Adaptive Fine-Grained Ocean Vessel Rescue Positioning Method Based on Severe Weather

Abstract

The present invention discloses a method for self-adaptive fine-grained positioning of ocean vessels based on severe weather, belonging to the technical field of ocean vessel rescue positioning, including: collecting photo data, extracting features of pollution elements, and fitting them with the photo data into a data set; based on a multi-scale linear attention decoupling head, decoupling fine-grained classification features and fine-grained regression features; using wavelet transform features to capture underlying texture information and aligning them with the features of the decoupling head; proposing a loss function, training the model, calculating the total loss function value, performing backpropagation, and optimizing through the selected optimizer and corresponding parameters to obtain the final vessel positioning model. The present invention introduces 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 to enhance the fine-grained learning ability of the method, improve the detection ability and robustness of the model, and ultimately improve the accuracy and efficiency of vessel positioning in severe weather.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine ship rescue positioning, and particularly relates to a fine-grained marine ship rescue positioning method based on adaptation to bad weather. Background Art

[0002] As the vast interconnected waters on the earth's surface, the ocean has a huge impact on human activities. With the acceleration of the processes 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 continuously, posing a huge threat to people's lives and property. In order to reduce the casualties and property losses after maritime accidents, a maritime rescue system has emerged.

[0003] However, in the face of a complex marine background, small targets, and crowded scenes, there have always been problems in maritime rescue such as long distances, long times, and difficult search and rescue. Especially during the search and rescue process, it will be affected by extreme conditions, resulting in a shift in the data distribution of the detected remote sensing images. Therefore, traditional ship positioning algorithms are vulnerable to factors such as heavy rain, sea fog, high exposure, and cloud occlusion, leading to missed detections, false detections, or a significant decline in classification accuracy of targets, making it difficult to achieve precise positioning of marine ships, prolonging the search and rescue time, and preventing shipwrecked ships from being rescued in a timely manner. Summary of the Invention

[0004] Aiming at the deficiencies in the background art, the purpose of the present invention is to propose a fine-grained marine ship rescue positioning method based on adaptation to bad weather. The ship positioning model is redesigned using a fine-grained method to enable it to have the ability to process fine-grained feature levels, thereby enhancing the detection ability and robustness of the model in bad weather. The classification features and boundary regression features are separated to achieve targeted fine-grained feature learning, enhancing the model's ability to capture regions of interest (ROIs) and improving the matching of positive and negative samples. In addition, a progressive reinforcement learning loss function is adopted to enhance the detection ability for difficult-to-detect categories through iterative reinforcement learning, enabling more accurate positioning of ship targets and solving the problem of difficult positioning of currently rescued objects at sea.

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

[0006] A fine-grained marine ship rescue positioning method based on adaptation to bad weather, comprising the following steps:

[0007] S1) Collect photo data of marine remote sensing ship positioning and marine bad weather remote sensing satellites to construct a marine ship positioning data set under bad weather;

[0008] S2) With the aid of a pollution element generation module, extract features for pollution elements and fit them with photos of marine remote sensing ship positioning to generate a polluted image data set with different data distributions;

[0009] S3) A multi-scale linear attention decoupling head based on the region proposal network takes the features related to classification features and boundary regression features as input features, and decouples and outputs fine-grained classification features Xcls and fine-grained regression features Xreg;

[0010] S4) Based on the wavelet transform feature decomposition space semantic alignment module, the wavelet transform feature is used to capture the underlying texture information and align the features with the decoupling head. The Nth layer feature of the feature X after FPN space alignment is aligned with the N-1th layer features of Xcls and Xreg in terms of spatial semantic information;

[0011] S5) Based on the progressive reinforcement loss function, the localization ability of difficult-to-localize categories is improved through iterative reinforcement learning;

[0012] S6) Train the polluted image dataset, input it into steps S3) and S4), then calculate the total loss function value, perform backpropagation, and optimize the connection weights through the selected optimizer and corresponding parameters to obtain an adaptive fine-grained marine ship rescue localization model based on bad weather.

[0013] Preferably, in the step S2), the specific process of generating the polluted image dataset is as follows:

[0014] Sa1) Embedding and initialization of pollution conditions: Based on the CLIP text encoder, the pollution elements are converted into conditional embedding vectors E. The remote sensing ship positioning photos are encoded by the VAE of the Stable Diffusion framework to obtain the initial latent expression , then Gaussian noise is injected to generate the noise latent expression , and the noise latent expressions are concatenated to generate the multi-pollution joint feature PF;

[0015] Sa2) Multi-scale pollution feature extraction: Input the multi-pollution joint feature, based on the U-Net skip connection, construct a pollution feature pyramid at three scales of 16×16, 32×32 and 64×64, align the conditional embedding vector E with the multi-pollution joint feature through hierarchical cross-attention, and use the feature pyramid network to fuse multi-scale features to simultaneously retain the local texture of heavy rain and the global morphology of clouds;

[0016] Sa3) Dynamic pollution intensity control: Based on the dual-path discriminant module, use the spatial attention path to control the intensity attenuation of pollution elements above the hull through the relative position matrix , use the channel attention path to learn the RGB channel weights , and through the learnable parameter via function fusion to generate a dynamic mask ;

[0017] Sa4) Reverse diffusion and data fusion: Based on the dynamic mask, iteratively update the noise latent vector in the latent space through the conditional reverse diffusion process, combine the conditional embedding vector E and the noise prediction network , and use the VAE decoder to generate the contaminated image , and construct a contaminated image dataset.

[0018] Preferably, in step S3), the process of processing the input features in the multi-scale linear attention decoupling head is as follows: The specific steps are as follows:

[0019] Sb1) Multi-scale feature extraction and fusion: Perform average pooling on the input features along the width and height dimensions respectively to obtain the feature after average pooling along the height and the feature after average pooling along the width ; Transpose and , and splice them along the channel dimension to obtain the spliced feature ; Apply a depth convolution with a kernel size of to to obtain the multi-scale feature in the height direction and the multi-scale feature in the width direction , where takes values 1, 3, 5, determines the receptive field, and gradually accumulate the features with a large receptive field to the features with a small receptive field. The assignment operation formula is as follows:

[0020]

[0021] ,

[0022] where is the height feature with a receptive field of 1×1, is the height feature with a receptive field of 3×3, is the height feature with a receptive field of 5×5, is the width feature with a receptive field of 1×1, is the width feature with a receptive field of 3×3, is the width feature with a receptive field of 5×5;

[0023] Sb2) Multi-scale attention mechanism calculation: By expanding , , and features and multiplying them, and then passing through the activation function Calculate the attention weights, and the assignment operation formula is as follows:

[0024]

[0025]

[0026] Among them, is an expansion function, is an activation function, is the feature after the attention weight in the height direction, is the feature after the attention weight in the width direction;

[0027] Divide and into G groups in the channel dimension. The total number of channels is such that the number of channels in each group is Perform average pooling on each group of channels respectively, and the assignment operation formula is as follows:

[0028]

[0029] Among them, is the average pooling function, is the pooled feature of after grouping, is the pooled feature of after grouping;

[0030] Sb3) Cross - matrix similarity and feature update: Calculate and , and 's cross - matrix similarity, and obtain the result through softmax. The assignment operation formula is as follows:

[0031]

[0032]

[0033] Among them, is the matrix multiplication function, is the normalized exponential function, which converts the input into a probability distribution;

[0034] Apply the sum of the normalized and to the function, and then multiply it with the input feature to update the final result, obtaining the fine - grained classification feature X cls and the fine - grained regression feature X reg , and the formula is as follows:

[0035]

[0036]

[0037] Preferably, in the step S4), the specific steps of the wavelet transform feature decomposition space semantic alignment module are as follows:

[0038] Sc1) Deep convolutional feature enhancement: Perform deep convolutional operations on the input features to enhance the overall correlation and non-linearity of the features while keeping the input dimension unchanged. The assignment operation formula is as follows:

[0039]

[0040] where is the deep convolutional operation function; is the output feature after deep convolutional feature enhancement;

[0041] Sc2) Haar wavelet transform feature decomposition: Perform a first-order Haar wavelet transform on the output feature of each channel to decompose it into a low-frequency component Ac and three high-frequency components. The assignment operation formula is as follows:

[0042]

[0043]

[0044]

[0045]

[0046] where and are the position indices of the feature matrix, is the low-frequency component, capturing the overall trend of the feature; is the horizontal high-frequency component, capturing the high-frequency detailed information of the feature in the horizontal direction; is the vertical high-frequency component, capturing the high-frequency detailed information of the feature in the vertical direction; is the diagonal high-frequency component, capturing the high-frequency detailed information of the feature in the diagonal direction;

[0047] Sc3) High-low frequency feature fusion: Process the low-frequency component and the high-frequency components respectively through point convolution, perform convolution processing and feature fusion on them, and align and accumulate the current layer output with the feature of the previous layer to obtain the finally fused low-frequency feature and the fused high-frequency feature. The formula is as follows:

[0048]

[0049]

[0050]

[0051] Among them, is the fused low-frequency feature, is the fused high-frequency feature, is the operation of fusing low and high-frequency features;

[0052] WTFD means repeating the above steps, that is, each output is spatially and semantically aligned with the N-1 layer.

[0053] Preferably, in the step S5), the specific steps are as follows:

[0054] The progressive reinforcement loss function includes the reinforcement molecular factor function Fl and the iteration factor , and the formula of the total loss function is:

[0055]

[0056] Among them, is the loss function, is the predicted probability of the model for the target category, is the balance factor, is the adjustment factor;

[0057] According to the length of the dataset and the set number of reinforcement learning rounds refine, iterative calculation is performed to obtain the iteration factor, and the assignment operation formula is as follows:

[0058]

[0059] Among them, is the Ap value of each classification, which belongs to hyperparameters, is the number of samples in the dataset, is the set number of reinforcement learning rounds, is the current iteration number.

[0060] Preferably, the present invention also proposes a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements a method for self-adaptive fine-grained marine ship rescue positioning based on bad weather.

[0061] Compared with the prior art, the present invention proposes a method for self-adaptive fine-grained marine ship rescue positioning based on bad weather, and the advantages of this method are:

[0062] (1) The present invention redesigns the ship positioning model using a fine-grained method, enabling it to have the ability to process at the fine-grained feature level, thereby enhancing the detection ability and robustness of the model in adverse weather. At the same time, after extracting the features of heavy rain, sea fog, high exposure, and cloud occlusion pollution elements and fusing them with remote sensing ship images, using the relatively stable characteristics of the ocean background and data distribution, the photos with shifted data distribution can be made more natural;

[0063] (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 separated classification features and boundary regression features is analyzed, and the decoupled output is the fine-grained classification feature Xcls and the fine-grained regression feature Xreg, realizing targeted fine-grained feature learning; this improvement enhances the model's ability to capture regions of interest (ROIs), improves the matching of positive and negative samples, and replaces the learning of ordinary convolution with fine-grained attention learning, enabling more accurate positioning of ship targets;

[0064] (3) The present invention introduces a wavelet transform feature decomposition space semantic alignment module, which uses wavelet transform features to capture low-level texture information and aligns the features with the decoupling head in the previous stage to solve the problem of insufficient semantic information extraction of the feature pyramid network (FPN); in specific operations, the Nth layer feature of the feature X aligned in the FPN space is aligned with the N-1th layer features of Xcls and Xreg in terms of spatial semantic information;

[0065] (4) The present invention introduces a training process for iterative reinforcement learning for difficult detection classification and proposes a new loss function to improve the problem of uneven distribution of the sample ratio of difficult detection classification and insufficient texture information features. Description of the Drawings

[0066] Figure 1 is the flowchart of the method for self-adaptive fine-grained ocean ship rescue positioning based on adverse weather of the present invention;

[0067] Figure 2 is the framework diagram of the self-adaptive fine-grained ocean remote sensing ship rescue positioning model proposed by the present invention;

[0068] Figure 3 is the structural block diagram of the multi-scale linear attention decoupling head;

[0069] Figure 4 is the processing flowchart of the wavelet transform feature decomposer and the space semantic alignment module;

[0070] Figure 5 is the comparison diagram of the enhanced effects of heavy rain, sea fog, high exposure, and cloud occlusion based on the ocean background;

[0071] Figure 6 This is the architecture diagram of the adaptive fine-grained marine ship rescue positioning method based on bad weather for the present invention. Specific implementation manners

[0072] Next, the technical solutions in the embodiments of the present application will be further clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. It should be noted that the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0073] In order to make the invention purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings of the specification: In order to better understand the above purposes, features and advantages of the present invention, the advantages of the present invention will be further illustrated through the comparison of embodiments in conjunction with the drawings and specific implementation manners.

[0074] The present invention proposes an adaptive fine-grained marine ship rescue positioning method based on bad weather, and the steps of this method will be described in detail:

[0075] S1) Collect photo data of marine remote sensing ship positioning and marine bad weather remote sensing satellites, and construct a marine ship positioning data set under bad weather;

[0076] S2) With the help of a pollution element generation module, extract features for pollution elements such as heavy rain, sea fog, high exposure, and cloud occlusion, and fit them with marine remote sensing ship positioning photos to generate a pollution image data set with different data distributions;

[0077] Specifically, in the step S2), the pollution element generation module includes the following:

[0078] 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:

[0079] The multi-scale pollution feature generation module performs multi-scale pollution feature extraction on pollution elements such as heavy rain, sea fog, high exposure, and cloud occlusion based on the Stable Diffusion framework, and generates pollution feature maps with different intensity distributions through a cascaded noise prediction network; uses a hierarchical attention mechanism to Concat splice the latent representations of each pollution element to form a multi-pollution joint feature PF;

[0080] The pollution intensity discrimination module inputs the multi-pollution joint feature PF into the intensity adjustment unit composed of the cross-attention mechanism, and through two-stage dynamic weight allocation: in the first stage, the occlusion relationship matrix Q = [q1, q2, q3, q4] (q1, q2, q3, q4 correspond to the pollution elements of heavy rain, sea fog, high exposure, and cloud occlusion) between the pollution elements and the ship target is calculated, and in the second stage, the effective pollution area is screened through the differentiable threshold gating to generate a dynamic pollution mask that conforms to physical laws;

[0081] The element fusion layer uses the adaptive hybrid rendering technology to fuse the pollution feature map and the marine remote sensing ship positioning photo; specifically, through the reverse diffusion process of the Stable Diffusion framework, conditional noise perturbation is applied to the remote sensing ship positioning photo 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 detection task after pollution;

[0082] More specifically, the specific process of the multi-scale pollution feature generation module is as follows:

[0083] 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 skip connection layer of U-Net to obtain pollution pattern maps with different resolutions from 16×16 to 256×256, that is, pollution feature maps; 3) Fuse the multi-scale pollution features through the feature pyramid network to form the multi-pollution joint feature PF, and retain the collaborative expression of the local texture of heavy rain and the global morphology of clouds.

[0084] More specifically, the specific process of the pollution intensity discrimination module is as follows:

[0085] The pollution intensity discrimination module further adopts a dual-path discrimination module, in which the spatial attention path calculates the relative position weight between the pollution area and the ship target, and the channel attention path evaluates the intensity ratio of different pollution types; through the recursive weight update algorithm, the intensity of heavy rain pollution is the largest in the area above the hull, while the sea fog pollution shows a global uniform distribution characteristic;

[0086] Specifically, in step S2), the specific process of generating the pollution image dataset is as follows:

[0087] 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 from pollution elements such as heavy rain, sea fog, high exposure, and cloud occlusion to form a multi-pollution joint feature PF. Through the pollution intensity discrimination module, a dual-path discrimination module is used to optimize and generate a dynamic pollution mask. Finally, through the reverse diffusion and data fusion of the element fusion layer, a pollution image dataset with a natural transition effect is finally output;

[0088] Sa1) Embedding and initialization of pollution conditions: Based on the CLIP text encoder, four types of pollution elements, namely heavy rain, sea fog, high exposure, and cloud occlusion, are transformed into conditional embedding vectors E. The remotely sensed ship positioning photo is encoded by the VAE of the Stable Diffusion framework to obtain an initial latent representation , and then Gaussian noise is injected to generate a noise latent representation . The noise latent representations are concatenated to generate a multi-pollution joint feature PF;

[0089] Sa2) Multi-scale pollution feature extraction: Input the multi-pollution joint feature. Based on the U-Net skip connection, a pollution feature pyramid is constructed at three scales of 16×16, 32×32, and 64×64. The conditional embedding vector E is aligned with the multi-pollution joint feature through hierarchical cross-attention, and the feature pyramid network is used to fuse multi-scale features to simultaneously retain the local texture of heavy rain and the global morphology of clouds;

[0090] Sa3) Dynamic pollution intensity control: Based on the dual-path discrimination module, the spatial attention path is used to control the intensity attenuation of pollution elements above the hull through the relative position matrix , and the channel attention path is used to learn the RGB channel weights , and through the learnable parameter via function fusion to generate a dynamic mask ;

[0091] Sa4) Reverse diffusion and data fusion: Based on the dynamic mask, the noise latent vector is iteratively updated through the conditional reverse diffusion process in the latent space. Combining the conditional embedding vector E and the noise prediction network , the VAE decoding is used to generate a pollution image , and a pollution image dataset is constructed.

[0092] S3) A multi-scale linear attention decoupling head based on the region proposal network, which takes features related to classification features and boundary regression features (such as coordinates, width, and height) as input features , and the decoupled output is X cls fine-grained classification features and Xreg Fine-grained regression features are used to achieve targeted fine-grained feature learning; the ability of the model to capture regions of interest (ROIs) is improved and enhanced, and the matching of positive and negative samples is improved.

[0093] Specifically, in step S3), the process of processing the input features in the multi-scale linear attention decoupling head is as follows:

[0094] Sb1) Multi-scale feature extraction and fusion: For the input features Average pooling is performed separately along the width and height dimensions to initially extract features from different spatial dimensions, obtaining the feature after average pooling along the height and the feature after average pooling along the width ; where and represent the functions for performing average pooling operations along the height and width dimensions; after transposing and and concatenating them along the channel dimension, the concatenated feature is obtained, where is the concatenation function, enabling the integration of feature information in different dimensions along the channel dimension; applying a depth convolution with a kernel size of to results in features and and at different scales (determined by to determine the receptive field), where can take values of 1, 3, or 5. The features with large receptive fields are gradually accumulated onto the features with small receptive fields to achieve the fusion of multi-scale features. The formula is as follows:

[0095]

[0096] ,

[0097] where is the height feature with a receptive field of 1×1, is the height feature with a receptive field of 3×3, is the height feature with a receptive field of 5×5, is the width feature with a receptive field of 1×1, is the width feature with a receptive field of 3×3, is the width feature with a receptive field of 5×5;

[0098] Sb2) Multi-scale attention mechanism calculation: By expanding , , , Multiply the following features and then pass through the activation function Calculate the attention weights. The assignment operation formula is as follows:

[0099]

[0100]

[0101] Among them, is the expansion function, is the activation function, also known as the Sigmoid function, is the feature after the attention weight in the height direction, is the feature after the attention weight in the width direction;

[0102] Divide and into G groups in the channel dimension. The total number of channels is such that the number of channels in each group is Perform average pooling on each group of channels. The assignment operation formula is as follows:

[0103]

[0104] Among them, is the average pooling function, is the pooled feature of after grouping, is the pooled feature of after grouping;

[0105] Sb3) Cross-matrix similarity and feature update: Calculate and , and 's cross-matrix similarity, and obtain the result through softmax. The assignment operation formula is as follows:

[0106]

[0107]

[0108] Among them, is the matrix multiplication function, is the normalized exponential function, which converts the input into a probability distribution;

[0109] Apply the sum of the normalized and to the function and multiply it with the input feature to update the final result and obtain the fine-grained classification feature X cls and the fine-grained regression featureX reg , the assignment operation formula is as follows:

[0110]

[0111]

[0112] S4) Based on the wavelet transform feature decomposition spatial semantic alignment module, use wavelet transform features to capture the underlying texture information and align the features with the decoupled head to solve the problem of insufficient semantic information in the feature layers extracted by the Feature Pyramid Network (FPN); the features after spatial alignment by FPN X The Nth layer feature of X cls and X reg The (N-1)th layer feature of are subjected to spatial semantic information alignment;

[0113] Specifically, in step S4), the wavelet transform feature decomposition spatial semantic alignment module (WTFD) has the following specific steps:

[0114] Sc1) Depth convolution feature enhancement: Perform depth convolution operation on the input feature (C represents the number of channels, H represents the height, and W represents the width) to enhance the overall correlation and nonlinearity of the feature while keeping the input dimension unchanged. The assignment operation formula is as follows:

[0115]

[0116] In the formula is the depth convolution operation function; is the output feature after depth convolution feature enhancement;

[0117] Sc2) Haar wavelet decomposition: Perform the first-order Haar wavelet transform on the output feature of each channel, 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:

[0118]

[0119]

[0120]

[0121]

[0122] In the formula and is the position index of the feature matrix, is the low-frequency component, capturing the overall trend of the feature; is the horizontal high-frequency component, capturing the high-frequency detailed information of the feature in the horizontal direction; is the vertical high-frequency component, capturing the high-frequency detailed information of the feature in the vertical direction; is the diagonal high-frequency component, capturing the high-frequency detailed information of the feature in the diagonal direction;

[0123] Sc3) High-low frequency feature fusion: Process the low-frequency component and the high-frequency component separately through point convolution, perform convolution processing and feature fusion on them, and output the current layer through WTFD Align and accumulate with the features of the previous layer to obtain the finally fused low-frequency feature and the fused high-frequency feature. The formula is as follows:

[0124]

[0125]

[0126]

[0127] Among them, is the fused low-frequency feature, is the fused high-frequency feature, is the operation of fusing high-low frequency features;

[0128] WTFD repeats the above steps, that is, each output is spatially semantically aligned with the N-1 layer.

[0129] S5) Based on the progressive reinforcement loss function, improve the localization ability of difficult-to-localize categories through iterative reinforcement learning;

[0130] Specifically, the progressive reinforcement loss function includes the reinforcement molecular factor function Fl and the iteration factor , and the formula of the total loss function is:

[0131]

[0132] Among them, is the loss function, is the prediction probability of the model for the target category (value range [0,1]), is the balance factor, is the adjustment factor;

[0133] Iteratively calculate according to the length of the dataset and the set number of reinforcement learning rounds refine to obtain the iteration factor for adjusting the loss function. The specific formula is as follows:

[0134]

[0135] Among them, is the Ap value for each classification, which belongs to hyperparameters, is the number of samples in the dataset, is the set number of reinforcement learning rounds, is the current iteration number.

[0136] S6) Train the polluted image dataset, input it into steps S3) and S4), then calculate the total loss function value, perform backpropagation, and optimize the connection weights through the selected optimizer and corresponding parameters. Obtain the harsh weather adaptive fine-grained ship model in the first 16 rounds, and perform reinforcement loss function training in the last 4 rounds to obtain the final harsh weather adaptive fine-grained marine remote sensing ship rescue positioning model.

[0137] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be interpreted to include the preferred embodiments as well as all changes and modifications falling within the scope of the present application.

[0138] 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 equivalent technologies, the present application also intends 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 the 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 operation 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 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