Ice region ship identification method and system combining visible light and short wave infrared remote sensing

Abstract

The invention provides an ice region ship identification method and system combining visible light and short wave infrared remote sensing, and the method comprises the following steps: S1, obtaining a visible light image and a short wave infrared image of a polar region channel under the same time and space in a remote sensing image library; s2, performing radiation correction, atmospheric correction, image sampling and image fusion processing on the visible light image and the short-wave infrared image to obtain a data set; s3, rotating frame labeling is conducted on ships in the data set, and a ship sample set is obtained; s4, inputting the training sample set into a ship identification model, performing iterative training on the model, and optimizing the ship identification model by using the verification sample set to obtain a ship identification model with a high fitting degree; and S5, detecting the polar navigation channel image to be identified by using the ship identification model with a relatively high fitting degree to obtain a ship identification result. The polar navigation channel ship identification method has high precision and efficiency in polar navigation channel ship identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent maritime affairs, in particular, especially relates to an ice area ship identification method and system combining visible light and short-wave infrared remote sensing. BACKGROUND

[0002] With global warming, shipping routes and oil and gas exploration and production are moving to higher latitudes. Sea ice brings great challenges to the shipping industry and marine resource development. Ship ice accidents and other situations pose a great threat to ship safety; the infrastructure on high-latitude routes is not yet perfect, and support services for navigation safety need to be strengthened. Developing research on ship identification in polar routes based on remote sensing images can provide ship information in the polar route environment, which is of great significance for maritime search and rescue, ship safety, high-latitude route development, and marine ecological environment protection.

[0003] Existing ship target identification technology is mostly based on open sea level, and there is little research on the identification of ships in polar routes. Polar routes have unique environmental characteristics, which pose many challenges to ship identification. Currently, although some research has proposed polar route identification based on visible light remote sensing images, there are still limitations in the accurate identification of ship targets. The cracking of ice layers and ice fragments exhibit similar features to ships in vision. The shape, texture, and other features of ice fragments and ships may have a large overlap and interference. This makes the boundary of the ship in the remote sensing image fuzzy, affecting the accurate identification and positioning of the ship, and leading to a significant reduction in the application effect of traditional single-image ship identification methods in the polar route environment. Especially under low image quality conditions, the extraction of ship targets becomes more difficult.

[0004] In recent years, due to the development of various sensors, research on multi-source image fusion identification has gradually emerged. Multi-source image fusion technology can fully combine the advantages of different data sources, making the target features more obvious, especially in complex environmental conditions. Visible light images make it easy to confuse targets with ice layers and ice gaps. Infrared images can provide clear target information in low visibility, at night, and in complex weather conditions. Especially in the polar route environment, infrared images can effectively identify ship targets, make up for the missing visual features of optical images, and help reduce misjudgments caused by sea ice. SUMMARY

[0005] Therefore, the purpose of the present application is to provide an ice area ship identification method and system combined with visible light and short-wave infrared remote sensing, which can retain environmental information while enhancing the target by fusing the features of visible light and infrared images, helping to improve the visibility of the target, reduce the interference of the background on the detection task, improve the identification ability of the detection model for low-contrast targets, and solve the problem of missed detection caused by relying solely on visible light data for ship detection, especially in the polar channel environment.

[0006] The technical means adopted by the present application are as follows: An ice area ship identification method combined with visible light and short-wave infrared remote sensing, comprising the following steps: S1. In the remote sensing image library, obtain visible light images and short-wave infrared images under the same space-time in the polar channel; S2. After radiometric correction, atmospheric correction, image sampling and image fusion processing of the visible light images and short-wave infrared images, use them as a data set; S3. Rotate the frame and label the ships in the data set to obtain a ship sample set, and divide the ship sample set into a training sample set and a verification sample set according to a predetermined proportion; S4. Input the training sample set into the ship identification model, iteratively train the model, optimize the ship identification model using the verification sample set, and obtain a ship identification model with high fitting degree; S41. Input the registered visible light images and short-wave infrared images in the training sample set into the two modal branches of the double-flow structure respectively; first, pre-process them through the frequency domain modal enhancement module SFA to suppress background interference and enhance ship edge response; the frequency domain enhanced feature maps generated after SFA processing are sent to the respective backbone networks for multi-scale hierarchical feature extraction to obtain modal semantic feature maps; S42. The modal semantic feature maps extracted in the visible light and infrared modalities are combined to jointly express the shallow structure information and deep semantic information, forming a fusion feature map containing multi-scale and multi-frequency features; S43. Input the fusion feature map into the detector to identify the ship target; S44. During the training process, periodically evaluate the model performance using the verification sample set, adjust the learning rate and training strategy according to the verification set accuracy; after training is completed, output a ship identification model with high fitting degree; S5. Use the ship identification model with high fitting degree to detect the polar channel image to be identified to obtain the ship identification result.

[0007] Further, in S41, the frequency domain modal enhancement module SFA includes: a lightweight high-low frequency separation sub-module EFSM, a short-wave infrared high-frequency attention mechanism sub-module FGA-S, and a visible light modal high-low frequency control sub-module FAM-VI.

[0008] Further, the lightweight high-low frequency separation sub-module EFSM performs average pooling on the input image, removes local high-frequency details, and obtains low-frequency features; then, the original image is upsampled to the original image size through transposed convolution instead of the interpolation process; the high-frequency components are extracted by performing difference between the original image and the upsampled image, and the calculation process is as follows:

[0009]

[0010]

[0011] wherein I is an input image, , is a batch size of the image, is a channel number of the image, represents a height of the image, represents a width of the image, represents a low-frequency component, represents an upsampled low-frequency component, represents a high-frequency component, represents an average pooling layer, represents a transposed convolution, k=3.

[0012] Further, the short-wave infrared high-frequency attention mechanism sub-module FGA-S introduces batch normalization to enhance the distribution stability of the high-frequency features separated from the short-wave infrared; then, the positive part in the high-frequency component is activated through a ReLU function, and the negative value response is suppressed; then, the Sigmoid function is used to map it to the attention interval of [0, 1]; finally, a threshold is introduced combined with a gating mask, the high-confidence area is retained, the weights less than the threshold are set to zero, the background interference is removed, the key area is guided to focus, the attention weight is obtained, and the calculation process is as follows:

[0013]

[0014] wherein, represents a high-frequency component of a short-wave infrared image, is an attention matrix obtained, represents batch normalization, represents a linear rectifier function, is an indicator function; is a threshold, and the weights less than the threshold are set to zero, and is ; represent the Hadamard product operation.

[0015] Further, the visible light mode high-low frequency control sub-module FAM-VI, for high-frequency information Processing path, using group convolution in guaranteeing calculation efficiency, smoothing through local filtering, reducing the local influence of high-frequency noise, obtaining the smoothed feature , the calculation formula is as follows:

[0016] Among them, representing group convolution, setting 4 groups; Using the reverse mechanism to suppress the ice layer area with significant high frequency, the calculation formula is as follows:

[0017]

[0018] Among them, representing global average pooling, and is a learnable parameter, , C is the number of channels, and r is the channel reduction rate, and represent the activation function, representing the feature after channel gating; Spatial gating weight Through a larger convolution operation to learn the global spatial relationship and generate, and the feature is subjected to Hadamard product operation, to obtain the high-frequency suppressed feature , the calculation formula is as follows:

[0019]

[0020] Among them, is the average-pooled input for each channel, representing convolution; For the separated low-frequency information path, introduce MBB module; in keeping the stability of global structure information, fusion multi-scale feature, is conducive to enhancing the representation ability of small target, the calculation formula is as follows:

[0021]

[0022]

[0023] wherein, represents one convolution, using as an activation function, and are two learnable parameters; The inhibited high frequency and the inhibited low frequency information are fused, and the calculation formula is as follows:

[0024]

[0025] wherein, , represents the fusion weight of low frequency and high frequency, and satisfies .

[0026] The application also provides an ice area ship identification system combining visible light and short-wave infrared remote sensing, which is used to realize any one of the ice area ship identification methods combining visible light and short-wave infrared remote sensing. An image acquisition module is configured to acquire visible light images and infrared images of the same space-time coverage area. An image processing module is configured to perform radiation correction, atmospheric correction, image sampling and image fusion processing on the images obtained by the image acquisition module, and perform a rotated frame label on the ships in the images to obtain a ship sample set. A model training module is configured to train the improved model using the labeled visible light and infrared image data to obtain a model with polar channel ship identification capability. A polar channel ship identification module is configured to load the trained polar channel ship identification model, infer new input visible light and infrared images, and accurately identify the positions and confidence of the ships in the images. A result output module is configured to output the polar channel ship detection result, including the target position, the confidence, the superimposed result of the identified image and the original image, and support the loading of a map visualization platform.

[0027] The application also provides a storage medium, which comprises a stored program, wherein the program, when executed, performs any one of the ice area ship identification methods combining visible light and short-wave infrared remote sensing.

[0028] The application also provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor performs any one of the ice area ship identification methods combining visible light and short-wave infrared remote sensing by executing the computer program.

[0029] Compared with the prior art, the application has the following advantages: The application effectively alleviates the recognition difficulty caused by ice and snow background interference, low light environment and target edge blur by introducing short-wave infrared remote sensing images, and modal complementation with visible light images, and enhances the detection stability of the model for ship targets under extreme conditions. The application innovatively designs a frequency domain modal enhancement module SFA, combines a high-frequency attention module FAG-S of a short-wave infrared image with a high-low frequency control module FAM-VI of a visible light image, and realizes efficient cooperation of multi-modal features through a frequency domain guidance mode. The method can be trained end to end under a deep learning framework, has high automation degree, fast inference speed, can be integrated into a remote sensing image automatic processing system, and is suitable for use in application scenarios such as Arctic channel monitoring and polar shipping safety management. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0031] Figure 1 It is a method flowchart of the application. Figure 2 It is a polar shipping channel ship identification model network structure schematic diagram of the application. Figure 3 It is a visible light image processing flowchart of the frequency domain modal enhancement module SFA in the application. Figure 4 It is a short-wave infrared image processing flowchart of the frequency domain modal enhancement module SFA in the application. Figure 5 It is a polar shipping channel ship identification system structure schematic diagram of the application. Figure 6a It is a visible light image ship identification result schematic diagram of the application. Figure 6b It is a short-wave infrared image ship identification result schematic diagram of the application. DETAILED DESCRIPTION

[0032] In order to better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, 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 labor should fall within the scope of protection of the present application.

[0033] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0034] As shown in Figure 1 The present application provides a polar shipping route identification method combining visible light and short-wave infrared remote sensing, comprising the following steps: S1. In the remote sensing image library, obtain visible light images and short-wave infrared images under the same space-time in the polar shipping route; S2. After radiation correction, atmospheric correction, image sampling and image fusion processing of the images obtained in S1, the images are used as a data set; Specifically, Sentinel-2 images of the Arctic shipping route region from December to January are selected in the remote sensing image library to determine the target region. The Sentinel-2 images are subjected to radiation correction and atmospheric correction. Visible light image data is made by selecting B2, B3 and B4, and short-wave infrared image is selected as B11 and resampled to 10m resolution. Image fusion is performed in the order of B2, B3, B4 and B11. All images are uniformly cropped to 400x400 pixels in size and used as a data set.

[0035] S3. The ships in the data set are labeled with a rotating frame to obtain a ship sample set, and the ship sample set is divided into a training sample set and a verification sample set according to a predetermined proportion; Specifically, the high complexity of sea ice background and the interference of a large number of floating ice blocks may affect the judgment of the ship. Therefore, in order to ensure the accuracy and consistency of the labeled data, a multi-stage labeling process is used: in the first stage, the joint data set is initially labeled. In the labeling process, the short-wave infrared image needs to be referred to to ensure one-to-one correspondence of the target. When drawing the rotating box, make the target box consistent with the direction of the ship, and try to make the rectangular box tightly surround the ship target. In the second stage, the labeling of the first stage is reviewed, and the problems are returned to the first stage for solution. In the third stage, expert reviewers review, and if the reviewers have doubts about the labeling in the image, the first stage is performed again for the image. After labeling is completed, the obtained ship sample set is divided into a training sample set and a verification sample set according to a 7:3 ratio.

[0036] S4, input the training sample set into the ship identification model, iteratively train the model, optimize the ship identification model by using the verification sample set, and obtain a ship identification model with high fitting degree; Specifically, please refer to Figure 2 , Figure 2 The polar route ship identification model network structure diagram of the application, the ship identification model training includes the following steps: S41, input the registered visible light image and short-wave infrared image into two modal branches of a double-flow structure respectively. First, pre-process through a frequency domain modal enhancement module SFA to suppress background interference and enhance ship edge response respectively. The frequency domain enhanced feature map generated after SFA processing is sent into the respective backbone network for multi-scale hierarchical feature extraction to obtain modal semantic feature maps; Specifically, the frequency domain modal enhancement module SFA in the ship identification model includes: a lightweight high-low frequency separation sub-module EFSM, a short-wave infrared high-frequency attention mechanism sub-module FGA-S, and a visible light modal high-low frequency control sub-module FAM-VI. From the frequency domain feature analysis of visible light and short-wave infrared, it can be seen that the visible light image provides rich color information, but in the complex ice background and low resolution conditions, the ship features are easily affected by sea ice; while the short-wave infrared image reduces the influence of the ice layer on the ship and provides a clear ship position. However, some data are limited by imaging conditions and sensor characteristics and have edge blur and insufficient contrast. These factors affect the effect of single-image detection and fusion detection. By enhancing the perception of the frequency domain, more hidden information is extracted from the frequency domain to enhance the visual ability.

[0037] Specifically, the ship identification model is an improved double-flow S 2 A-Net.

[0038] Specifically, in S4, the lightweight high-low frequency separation sub-module EFSM in the ship identification model performs average pooling on the input image, removes local high-frequency details, and obtains low-frequency features. Then, the original image size is restored by transposed convolution instead of interpolation process upsampling. The original image and the upsampled image are differentiated to extract high-frequency components. The calculation process is as follows:

[0039]

[0040]

[0041] wherein I is an input image, wherein, is the batch size of the image, is the number of image channels, and represent the height and width of the image, respectively, represents a low-frequency component, represents an upsampled low-frequency component, represents a high-frequency component, represents an average pooling layer, represents transposed convolution, k=3.

[0042] S42, the multiscale features (modal semantic feature maps) extracted in the visible light and infrared modalities are fused, so that the shallow structure information and the deep semantic information are jointly expressed, forming a fusion feature map containing multiscale and multi-frequency features. Please refer to Figure 4 , Figure 4 is the short-wave infrared image processing flowchart of the frequency domain modal enhancement module SFA in the present application.

[0043] Specifically, the short-wave infrared high-frequency attention mechanism sub-module FGA-S has the following process: For the high-frequency features separated from the short-wave infrared, batch normalization is introduced to enhance the stability of the distribution. Then, the positive part of the high-frequency component is activated by the ReLU function, and the negative response is suppressed. Then, the Sigmoid function is used to map it to the attention interval [0, 1]. Finally, in order to remove non-significant information with low weights, a threshold is introduced combined with a gating mask to retain high-confidence areas, and weights less than the threshold are set to zero to remove background interference, guide the focus area, and obtain attention weights. The calculation process is as follows:

[0044]

[0045] wherein, represents a high-frequency component of a short-wave infrared image, is the attention matrix, represents batch normalization (BatchNorm), represents linear rectifier function (ReLU), is an indicator function, is a threshold value, less than which is zero, here , represents Hadamard product operation.

[0046] Please refer to Figure 3 , Figure 3 is the visible light image processing flowchart of the frequency domain modal enhancement module SFA in the application.

[0047] Specifically, the high and low frequency control sub-module FAM-VI of the visible light modal, the flow is as follows: For high frequency information processing path, using group convolution, while ensuring the calculation efficiency, through local filtering for smoothing, reducing the local influence of high frequency noise, and retaining local important details, getting the smoothed feature , the calculation formula is as follows:

[0048] wherein, represents group convolution, set into 4 groups.

[0049] In order to suppress the ice layer area with significant high frequency, using the reverse mechanism, through the reverse processing to achieve suppression, the calculation formula is as follows:

[0050]

[0051] wherein, represents global average pooling, and are learnable parameters, , C is the number of channels, r is the channel reduction rate, in order to reduce the amount of calculation, here set to 8. and represent the activation function, represents the feature after channel gating.

[0052] Spatial gating weight Through the larger convolution operation to learn the global spatial relationship and generate, and the feature is Hadamard product operation, getting the high frequency suppressed feature , the calculation formula is as follows:

[0053]

[0054] wherein, is the input after average pooling for each channel. represents convolution.

[0055] For the separated low-frequency information path, the MBB module is introduced. In order to keep the stability of global structure information and fuse multi-scale features, it is beneficial to enhance the representation ability of small targets. The calculation formula is as follows:

[0056]

[0057]

[0058] wherein, represents a convolution, using function as the activation function, and are two learnable parameters.

[0059] Finally, the suppressed high-frequency and the suppressed low-frequency information are fused, and the calculation formula is as follows:

[0060]

[0061] wherein , indicates the fusion weight of low frequency and high frequency, and satisfies .

[0062] S43, input the fused feature map into the detector to identify the ship target; S44, in the training process, the model performance is evaluated periodically using the verification sample set, and the learning rate and the training strategy are adjusted according to the verification set accuracy to avoid overfitting. After training, a ship identification model with high fitting degree is output; Specifically, the precision curve generated in the training process is analyzed to check whether overfitting occurs and whether the precision meets the requirements of ship identification. The threshold is generally set to 0.90, and greater than or equal to 0.90 is qualified and less than 0.90 is unqualified. If qualified, the training is ended, otherwise the parameters are adjusted and the training is continued.

[0063] The specific training method is as follows: A learning rate strategy with warm-up and phased decay is adopted: initially, linear warm-up is used, gradually increasing the learning rate from one-third of the initial value to the set learning rate of 0.001 in the first 500 iterations to avoid model instability at the beginning of training; then, a step strategy is adopted, decaying the learning rate (usually divided by 10) at the 16th and 22nd epochs respectively, so as to finely adjust the model parameters in later stages, thereby improving model performance. The SGD optimizer (momentum 0.9, weight decay 0.0001) is used, and gradient clipping is added to prevent gradient explosion. The entire training process lasts for 24 epochs.

[0064] S5. Using the ship identification model with a high fitting degree obtained in S4, the polar waterway image to be identified is detected to obtain the ship identification result. Please see Figure 6a , Figure 6b , Figure 6a This is a schematic diagram of the visible light image ship recognition results of the present invention. Figure 6b This is a schematic diagram of the ship identification results from shortwave infrared images according to the present invention. It can be seen that the polar waterway ship identification method combining visible light and shortwave infrared remote sensing provided by the present invention can achieve high detection accuracy and efficiency in identifying ships from remote sensing images of polar waterway areas.

[0065] like Figure 5 As shown, the present invention also provides a polar waterway vessel identification system combining visible light and shortwave infrared remote sensing, the system comprising: The image acquisition module acquires visible light and infrared images of the same spatiotemporal coverage area; The image processing module performs radiometric correction, atmospheric correction, image sampling, and image fusion processing on the images obtained by the image acquisition module, and adds rotating bounding boxes to the ships in the images to obtain a ship sample set. The model training module uses labeled visible light and infrared image data to train the improved model and obtain a model with the ability to identify ships in polar waterways. The polar waterway vessel identification module loads a pre-trained polar waterway vessel identification model and performs inference on newly input visible light and infrared images to accurately identify the position and confidence level of vessels in the images. The results output module outputs the detection results of ships in polar waterways, including target location and confidence level, and exports the result of overlaying the recognized image with the original image, which supports loading on map visualization platforms.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying ships in ice-covered areas using a combination of visible light and shortwave infrared remote sensing, characterized in that, Includes the following steps: S1. Obtain visible light and shortwave infrared images of the polar shipping route in the same time and space from the remote sensing image database. S2. After performing radiometric correction, atmospheric correction, image sampling, and image fusion processing on visible light images and shortwave infrared images, the resulting dataset is used. S3. Rotate the bounding boxes to label the ships in the dataset to obtain the ship sample set. Divide the ship sample set into a training sample set and a validation sample set according to a preset ratio. S4. Input the training sample set into the ship recognition model, iteratively train the model, and optimize the ship recognition model using the validation sample set to obtain a ship recognition model with high fitting degree. S41. The visible light image and shortwave infrared image after the training samples are registered together are used as inputs and sent to the two modal branches of the dual-stream structure. First, they are preprocessed by the frequency domain modal enhancement module SFA to suppress background interference and enhance the ship edge response respectively. The frequency domain enhanced feature maps generated after SFA processing are fed into their respective backbone networks for multi-scale hierarchical feature extraction to obtain modal semantic feature maps. S42. The modal semantic feature maps extracted from the visible light and infrared modes are fused to enable the shallow structural information and deep semantic information to be jointly expressed, forming a fused feature map containing multi-scale and multi-frequency features. S43. Input the fused feature map into the detector to identify ship targets; S44. During the training process, the model performance is periodically evaluated using the validation sample set, and the learning rate and training strategy are adjusted according to the accuracy of the validation set. After training, a ship recognition model with high fitting accuracy is output. S5. Use a ship identification model with high fitting degree to detect the polar waterway image to be identified and obtain the ship identification results.

2. The method for identifying ships in ice-covered areas using combined visible light and shortwave infrared remote sensing according to claim 1, characterized in that, In S41, the frequency domain modal enhancement module SFA includes: the lightweight high and low frequency separation submodule EFSM, the short-wave infrared high-frequency attention mechanism submodule FGA-S, and the visible light modal high and low frequency control submodule FAM-VI.

3. The method for identifying ships in ice-covered areas using combined visible light and shortwave infrared remote sensing according to claim 2, characterized in that, The lightweight high-low frequency separation submodule EFSM performs average pooling on the input image to remove local high-frequency details and obtain low-frequency features. Then, the image is upsampled back to its original size using transposed convolution instead of interpolation. The original image and the upsampled image are then subtracted to extract high-frequency components. The calculation process is as follows: Where I is the input image, , For image batch size, The number of image channels. Indicates the height of the image. Indicates the width of the image. Represents low-frequency components. This represents the low-frequency component after upsampling. Represents high-frequency components. Represents the average pooling layer. This represents the transposed convolution, k=3.

4. The method for identifying ships in ice-covered areas using combined visible light and shortwave infrared remote sensing according to claim 2, characterized in that, The shortwave infrared high-frequency attention mechanism submodule FGA-S introduces batch normalization to enhance the distribution stability of the high-frequency features separated from the shortwave infrared; then, it activates the positive part of the high-frequency components and suppresses the negative response through the ReLU function; and finally, it maps them to the attention interval [0,1] with the help of the Sigmoid function. Finally, a threshold combined with a gating mask is introduced to retain high-confidence regions, while weights below the threshold are reset to zero. This removes background interference, guides attention to key regions, and yields attention weights. The calculation process is as follows: in, High-frequency components representing shortwave infrared images. This is to obtain the attention matrix. Represents batch normalization. Represents the linear rectification function. For indicator functions; Set as a threshold; values ​​below this threshold are set to zero. ; This represents the Hadamard product operation.

5. The method for identifying ships in ice-covered areas using combined visible light and shortwave infrared remote sensing according to claim 2, characterized in that, The visible light mode high and low frequency control submodule FAM-VI, for high frequency information The processing path uses grouped convolution to ensure computational efficiency, and local filtering is used for smoothing to reduce the local influence of high-frequency noise, resulting in smoothed features. The calculation formula is as follows: in, This represents grouped convolution, set to be divided into 4 groups; The inverse mechanism is used to suppress significantly high-frequency ice regions, and the calculation formula is as follows: in, Represents global average pooling. and For learnable parameters, C is the number of channels, and r is the channel reduction rate. and Represents the activation function. The characteristics represented after channel gating; Spatial gating weights The features are generated by learning global spatial relationships through large convolution operations, and then subjected to Hadamard product with the features to obtain high-frequency suppressed features. The calculation formula is as follows: in, It is the input after average pooling of each channel. represent Convolution; For the separated low-frequency information The path is determined by introducing the MBB module; while maintaining the stability of global structural information, the fusion of multi-scale features enhances the representation capability of small targets. The calculation formula is as follows: in, Represents a Convolution, using The function is used as an activation function. and These are two learnable parameters; The suppressed high-frequency and suppressed low-frequency information are fused together, and the calculation formula is as follows: in, , representing the fusion weight of low and high frequencies, satisfying .

6. A combined visible light and shortwave infrared remote sensing system for identifying ships in ice-covered areas, used to implement the combined visible light and shortwave infrared remote sensing method for identifying ships in ice-covered areas as described in any one of claims 1-5, characterized in that, include: The image acquisition module acquires visible light and infrared images of the same spatiotemporal coverage area; The image processing module performs radiometric correction, atmospheric correction, image sampling, and image fusion processing on the images obtained by the image acquisition module, and adds rotating bounding boxes to the ships in the images to obtain a ship sample set. The model training module uses labeled visible light and infrared image data to train the improved model and obtain a model with the ability to identify ships in polar waterways. The polar waterway vessel identification module loads a pre-trained polar waterway vessel identification model and performs inference on newly input visible light and infrared images to accurately identify the position and confidence level of vessels in the images. The results output module outputs the detection results of ships in polar waterways, including target location and confidence level, and exports the result of overlaying the recognized image with the original image, which supports loading on map visualization platforms.

7. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it performs the ice-zone ship identification method based on combined visible light and shortwave infrared remote sensing as described in any one of claims 1 to 5.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the ice-zone ship identification method based on the combined visible light and shortwave infrared remote sensing as described in any one of claims 1 to 5 through the computer program.

Images