Hyperspectral identification method and system for distress target in complex sea state with enhanced spectral features

Abstract

The application discloses a complex sea condition distress target hyperspectral identification method and system with enhanced spectral features, and the method comprises the following steps: acquiring an initial complex sea condition RGB image in a search and rescue process, converting the initial complex sea condition RGB image into a hyperspectral image through a hyperspectral image generation model, comparing the generated hyperspectral image with a maritime search and rescue target spectral library, identifying a search and rescue target, obtaining the spatial position and size of the search and rescue target, and performing rescue according to the identification result. The application introduces a spectral and spatial double attention module into the hyperspectral image generation model to extract features, and uses a weighted sum of a reconstruction loss and a physical constraint loss as a loss function for training, so that the model can automatically focus on the spatial region of a search and rescue target such as a fallen person or a lifesaving device, suppress background interference such as seawater and waves, enhance the expression ability of spectral features, and ensure the balance between the physical rationality and the visual quality of the generated hyperspectral image.

Description

Technical Field

[0001] This invention relates to the field of maritime search and rescue technology, and in particular to a hyperspectral identification method and system for distressed targets in complex sea conditions with enhanced spatial spectral features. Background Technology

[0002] Maritime search and rescue is a crucial link in ensuring the safety of life and property at sea. However, complex sea conditions such as nighttime, fog, and high waves are the core factors restricting search and rescue efficiency. In such environments, the detection signal of distressed targets (such as people in the water and small life-saving equipment) is weak, and long-distance identification is difficult, directly leading to insufficient success rates in maritime rescue. Traditional visible light surveillance systems are severely limited in performance under these complex conditions. While active detection equipment such as radar can locate large vessels, their detection capability for small targets crucial to search and rescue—such as people in the water, life rings, and small floating debris—is extremely poor, failing to meet the needs of modern maritime search and rescue. Hyperspectral remote sensing technology has the unique advantage of simultaneously acquiring the spectral and spatial information of objects, providing a new technical approach to solving the target detection problem in complex search and rescue environments. Compared with traditional RGB three-channel color images, hyperspectral images not only contain rich spatial information but, more importantly, have nanometer-level spectral resolution, enabling the acquisition of hundreds of continuous narrow-band data to form a complete spectral reflectance curve. Due to differences in internal structure and external shape, objects possess unique spectral characteristics. This characteristic enables precise differentiation of the material differences among various search and rescue targets, such as people in the water (human tissue / clothing), life jackets (polymer materials), ship wreckage (metal / wood), and seawater. Even targets that are small or partially obscured by waves can be effectively identified through spectral characteristics, overcoming the limitations of traditional image recognition that relies on visual morphology and providing "material-level" target information support for search and rescue command and decision-making.

[0003] However, the application of hyperspectral technology in maritime search and rescue still faces technical bottlenecks. First, data acquisition in search and rescue scenarios is difficult. Under complex sea conditions, the acquisition of hyperspectral data is severely limited. Wave-induced platform swaying causes image blurring and registration errors; low light conditions significantly reduce the signal-to-noise ratio; and sea foam and complex lighting environments cause spectral distortion. These problems make it difficult to obtain high-quality hyperspectral data in actual search and rescue scenarios, and the high acquisition cost limits the full realization of the technology's advantages. Second, the utilization of search and rescue target features is insufficient. Traditional hyperspectral maritime identification methods mostly focus on larger targets such as ships, failing to fully exploit the spatial-spectral synergistic features of small targets, multiple materials, and easy obstruction in search and rescue scenarios. Their ability to extract features from small search and rescue targets such as people in the water and lifebuoys is weak. Third, a dedicated spectral library for search and rescue is lacking. Existing spectral libraries are mostly built based on normal sea conditions, lacking systematic spectral data on people in the water, rescue equipment, and floating debris under complex sea conditions. This results in insufficient prior knowledge to support the identification model, making it impossible to establish a reliable correspondence between "spectral features and search and rescue target categories." Fourth, the technology chain is disconnected from search and rescue needs. Data acquisition, feature analysis, and target identification are processed independently, failing to form a closed loop "from data generation in the search and rescue scenario to identification of the distressed target," making it difficult to meet the practical requirements of "rapid response and precise positioning" in maritime search and rescue. Summary of the Invention

[0004] To address the issues of scarce data and insufficient identification capabilities in existing search and rescue scenarios, this invention provides a hyperspectral identification method and system for distressed targets in complex sea conditions with enhanced spatial spectral features, thereby improving the identification efficiency and accuracy of distressed targets under complex sea conditions.

[0005] Therefore, the technical solution adopted by the present invention is as follows: A hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features is provided, characterized in that the method includes: The initial complex sea state RGB image is acquired during the search and rescue process, and the initial complex sea state RGB image is converted into a hyperspectral image through a hyperspectral image generation model. The hyperspectral image generation model is based on a conditional diffusion probability model, and generates hyperspectral images by introducing dual attention modules of spectral and spatial data. The hyperspectral image generation model uses simulated images obtained from physical simulation as the training dataset, and is trained using a weighted sum of reconstruction loss and physical constraint loss as the loss function. A spectral database of maritime search and rescue targets for complex sea conditions is constructed. Hyperspectral images generated during the search and rescue process are compared with the spectral database of maritime search and rescue targets to obtain the spatial location and size of the search and rescue targets. Rescue operations are carried out based on the identification results.

[0006] According to the above scheme, the spectral library of maritime search and rescue targets is obtained in the following ways: Hyperspectral images of typical targets are selected, and the spectral data are filtered and denoised. Then, multivariate scattering correction is performed on the filtered data. For the processed images, regions of interest are selected, and the mean spectral data of the selected regions of interest is calculated to obtain the average spectral curve of the target. The average spectral curves of all typical targets are integrated to establish a spectral library of maritime search and rescue targets in complex sea conditions.

[0007] According to the above plan, the identification of maritime search and rescue targets is achieved through the following methods: Using a spectral angle matching algorithm, the generated hyperspectral image of the distressed target is compared with the spectral library of maritime search and rescue targets to initially screen out the pixel regions of suspected targets and initially separate seawater, ships and small floating objects. Ellipse fitting method is used to identify ship outlines in suspected target pixel regions, calculate ship length and width, and remove ship region pixels. For the hyperspectral image region of distressed targets after removing ships, N-FINDR spectral unmixing technology is applied to extract endmembers of small floating objects. The abundance ratio of each endmember in the region pixels is inverted using the fully constrained least squares method. The spatial location of the small floating objects is determined according to the abundance threshold, and their actual size is calculated in combination with the image spatial resolution to complete the identification of maritime search and rescue targets.

[0008] According to the above scheme, the similarity comparison is specifically calculated based on the angle formed in high-dimensional space between the generated hyperspectral image of the distressed target and the hyperspectral image of the maritime search and rescue target spectral library. Pixel regions with an angle smaller than a certain threshold are regarded as pixel regions of suspected targets. The angle is calculated based on the radiance values ​​of the target pixels in each band of the distressed target hyperspectral image, the radiance values ​​of the reference samples in each band of the maritime search and rescue target spectral library, and the effective number of bands of the hyperspectral sensor of the acquired image.

[0009] According to the above scheme, the denoising network of the hyperspectral image generation model is specifically a denoising network based on the U-Net architecture in the reverse generation process of DDPM. The denoising network based on the U-Net architecture specifically denoises the hyperspectral data after noise is added in the forward diffusion process of DDPM. Among them, the forward diffusion process of DDPM adopts a cosine noise scheduling strategy, which maintains a certain noise addition rate in the early stage of training and increases the noise addition rate in the later stage of training.

[0010] According to the above scheme, after the hyperspectral image generation model generates a hyperspectral image, it is verified by the physical consistency criterion; the physical consistency criterion includes spectral rationality test and radiometric consistency test.

[0011] According to the above scheme, the reconstruction loss is specifically calculated based on the sum of the squares of the differences in each dimension between the predicted noisy image and the hyperspectral image; the physical constraint loss is specifically calculated based on the sum of the absolute errors of each pixel channel between the complex sea state RGB image and the generated hyperspectral image.

[0012] According to the above scheme, the predicted noise is obtained by forward diffusion of the generated hyperspectral image; the complex sea state RGB image is obtained by conversion of hyperspectral image data based on the spectral response characteristics of the RGB camera through integral calculation.

[0013] A hyperspectral identification system for distressed targets in complex sea states with enhanced spatial spectral features is also provided, the system comprising: A hyperspectral image generation module is used to acquire initial complex sea state RGB images during the search and rescue process, and to convert the initial complex sea state RGB images into hyperspectral images through a hyperspectral image generation model. The hyperspectral image generation model is based on a conditional diffusion probability model, and generates hyperspectral images by introducing a dual attention module of spectral and spatial data. The hyperspectral image generation model uses simulated images obtained from physical simulation as the training dataset, and is trained using a weighted sum of reconstruction loss and physical constraint loss as the loss function. The target recognition module is used to build a spectral library of maritime search and rescue targets for complex sea conditions. It compares the hyperspectral images generated during the search and rescue process with the spectral library of maritime search and rescue targets to obtain the spatial location and size of the search and rescue targets, and carries out rescue operations based on the recognition results.

[0014] A computer storage medium is also provided, which stores a computer program that can be executed by a processor. The computer program executes the hyperspectral identification method for distressed targets in complex sea conditions with enhanced spatial spectral features as described above.

[0015] The beneficial effects of this invention are as follows: By introducing a dual attention module (spectral and spatial) into the hyperspectral image generation model to extract features, the model can automatically focus on the spatial region of search and rescue targets such as people in the water and rescue equipment, suppress background interference from seawater and waves, and enhance the expressive power of spectral features, thus optimizing the spatial details and spectral fidelity of the generated hyperspectral images. Furthermore, by using simulated images as the training dataset and a weighted sum of reconstruction loss and physical constraint loss as the loss function, the generated hyperspectral images maintain a balance between physical plausibility and visual quality. Additionally, by comparing the generated images with a constructed spectral library of maritime search and rescue targets, the invention identifies rescue targets, achieving robust target identification under different conditions and significantly improving the accuracy, efficiency, and environmental adaptability of target identification in complex sea conditions.

[0016] Furthermore, the present invention verifies the generated hyperspectral image through physical consistency criteria. The spectral rationality test ensures that the generated hyperspectral data is within the physically possible range, while the radiometric consistency test verifies the consistency between the generated result and the input image, thereby improving the reliability of the generated result and providing a guarantee for the credibility of the present invention in practical applications.

[0017] Furthermore, this invention identifies search and rescue targets by sequentially identifying pixel areas of suspected targets, recognizing and eliminating ship outlines, and finally determining small floating objects. This eliminates the interference of seawater pixels and ship pixels on the identification of small floating objects, and enhances the adaptability and identification accuracy of small targets such as people who have fallen into the water and life-saving equipment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the hyperspectral identification method for distressed targets in complex sea conditions with enhanced spatial spectral features according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of the hyperspectral image generation process according to an embodiment of the present invention; Figure 3 This is a schematic flowchart of the maritime search and rescue target identification process according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the system structure of the hyperspectral identification system for distressed targets in complex sea conditions with enhanced spatial spectral features, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] To address the problems in existing technologies, such as the scarcity of high-quality hyperspectral data in complex sea states, insufficient utilization of the "space-spectral" synergistic characteristics of hyperspectral images, weak spatial focusing and spectral feature enhancement capabilities for small targets in search and rescue, low accuracy in identifying distressed targets in complex sea states, and insufficient adaptability to small targets such as people in the water and rescue equipment, this invention provides a hyperspectral identification method for distressed targets in complex sea states with enhanced space-spectral features, such as... Figure 1 As shown, the method includes: S1. Obtain the initial complex sea state RGB image during the search and rescue process, and convert the initial complex sea state RGB image into a hyperspectral image using a hyperspectral image generation model.

[0021] S2. Construct a spectral database of maritime search and rescue targets for complex sea conditions. Compare the hyperspectral images generated during the search and rescue process with the spectral database of maritime search and rescue targets to obtain the spatial location and size of the search and rescue targets. Conduct rescue operations based on the identification results.

[0022] Specifically, the hyperspectral image generation model is based on the conditional diffusion probability model. It generates hyperspectral images by introducing dual attention modules of spectral and spatial data. The hyperspectral image generation model uses simulated images obtained from physical simulation as the training dataset and is trained using a weighted sum of reconstruction loss and physical constraint loss as the loss function. The simulated images include simulated hyperspectral images and their corresponding simulated RGB images.

[0023] like Figure 2 As shown, before applying the hyperspectral image generation model in practice, the model is first optimized and trained. Specifically, in the data preparation stage, the radiative transfer equation is solved using the radiative transfer model to generate hyperspectral data covering 200 bands in the wavelength range of 400-800nm. This yields a large number of paired, physically realistic simulated hyperspectral images obtained from physical simulation. The generation of each band strictly considers the optical properties of water components, the geometric characteristics of sea waves, and the influence of atmospheric conditions. The radiative transfer model is based on rigorous water optics theory and can accurately simulate the propagation process of light in the marine environment, including the absorption and scattering effects of water, the reflection and transmission effects of the air-sea interface, etc. It can also ensure that the generated data fully reflects the complexity of the real marine environment by setting multiple parameters such as wind speed, suspended matter concentration, and solar altitude angle. Secondly, in the RGB image conversion stage, based on the spectral response characteristics of a real RGB camera, the hyperspectral data is converted into a three-channel RGB image through integral calculation, resulting in the RGB image corresponding to the simulated hyperspectral image. This process accurately simulates the response of spectral signals in different bands on the camera sensor, ensuring the physical authenticity of the generated complex sea state RGB image. Finally, model parameters for calm sea conditions are set to generate corresponding baseline hyperspectral data, providing the model with hyperspectral image references for calm sea conditions under normal calm sea conditions.

[0024] Specifically, during the training phase, the training objective is to establish an accurate mapping relationship from multi-source input data to the target hyperspectral image using deep learning methods. To this end, this embodiment designs an advanced framework based on conditional DDPM, which fully considers the characteristics of hyperspectral data, can effectively handle the special properties of hyperspectral data, and ensures the physical rationality of the generated results.

[0025] Specifically, the advanced framework based on conditional DDPM employs a cosine noise scheduling strategy during the forward diffusion process of the model. In the early stages of training, a low noise addition rate is maintained, which is beneficial for the model to learn detailed features. In the later stages of training, the noise addition rate is appropriately increased to ensure effective denoising. During the reverse generation process, a denoising network based on the U-Net architecture is introduced. This network takes noisy hyperspectral images, time step information, and conditional vectors as inputs to predict the noise components of the current step. The construction of the conditional vectors is particularly important, as it provides strong guidance for the generation process by performing deep feature extraction and fusion on the input calm sea state hyperspectral images and complex sea state RGB images.

[0026] Furthermore, the denoising network based on the U-Net architecture employs an encoder-decoder architecture. The encoder extracts deep semantic features progressively through multiple downsampling stages, while the decoder recovers spatial details through upsampling and skip connections. This embodiment also introduces spectral-spatial attention into the U-Net-based denoising network, adding dual attention modules for both spectral and spatial aspects to fully consider the dual spatial and spectral characteristics of hyperspectral data. Specifically, the spatial attention module calculates the correlation between different locations in the feature map, enabling the model to automatically focus on the spatial region of search and rescue targets such as people in the water and rescue equipment, while suppressing background interference from seawater and waves. The spectral attention module enhances the expressive power of spectral features by analyzing the correlation between features in different bands. These two attention modules are integrated with the backbone network through residual connections, ensuring effective information transmission while achieving focused feature enhancement.

[0027] Furthermore, this hyperspectral image generation model is trained using a weighted sum of reconstruction loss and physical constraint loss as the loss function. The reconstruction loss employs mean squared error to ensure pixel-level consistency between the generated and real images. The physical constraint loss is calculated using a forward radiative transfer model to determine the difference between the generated RGB image (after radiative transfer model transformation) and the input RGB image, ensuring the physical validity of the generated result. This design cleverly integrates physical knowledge into the deep learning framework. Specifically, the loss function can be expressed as:

[0028]

[0029]

[0030] In the formula, Represents the loss function. Indicates the losses incurred during reconstruction. Represents physical constraint loss. This represents the weighting coefficients determined through a grid search, ensuring physical plausibility without compromising visual quality. For mathematical expectation, For diffusion time step, For real images, It is Gaussian noise. For noise prediction networks, The noisy image at time step t is derived from a real hyperspectral image. The intermediate state image containing noise obtained after t-step forward diffusion. For conditional vectors, The square of the L2 norm is used to calculate the sum of squares of the differences in each dimension between the predicted noisy image and the real hyperspectral noisy image; As a forward radiative transfer model, the generated hyperspectral image is converted to RGB space. For the generated hyperspectral image, For input, a complex sea state RGB image, Given the L1 norm, calculate the sum of the absolute errors of each pixel channel between two images.

[0031] After the hyperspectral image generation model is trained, the trained model is applied to practical applications using a sampling generation stage to generate high-quality hyperspectral images from input RGB images of complex sea states and hyperspectral reference images of calm sea states. This embodiment employs the DDIM accelerated sampling strategy, a deterministic sampling method that significantly reduces the number of sampling steps while maintaining generation quality. Traditional DDPM sampling requires thousands of iterations, while the DDIM method reduces the number of sampling steps, greatly improving the method's practicality. During the sampling process, each time step is guided by the aforementioned conditional vector, ensuring that the generation process always proceeds in a direction that meets the input conditions.

[0032] Preferably, after the hyperspectral image generation model generates a hyperspectral image, it is verified using physical consistency criteria. These criteria include spectral rationality checks and radiometric consistency checks. The spectral rationality check ensures that the generated hyperspectral data is within physically possible limits, while the radiometric consistency check verifies the consistency between the generated result and the input RGB image using a forward radiative transfer model. These verification criteria not only improve the reliability of the generated results but also ensure the credibility of the method in practical applications.

[0033] Specifically, the process for identifying maritime search and rescue targets in complex sea conditions is as follows: Figure 3As shown, this embodiment innovatively constructs a maritime search and rescue target spectral database for search and rescue missions in complex sea conditions. Through a dual recognition mechanism combining database-based and database-free recognition, robust identification of maritime targets under different prior knowledge conditions is achieved. Specifically, the maritime search and rescue target spectral database is obtained in the following ways: Hyperspectral images of typical targets are selected, and the spectral data are filtered and denoised. Then, multivariate scattering correction is performed on the filtered data. For the processed images, regions of interest are selected, and the mean spectral data of the selected regions of interest is calculated to obtain the average spectral curve of the target. The average spectral curves of all typical targets are integrated to establish a spectral library of maritime search and rescue targets in complex sea conditions.

[0034] Specifically, targeting a large number of hyperspectral images under complex sea conditions, and focusing on maritime search mission scenarios in complex environments such as sea fog, high waves, and nighttime, this invention, after surveying and statistically analyzing accident cases, selected some representative typical targets in maritime search and rescue missions and constructed a standard spectral library containing objects such as ships (iron), simulated drowning persons (dummies), life rings, life jackets, life rafts, flotation devices, hair, oil booms, foam boards, clothing of distressed persons (cotton), and seawater.

[0035] During construction, the hyperspectral images of the standard spectral library are first preprocessed. First, the Savitzky-Golay smoothing filter is used to filter and denoise the spectral data. Then, multivariate scattering correction is performed on the filtered data to correct for the scattering effects caused by material composition inhomogeneity. For the preprocessed images, a Region of Interest (ROI) is selected, and the mean spectral data of the selected ROI is calculated using MATLAB to obtain the average spectral curve of the target, thus establishing a spectral library of maritime search and rescue targets in complex sea conditions.

[0036] Specifically, this embodiment addresses the "target separation-interference removal-precise identification" technology system for maritime search and rescue, achieving efficient positioning and size estimation of maritime targets. Target identification for maritime search and rescue is achieved through the following methods: Using a spectral angle matching algorithm, the hyperspectral image of the distressed target is compared with the spectral database of maritime search and rescue targets to initially screen out the pixel regions of suspected targets and initially separate seawater, ships and small floating objects. Ellipse fitting method is used to identify ship outlines in suspected target pixel regions, calculate ship length and width, and remove ship region pixels. For the hyperspectral image region of distressed targets after removing ships, N-FINDR spectral unmixing technology is applied to extract endmembers of small floating objects. The abundance ratio of each endmember in the region pixels is inverted using the fully constrained least squares method. The spatial location of the small floating objects is determined according to the abundance threshold, and their actual size is calculated in combination with the image spatial resolution to complete the identification of maritime search and rescue targets.

[0037] Specifically, the Spectral Angle Matching (SAM) algorithm is used to compare the generated hyperspectral image of the distressed target with a spectral database of maritime search and rescue targets to initially screen out pixel regions of suspected targets and exclude most pure seawater pixels. The spectrum of each pixel in the hyperspectral image can be considered as a vector in a high-dimensional space. The similarity between the target spectrum (image pixel) and the reference spectrum (spectrum of floating objects measured in situ) can be measured by the angle formed between them in high-dimensional space. This algorithm assumes that the magnitude of the spectral vector does not affect the similarity. The angle is inversely calculated by calculating the cosine of the angle; the smaller the angle, the closer the shapes of the two spectra are, and the higher the consistency between the target and the reference object. Specifically, this can be expressed as:

[0038] In the formula, Let be the radiance value of the target pixel in the i-th band of the hyperspectral image. Let be the radiance value of the reference sample in the i-th band of the in-situ spectral database. This represents the number of effective bands for the hyperspectral sensor. The calculated spectral angle is shown below.

[0039] Principal component analysis (PCA) was used to reduce the dimensionality of the hyperspectral image, preserving core spectral features and suppressing noise. Based on N-FINDR spectral unmixing technology, "pure endmember" spectra of seawater, ships, and small floating objects were extracted, achieving preliminary separation of the three target types. To address the issue of spectral similarity between ship structures and small floating objects, ellipse fitting was used to identify ship outlines, calculate ship length and width, and remove pixels in the ship region to eliminate their interference with small floating object identification. For the image region after ship removal, N-FINDR was applied again to extract small floating object endmembers. The abundance proportion of each endmember in the pixel was inverted using fully constrained least squares (FCLS), and the spatial location of the small floating objects was determined based on the abundance threshold. Their actual size was estimated by combining the image's spatial resolution.

[0040] Furthermore, this embodiment of the invention also provides a hyperspectral identification system for distressed targets in complex sea states with enhanced spatial spectral features, used in conjunction with the hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features described in this embodiment. Figure 4 As shown, the system includes: The hyperspectral image generation module is used to acquire initial complex sea state RGB images during the search and rescue process, and to convert the initial complex sea state RGB images into hyperspectral images through the hyperspectral image generation model; The target recognition module is used to build a spectral library of maritime search and rescue targets for complex sea conditions. It compares the hyperspectral images generated during the search and rescue process with the spectral library of maritime search and rescue targets to obtain the spatial location and size of the search and rescue targets, and carries out rescue operations based on the recognition results.

[0041] The various modules or mechanisms of the system are mainly used to implement the various steps of the above method embodiments, and will not be described in detail here.

[0042] Furthermore, this embodiment of the invention also provides a computer storage medium storing a computer program that can be executed by a processor. This computer program executes the hyperspectral identification method for distressed targets in complex sea conditions with enhanced spatial spectral features as described above in this embodiment.

[0043] This invention introduces a dual attention module (spectral and spatial) to extract features in the hyperspectral image generation model. This allows the model to automatically focus on the spatial region of search and rescue targets such as people in the water and rescue equipment, suppressing background interference from seawater and waves, and enhancing the expressive power of spectral features. This optimizes the spatial details and spectral fidelity of the generated hyperspectral images. Furthermore, by using simulated hyperspectral images and their RGB images as training datasets and employing a weighted sum of reconstruction loss and physical constraint loss as the loss function, the generated hyperspectral images maintain a balance between physical plausibility and visual quality. The generated images are then compared with a constructed spectral library of maritime search and rescue targets to identify rescue targets. This robust identification of search and rescue targets under different conditions significantly improves the accuracy, efficiency, and environmental adaptability of search and rescue target identification in complex sea conditions.

[0044] Furthermore, in this embodiment of the invention, the generated hyperspectral image is verified through physical consistency criteria. The spectral rationality test ensures that the generated hyperspectral data is within the physically possible range, while the radiometric consistency test verifies the consistency between the generated result and the input image, thereby improving the reliability of the generated result and providing a guarantee for the credibility of the invention in practical applications.

[0045] Furthermore, this embodiment of the invention identifies search and rescue targets by sequentially identifying pixel areas of suspected targets, recognizing and eliminating ship outlines, and finally determining small floating objects. This eliminates the interference of seawater pixels and ship pixels on the identification of small floating objects, and enhances the adaptability and recognition accuracy of small targets such as people who have fallen into the water and life-saving equipment.

[0046] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0047] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0048] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features, characterized in that, The method includes: The initial complex sea state RGB image is acquired during the search and rescue process, and the initial complex sea state RGB image is converted into a hyperspectral image through a hyperspectral image generation model; wherein, the hyperspectral image generation model is based on the conditional diffusion probability model, and generates hyperspectral images by introducing a dual attention module of spectral and spatial data, and uses the weighted sum of reconstruction loss and physical constraint loss as the loss function for training. A spectral database of maritime search and rescue targets for complex sea conditions is constructed. Hyperspectral images generated during the search and rescue process are compared with the spectral database of maritime search and rescue targets to obtain the spatial location and size of the search and rescue targets. Rescue operations are carried out based on the identification results.

2. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 1, characterized in that, The maritime search and rescue target spectral library was obtained through the following methods: Hyperspectral images of typical targets are selected, and the spectral data are filtered and denoised. Then, multivariate scattering correction is performed on the filtered data. For the processed images, regions of interest are selected, and the mean spectral data of the selected regions of interest is calculated to obtain the average spectral curve of the target. The average spectral curves of all typical targets are integrated to establish a spectral library of marine targets in complex sea conditions under maritime search and rescue.

3. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 1, characterized in that, Target identification in maritime search and rescue is achieved through the following methods: Using a spectral angle matching algorithm, the generated hyperspectral image of the distressed target is compared with the spectral library of maritime search and rescue targets to initially screen out the pixel regions of suspected targets and initially separate seawater, ships and small floating objects. Ellipse fitting method is used to identify ship outlines in suspected target pixel regions, calculate ship length and width, and remove ship region pixels. For the hyperspectral image region of distressed targets after removing ships, N-FINDR spectral unmixing technology is applied to extract endmembers of small floating objects. The abundance ratio of each endmember in the region pixels is inverted using the fully constrained least squares method. The spatial location of the small floating objects is determined according to the abundance threshold, and their actual size is calculated in combination with the image spatial resolution to complete the identification of maritime search and rescue targets.

4. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 1 or 3, characterized in that, The similarity comparison is specifically calculated based on the angle formed in high-dimensional space between the generated hyperspectral image of the distressed target and the hyperspectral image of the maritime search and rescue target spectral library. Pixel regions with an angle smaller than a certain threshold are regarded as pixel regions of suspected targets. The included angle is calculated based on the radiance values ​​of the target pixels in each band of the hyperspectral image of the distressed target, the radiance values ​​of the reference samples in each band of the maritime search and rescue target spectral library, and the effective number of bands of the hyperspectral sensor of the acquired image.

5. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 1, characterized in that, The denoising network of the hyperspectral image generation model is specifically a U-Net-based denoising network in the reverse generation process of DDPM. The U-Net-based denoising network specifically denoises the hyperspectral data after noise is added in the forward diffusion process of DDPM. The forward diffusion process of DDPM adopts a cosine noise scheduling strategy, which maintains a certain noise addition rate in the early stage of training and increases the noise addition rate in the later stage of training.

6. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 1, characterized in that, After the hyperspectral image generation model generates a hyperspectral image, it is verified by physical consistency criteria; the physical consistency criteria include spectral rationality test and radiometric consistency test.

7. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 1, characterized in that, The reconstruction loss is specifically calculated based on the sum of the squares of the differences in each dimension between the predicted noisy image and the hyperspectral image; the physical constraint loss is specifically calculated based on the sum of the absolute errors of each pixel channel between the complex sea state RGB image and the generated hyperspectral image.

8. The hyperspectral identification method for distressed targets in complex sea states with enhanced spatial spectral features according to claim 7, characterized in that, The predicted noise is obtained by forward diffusion of the generated hyperspectral image; the complex sea state RGB image is obtained by converting the hyperspectral image data based on the spectral response characteristics of the RGB camera through integral calculation.

9. A hyperspectral identification system for distressed targets in complex sea states with enhanced spatial spectral features, characterized in that, The system includes: A hyperspectral image generation module is used to acquire initial complex sea state RGB images during the search and rescue process, and to convert the initial complex sea state RGB images into hyperspectral images through a hyperspectral image generation model. The hyperspectral image generation model is based on a conditional diffusion probability model, and generates hyperspectral images by introducing a dual attention module of spectral and spatial data. The hyperspectral image generation model uses simulated images obtained from physical simulation as the training dataset, and is trained using a weighted sum of reconstruction loss and physical constraint loss as the loss function. The target recognition module is used to build a spectral library of maritime search and rescue targets for complex sea conditions. It compares the hyperspectral images generated during the search and rescue process with the spectral library of maritime search and rescue targets to obtain the spatial location and size of the search and rescue targets, and carries out rescue operations based on the recognition results.

10. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the hyperspectral identification method for distressed targets in complex sea conditions with enhanced spatial spectral features as described in any one of claims 1-8.

Images