea fog, high exposure and cloud occlusion effects based on ocean background; Figure 6 This is an architecture diagram of the present invention's method for adaptive fine-grained marine ship rescue positioning based on severe weather. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present application to further clearly and completely describe the technical solutions in the embodiments of the present application. It should be noted that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0033] In order to make the invention objectives, technical solutions and advantages of the present application clearer, the embodiments of the present application are further described in detail in conjunction with the drawings in the specification: In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the advantages of the present invention will be further illustrated by comparing the embodiments in conjunction with the drawings and specific implementation methods.

[0034] The present invention proposes a method for adaptive fine-grained marine ship rescue positioning based on severe weather, and the steps of the method are described in detail: S1) Collect ocean remote sensing ship positioning and ocean severe weather remote sensing satellite photo data to build an ocean ship positioning dataset under severe weather conditions; S2) With the help of the pollution element generation module, the features of pollution elements such as heavy rain, sea fog, high exposure and cloud cover are extracted, and fitted with the marine remote sensing ship positioning photos to generate pollution image datasets with different data distributions; Specifically, in step S2), the pollution element generation module includes the following: The pollution element generation module includes three parts: a multi-scale pollution feature generation module, a pollution intensity discrimination module and an element fusion layer, including the following: The multi-scale pollution feature generation module extracts multi-scale pollution features of rainstorm, sea fog, high exposure and cloud occlusion pollution elements based on the Stable Diffusion framework, generates pollution feature maps with different intensity distributions through a cascaded noise prediction network, and uses a hierarchical attention mechanism to concatenate the potential representations of each pollution element to form a multi-pollution joint feature PF. The pollution intensity discrimination module inputs the multi-pollution joint feature PF into the intensity adjustment unit composed of the cross attention mechanism, and performs two-stage dynamic weight allocation: the first stage calculates the occlusion relationship matrix Q=[q1, q2, q3, q4] between the pollution element and the ship target (q1, q2, q3, q4 corresponds to the pollution elements of heavy rain, sea fog, high exposure and cloud occlusion), and the second stage screens the effective pollution area through differentiable threshold gating to generate a dynamic pollution mask that conforms to the laws of physics; The element fusion layer uses adaptive hybrid rendering technology to fuse the pollution feature map with the ocean remote sensing ship positioning photos. Specifically, through the reverse diffusion process of the Stable Diffusion framework, conditional noise perturbation is applied to the remote sensing ship positioning photos in the latent space, and finally a pollution image dataset with a natural transition effect is output. The bounding box annotation information of the ship target is retained during the rendering process to ensure the feasibility of the post-pollution detection task. More specifically, the specific process of the multi-scale pollution feature generation module is as follows: The specific process of the multi-scale pollution feature generation module is as follows: 1) Use the text encoder of Stable Diffusion to generate the conditional embedding vector of the pollution element; 2) Perform multi-scale feature extraction in the jump connection layer of U-Net to obtain pollution pattern maps with different resolutions from 16×16 to 256×256, namely pollution feature maps; 3) Fuse multi-scale pollution features through the feature pyramid network to form a multi-pollution joint feature PF, which retains the coordinated expression of the local texture of the rainstorm and the global morphology of the cloud layer.

[0035] More specifically, the specific process of the pollution intensity determination module is as follows: The pollution intensity discrimination module further adopts a dual-path discrimination module, in which the spatial attention path calculates the relative position weight of the polluted area and the ship target, and the channel attention path evaluates the intensity ratio of different pollution types; through a recursive weight update algorithm, the intensity of rainstorm pollution is the largest in the area above the hull, while the sea fog pollution presents a global uniform distribution characteristic; Specifically, in step S2), the specific process of generating the contaminated image data set is as follows: The pollution element generation module includes three parts: a multi-scale pollution feature generation module, a pollution intensity discrimination module and an element fusion layer. The multi-scale pollution feature generation module extracts multi-pollution features of rainstorm, sea fog, high exposure and cloud occlusion pollution elements to form a multi-pollution joint feature PF. The pollution intensity discrimination module adopts a dual-path discrimination module to optimize and generate a dynamic pollution mask. Finally, after reverse diffusion and data fusion in the element fusion layer, a pollution image data set with a natural transition effect is finally output; Sa1) Embedding and initialization of pollution conditions: Based on the CLIP text encoder, the four types of pollution elements, namely heavy rain, sea fog, high exposure and cloud occlusion, are converted into conditional embedding vectors E, and the remote sensing ship positioning photos are encoded by the VAE of the Stable Diffusion framework to obtain the initial potential expression , and then inject Gaussian noise to generate the noise potential expression , concatenate the noise potential expressions to generate multi-pollution joint features PF; Sa2) Multi-scale pollution feature extraction: Input multi-pollution joint features, build pollution feature pyramids at three scales of 16×16, 32×32 and 64×64 based on U-Net skip connection, align the conditional embedding vector E with the multi-pollution joint features through hierarchical cross attention, and use the feature pyramid network to fuse multi-scale features to simultaneously retain the local texture of the rainstorm and the global morphology of the cloud layer; Sa3) Dynamic pollution intensity control: Based on the dual-path discrimination module, the spatial attention path is used to pass the relative position matrix Control the intensity attenuation of pollution elements above the hull and learn RGB channel weights using channel attention path , and through the learnable parameters through Function fusion to generate dynamic masks ; Sa4) Reverse diffusion and data fusion: Based on the dynamic mask, the noise latent vector is iteratively updated in the latent space through the conditional reverse diffusion process, combining the conditional embedding vector E and the noise prediction network , using VAE decoding to generate polluted images , build a pollution image dataset.

[0036] S3) A multi-scale linear attention decoupling head based on a region proposal network, which takes features related to classification features and boundary regression features (such as coordinates, width and height) as input features , the decoupled output is X cls Fine-grained classification features and X reg Fine-grained regression features enable targeted fine-grained feature learning; improve and enhance the model's ability to capture regions of interest (ROIs) and improve the matching of positive and negative samples; Specifically, in step S3), the process of processing the input features in the multi-scale linear attention decoupling head is as follows: Sb1) Multi-scale feature extraction and fusion: Input features Average pooling is performed along the width and height dimensions respectively to extract features from different spatial dimensions and obtain the features after average pooling along the height. And the average pooling feature along the width ;in and Represents a function that performs average pooling operations along the height and width dimensions; and After transposition, concatenate along the channel dimension to obtain the concatenated features ,in is the concatenation function, which integrates the feature information of different dimensions in the channel dimension; The application kernel size is The depth of the convolution is obtained at different scales (by Determines the receptive field) and ,in The possible values ​​are 1, 3, and 5. The features of the large receptive field are gradually added to the features of the small receptive field to achieve the fusion of multi-scale features. The formula is as follows:

[0037] ,

[0038] in, is a height feature with a receptive field of 1×1, is a height feature with a receptive field of 3×3, is a height feature with a receptive field of 5×5, is the width feature with a receptive field of 1×1, is a width feature with a receptive field of 3×3, is a width feature with a receptive field of 5×5; Sb2) Multi-scale attention mechanism calculation: by extending , , , Equal features and multiply them, and then pass through the activation function Calculate the attention weight and assign the following formula:

[0039]

[0040] in, is an extension function, is the activation function, also known as the Sigmoid function, is the feature after high directional attention weight, It is the feature after attention weight in width direction; Will and It is divided into G groups in the channel dimension, and the total number of channels is , so that the number of channels in each group is , average pooling is performed on each group of channels, and the assignment operation formula is as follows:

[0041] in, is the average pooling function, After grouping The pooling features, After grouping Pooling features; Sb3) Cross-matrix similarity and feature update: Calculation and , and The cross matrix similarity is calculated and the result is obtained through softmax. The assignment operation formula is as follows:

[0042]

[0043] in, is the matrix multiplication function, It is a normalized exponential function that converts the input into a probability distribution; After normalization and and application After the function, with the input features Multiply and update the final result to obtain fine-grained classification features X cls and fine-grained regression features Xreg , the assignment operation formula is as follows:

[0044]

[0045] S4) Based on the wavelet transform feature decomposition spatial semantic alignment module, the wavelet transform feature is used to capture the underlying texture information, and the feature alignment is performed with the decoupling head to solve the problem of insufficient semantic information in the feature layer extracted by the feature pyramid network (FPN). X The Nth layer features and X cls and X reg Align the N-1th layer features with spatial semantic information; Specifically, in step S4), the wavelet transform feature decomposition spatial semantic alignment module (WTFD) comprises the following steps: Sc1) Deep convolution feature enhancement: input features (C represents the number of channels, H represents the height, and W represents the width) performs a deep convolution operation to enhance the overall correlation and nonlinearity of the features while keeping the input dimension unchanged. The assignment operation formula is as follows:

[0046] In the formula is the depth convolution operation function; is the output feature after deep convolution feature enhancement; Sc2) Haar wavelet decomposition: output features for each channel Perform a first-order Haar wavelet transform and decompose it into a low-frequency component Ac and three high-frequency components (horizontal high-frequency , vertical high frequency , diagonal high frequency ), to capture the underlying texture information and detail information of the image, the formula is:

[0047]

[0048]

[0049]

[0050] In the formula and is the position index of the feature matrix, It is a low-frequency component that captures the overall trend of the feature; It is the horizontal high-frequency component, capturing the high-frequency detail information of the feature in the horizontal direction; It is the vertical high-frequency component, capturing the high-frequency detail information of the feature in the vertical direction; It is the diagonal high-frequency component, which captures the high-frequency detail information of the feature in the diagonal direction; Sc3) High- and low-frequency feature fusion: The low-frequency component and high-frequency component are processed separately through point convolution, and convolution processing and feature fusion are performed on them, and the current layer output is output through WTFD With the features of the previous layer Align and accumulate to obtain the final fused low-frequency features and fused high-frequency features. The formula is as follows:

[0051]

[0052]

[0053] in, is the low-frequency feature after fusion, is the high-frequency feature after fusion, To fuse high and low frequency feature operations; WTFD repeats the above steps, that is, each output is spatially semantically aligned with the N-1 layer.

[0054] S5) Based on the progressive reinforcement loss function, the localization ability of difficult-to-localize categories is improved through iterative reinforcement learning; Specifically, the progressively enhanced loss function includes an enhanced numerator factor function Fl and iteration factor , the formula of the total loss function is:

[0055] in, is the loss function, is the model's predicted probability for the target category (value range [0,1]), is the balance factor, is the regulating factor; According to the length of the data set and the set number of reinforcement learning rounds, refine is iterated to obtain the iteration factor, which is used to adjust the loss function. The specific formula is as follows:

[0056] in, For each classification Ap value, it is a hyperparameter. is the number of samples in the dataset, is the set reinforcement learning round, is the current iteration number.

[0057] S6) The pollution image dataset is trained and input into step S3) and step S4), and then the total loss function value is calculated, and back propagation is performed. The connection weights are optimized through the selected optimizer and corresponding parameters. The first 16 rounds obtain the severe weather adaptive fine-grained ship model, and the last 4 rounds perform enhanced loss function training to obtain the final severe weather adaptive fine-grained ocean remote sensing ship rescue positioning model.

[0058] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0059] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for adaptive fine-grained marine ship rescue positioning based on severe weather, characterized in that: include: S1) Collect ocean remote sensing ship positioning and ocean severe weather remote sensing satellite photo data to build an ocean ship positioning dataset under severe weather conditions; S2) With the help of the pollution element generation module, feature extraction is performed on the pollution elements, and they are fitted with the marine remote sensing ship positioning photos to generate pollution image data sets with different data distributions; S3) A multi-scale linear attention decoupling head based on the region proposal network takes features related to classification features and boundary regression features as input features , the decoupled output is a fine-grained classification feature X cls and fine-grained regression features X reg ; S4) Based on the wavelet transform feature decomposition spatial semantic alignment module WTFD, the wavelet transform feature is used to capture the underlying texture information and align the features with the decoupling head. X The Nth layer features and X cls and X reg Align the N-1th layer features with spatial semantic information; S5) Based on the progressive reinforcement loss function, the localization ability of difficult-to-localize categories is improved through iterative reinforcement learning; S6) The polluted image dataset is trained and input into step S3) and step S4), and then the total loss function value is calculated and back propagated. The connection weights are optimized through the selected optimizer and corresponding parameters to obtain a fine-grained marine ship rescue positioning model based on severe weather adaptation.

2. According to claim 1, a method for adaptive fine-grained marine ship rescue positioning based on severe weather is characterized in that: In step S2), the specific process of generating the contaminated image data set is as follows: Sa1) Embedding and initialization of pollution conditions: Based on the CLIP text encoder, the pollution elements are converted into conditional embedding vectors E, and the remote sensing ship positioning photos are encoded by the VAE of the Stable Diffusion framework to obtain the initial potential expression , and then inject Gaussian noise to generate the noise potential expression , concatenate the noise potential expressions to generate multi-pollution joint features PF; Sa2) Multi-scale pollution feature extraction: Input multi-pollution joint features, build pollution feature pyramids at three scales of 16×16, 32×32 and 64×64 based on U-Net skip connection, align the conditional embedding vector E with the multi-pollution joint features through hierarchical cross attention, and use the feature pyramid network to fuse multi-scale features to simultaneously retain the local texture of the rainstorm and the global morphology of the cloud layer; Sa3) Dynamic pollution intensity control: Based on the dual-path discrimination module, the spatial attention path is used to pass the relative position matrix Control the intensity attenuation of pollution elements above the hull and learn RGB channel weights using channel attention path , and through the learnable parameters through Function fusion to generate dynamic masks ; Sa4) Reverse diffusion and data fusion: Based on the dynamic mask, the noise latent vector is iteratively updated in the latent space through the conditional reverse diffusion process, combining the conditional embedding vector E and the noise prediction network , using VAE decoding to generate polluted images , build a pollution image dataset.

3. The method for marine ship rescue positioning based on severe weather adaptive fine-grained positioning according to claim 1 is characterized in that: In step S3), the input features are processed in the multi-scale linear attention decoupling head. The specific steps of the processing are as follows: Sb1) Multi-scale feature extraction and fusion: Input features Perform average pooling along the width and height dimensions respectively to obtain the feature after average pooling along the height. And the average pooling feature along the width ;Will and After transposition, concatenate along the channel dimension to obtain the concatenated features ;right The application kernel size is The depth convolution obtains multi-scale features in the height direction and multi-scale features in width direction ,in The values ​​are 1, 3, and 5. Determine the receptive field, gradually add the features of the large receptive field to the features of the small receptive field, and the assignment operation formula is as follows: ； ， ； in, is a height feature with a receptive field of 1×1, is a height feature with a receptive field of 3×3, is a height feature with a receptive field of 5×5, is the width feature with a receptive field of 1×1, is a width feature with a receptive field of 3×3, is a width feature with a receptive field of 5×5; Sb2) Multi-scale attention mechanism calculation: by extending , , and And multiply each other, and then pass through the activation function Calculate the attention weight and assign the following formula: ； ； in, is an extension function, is the activation function, is the feature after high directional attention weight, It is the feature after attention weight in width direction; Will and It is divided into G groups in the channel dimension, and the total number of channels is , so that the number of channels in each group is , average pooling is performed on each group of channels, and the assignment operation formula is as follows: ； in, is the average pooling function, After grouping The pooling features, After grouping Pooling features; Sb3) Cross-matrix similarity and feature update: Calculation and , and The cross matrix similarity is calculated and the result is obtained through softmax. The assignment operation formula is as follows: ； ； in, is the matrix multiplication function, It is a normalized exponential function that converts the input into a probability distribution; After normalization and Apply separately After the function, with the input features Multiply and update the final result to obtain fine-grained classification features X cls and fine-grained regression features X reg , the assignment operation formula is as follows: ； 。 4. The method for rescuing marine vessels based on severe weather adaptive fine-grained positioning according to claim 1 is characterized in that: In step S4), the specific steps of the wavelet transform feature decomposition space semantic alignment module are as follows: Sc1) Deep convolution feature enhancement: input features Perform deep convolution operations to enhance the overall correlation and nonlinearity of features while keeping the input dimension unchanged. The assignment operation formula is as follows: ； In the formula is the depth convolution operation function; is the output feature after deep convolution feature enhancement; Sc2) Haar wavelet transform feature decomposition: output features for each channel Perform a first-order Haar wavelet transform and decompose it into a low-frequency component Ac and three high-frequency components. The assignment operation formula is as follows: ； ； ； ； In the formula and is the position index of the feature matrix, It is a low-frequency component that captures the overall trend of the feature; It is the horizontal high-frequency component, capturing the high-frequency detail information of the feature in the horizontal direction; It is the vertical high-frequency component, capturing the high-frequency detail information of the feature in the vertical direction; It is the diagonal high-frequency component, which captures the high-frequency detail information of the feature in the diagonal direction; Sc3) High- and low-frequency feature fusion: The low-frequency component and high-frequency component are processed separately through point convolution, and convolution processing and feature fusion are performed on them, and the current layer output is output through WTFD With the features of the previous layer Align and accumulate to obtain the final fused low-frequency features and fused high-frequency features. The assignment operation formula is as follows: ； ； ; in, is the low-frequency feature after fusion, is the high-frequency feature after fusion, To fuse high and low frequency feature operations; WTFD repeats the above steps, that is, each output is spatially semantically aligned with the N-1 layer.

5. The method for marine ship rescue positioning based on severe weather adaptive fine-grained positioning according to claim 1 is characterized in that: In the step S5), the specific steps are as follows: Progressively reinforced loss function, including reinforced numerator function Fl and the iteration factor , the formula of the total loss function is: ； in, is the loss function, is the model’s predicted probability for the target category, is the balance factor, is the regulating factor; According to the length of the data set and the set number of reinforcement learning rounds, refine is iterated to obtain the iteration factor. The assignment operation formula is as follows: ； in, For each classification Ap value, it is a hyperparameter. is the number of samples in the dataset, is the set reinforcement learning round, is the current iteration number.

6. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images