Three-dimensional reconstruction method for underwater environment

Abstract

The present invention relates to a method for three-dimensional reconstruction of an underwater environment. Project multi-band structured light patterns in the underwater environment, and at the same time use a multispectral camera to capture images; through a spectral matching algorithm, fuse images of different bands, extract depth information to generate basic data for three-dimensional reconstruction; use depth compensation technology to correct depth errors caused by underwater scattering and absorption; use underwater sensors to monitor environmental parameters such as light, turbidity, and temperature in real time to generate an environmental feature model; according to the environmental feature model, adaptively adjust image acquisition and processing parameters including exposure, contrast, and color balance; through an intelligent correction algorithm, detect distorted areas in the image caused by environmental changes, and perform local correction and overall optimization; generate a multi-scale image pyramid, and decompose the structural information of the three-dimensional image layer by layer from coarse to fine; through an adaptive detail enhancement technology, refine key areas to generate an underwater three-dimensional image.

Description

Technical Field

[0001] The present invention relates to the field of underwater environment modeling, and particularly to a method for three-dimensional reconstruction of an underwater environment. Background Art

[0002] Although the current methods for three-dimensional reconstruction of underwater environments have made some progress in technology, there are still some significant deficiencies and drawbacks in practical applications. These problems have largely limited the application effect of three-dimensional reconstruction technology in underwater environments. First, the complexity of the underwater environment poses a huge challenge to traditional three-dimensional reconstruction methods. Underwater optical imaging is affected by light attenuation, scattering, and absorption, and these factors become more complex with changes in depth, turbidity, and ambient light conditions. Traditional methods usually rely on single-band optical imaging or sonar imaging technology, and these methods perform poorly in highly turbid waters and are easily affected by enhanced light attenuation and scattering, resulting in inaccurate or lost depth information. At the same time, traditional three-dimensional reconstruction methods often assume that the propagation of light in water is uniform, but in practical applications, the heterogeneity and dynamic changes of water bodies make this assumption invalid, resulting in relatively large errors in the reconstruction results.

[0003] Secondly, most of the current three-dimensional reconstruction methods rely on static or offline environmental parameter monitoring and modeling, lacking the ability to dynamically respond to real-time data. When dealing with rapidly changing conditions such as light, turbidity, and temperature in the underwater environment, this method often cannot quickly adjust the modeling strategy, resulting in the generated three-dimensional model being prone to distortion or instability. Traditional technologies often use fixed calibration parameters in the underwater environment, and these parameters are difficult to adapt to complex and changing underwater conditions, especially when there are significant differences in environmental conditions at different depths and regions. The generalization ability of the model is poor and the performance is unstable. The application of fixed parameters will result in obvious differences in the performance of the model under different water conditions, and this inconsistency seriously affects the accuracy and reliability of three-dimensional reconstruction. In addition, traditional methods usually adopt a single-level modeling and optimization strategy, ignoring the changes in environmental features at different scales and being difficult to capture subtle local features, resulting in insufficient global consistency and local detail retention ability of the model. Especially in complex underwater environments, local feature changes caused by factors such as water flow and particulate matter will have a significant impact on the effect of three-dimensional reconstruction, and a single-level modeling method is difficult to cope with this diversity and variability. In terms of image processing and feature extraction, traditional methods mainly rely on manually designed feature extraction methods, which usually extract simple low-level features (such as edges, textures, etc.) and are difficult to effectively capture complex non-linear distortion features. This simple data fusion method often performs poorly in the face of complex environmental changes and is difficult to achieve in-depth analysis and optimization of multi-source data, resulting in inaccurate and unstable reconstruction results. In terms of intelligence level, there are still significant deficiencies in traditional underwater three-dimensional reconstruction technology.

[0004] Currently, most methods rely mainly on preset rules and parameters when dealing with image distortion and environmental changes, lacking the ability of adaptive learning and optimization. Additionally, most existing underwater three-dimensional reconstruction methods do not fully utilize the latest advances in deep learning and artificial intelligence. Although deep learning technology has achieved remarkable results in the field of computer vision, its application in underwater environments is still limited. Summary of the Invention

[0005] The objective of the present invention is to provide a three-dimensional reconstruction method for underwater environments, thereby solving some of the drawbacks and deficiencies pointed out in the background art.

[0006] The present invention adopts the following technical solutions to solve its above-mentioned technical problems: A three-dimensional reconstruction method for underwater environments. S1. Project an optimally designed spectral pattern onto the target surface, then capture the reflected light by a camera, and inversely deduce the three-dimensional structure of the object surface through the deformation of light.

[0007] Among them:

[0008] S1.1. Project structured light patterns of multiple frequency bands in the underwater environment, and simultaneously capture images using a multispectral camera.

[0009] S1.2. Through a spectral matching algorithm, fuse images of different bands, extract depth information, and generate basic data for three-dimensional reconstruction.

[0010] S1.3. Use depth compensation technology to correct depth errors caused by underwater scattering and absorption.

[0011] S2. Dynamically adjust the image correction strategy by real-time monitoring of underwater environmental parameters.

[0012] Among them:

[0013] S2.1. Use underwater sensors including light intensity, turbidity, and temperature to real-time monitor environmental parameters and generate an environmental feature model.

[0014] S2.2. According to the environmental feature model, adaptively adjust the processing parameters of image acquisition and including exposure, contrast, and color balance.

[0015] S2.3. Through an intelligent correction algorithm, detect the distorted areas in the image caused by environmental changes, and perform local correction and overall optimization.

[0016] S3. By constructing a multi-level image pyramid model, first capture the overall structure at a large scale, and then gradually enhance the texture details at a small scale.

[0017] Among them:

[0018] S3.1. Generate a multi-scale image pyramid to decompose the structural information of the three-dimensional image layer by layer from coarse to fine;

[0019] S3.2. Use a generative adversarial network to enhance details and restore textures on each layer;

[0020] S3.3. Through an adaptive detail enhancement technique, refine the key areas to generate an underwater three-dimensional image.

[0021] Furthermore, the construction process of the spectral matching algorithm includes:

[0022] S1. Utilize multi-spectral imaging technology to integrate various spectral information such as visible light, infrared, and ultraviolet into a multi-dimensional spectral feature matrix, and calculate the depth information of each pixel point through the spectral response differences of each band; its calculation formula is expressed as:

[0023]

[0024] Among them, D(x, y) represents the depth value at the pixel point (x, y), λ is the spectral wavelength, S i (λ) is the spectral response of the i-th band, α i and β i (λ) are the coefficients and absorption characteristics used to weigh the contributions of different bands;

[0025] S2. According to the real-time changes in the underwater environment, adjust the parameters of the spectral matching algorithm; by monitoring environmental parameters such as light and turbidity, dynamically update the matching function to ensure the accuracy of the image; its correction process is achieved through an adaptive adjustment function:

[0026]

[0027] Among them, Φ(t) is the dynamically adjusted spectral matching function, γ(t) represents the influence of environmental parameters on the correction function, κ is the adjustment frequency factor, and Γ(t) is the composite function of environmental light intensity and turbidity;

[0028] S3. Utilize a non-linear fusion model to comprehensively form a multi-dimensional feature space from the spectral information of different bands, and perform multi-level feature mapping and depth information extraction through a deep learning network; the process of non-linear fusion and depth extraction is expressed by the following formula:

[0029]

[0030] Among them, Z(x, y) represents the non-linearly fused depth information at the pixel point (x, y), Λ is the spectral range, δ j and ω j are the weight coefficients and frequency factors of different bands, T jThe non - linear spectral response function of the j - th band is \(η(λ)\).

[0031] Furthermore, the depth compensation technology adoption scheme includes:

[0032] S1. By establishing an underwater optical transmission model, analyzing the scattering path and absorption coefficient during the propagation of light underwater, dynamically calculating the non - linear effects of scattering and absorption on depth measurement; and simulating through the path - integral formula, which is expressed as follows:

[0033]

[0034] Among them, \(I(x,y,z)\) represents the light intensity at the point \((x,y,z)\), describing the energy of light propagation in the underwater environment; \(\lambda\) is the wavelength of light, and \(\alpha(\lambda)\) is the absorption coefficient, representing the attenuation degree of light at a specific wavelength;

[0035] The spectral response function \(\eta(\lambda)\) of the scattering path describes the scattering behavior of light with different wavelengths when passing through the water body, \(\theta(s)\) represents the angular change of the scattering path, and the integral path \(L\) represents the path length of light propagation in the water body;

[0036] S2. According to the real - time monitored underwater environmental parameters, dynamically adjust the parameters of the correction model; through a multi - level correction mechanism, adopt linear correction of the light scattering coefficient in the shallow layer area, and use depth compensation based on the non - linear scattering path in the deep layer and high - turbidity areas; which is achieved through the following formula:

[0037]

[0038] Among them, \(D(z)\) represents the corrected depth value at depth \(z\), which is dynamically calculated through a multi - level correction mechanism; in the formula \(D(z)\) represents the linear correction in the shallow layer, where \(\gamma\) n is the correction coefficient, and \(\beta\) n is the attenuation factor related to depth; and introduces the non - linear correction term \(\delta\) n (z) for processing non - linear depth compensation in the water body environment;

[0039] S3. Through a parallel processing architecture, perform scattering compensation and depth reconstruction in real - time, and dynamically adjust the compensation parameters using a depth error feedback mechanism; and the synchronization process of compensation and reconstruction is achieved through the following formula:

[0040]

[0041] Among them, E(t) represents the depth error after real-time compensation, reflecting the error correction effect during the entire compensation and reconstruction process; ξ(t) in the formula E(t) is the scattering compensation coefficient, used to adjust the compensation intensity in real time, ω is the synchronization frequency factor, determining the coordination frequency between compensation and reconstruction; and the reconstruction feedback adjustment term ζ(t) further optimizes the compensation effect by performing differential operations on the real-time feedback during the reconstruction process.

[0042] Furthermore, the process of generating the environmental feature model includes:

[0043] S1. First, in the underwater environment, multi-dimensional parameters such as light intensity, turbidity, and temperature are monitored in real time. Using dynamic environmental feature extraction technology, multi-source data are subjected to non-linear analysis and fusion; based on the real-time data, an initial model of environmental features is constructed using an adaptive algorithm; through formula-based dynamic adjustment, it adapts to environmental changes during the 3D reconstruction process:

[0044]

[0045] Among them, M(t) represents the dynamic response parameter of the environmental feature model at time t, α(t) represents the weight coefficients of different environmental parameters including light intensity, turbidity, and temperature, which are dynamically adjusted according to the real-time monitored data, ω is the frequency factor of the model response, reflecting the periodicity of environmental changes, and β(t) is the characteristic function describing the non-linear changes of environmental parameters;

[0046] S2. Then, local models are respectively established for light intensity, turbidity, and temperature to capture the features of different depths and regions; the local models are integrated into a global environmental feature model and dynamically adjusted through a multi-level optimization algorithm; the optimization process is as follows:

[0047]

[0048] Among them, G(x, y, z) is the global environmental feature model at the spatial position (x, y, z), γ n is the local environmental feature parameter, representing the initial value of the environmental feature under a specific region or depth, δ n is the depth-related environmental impact factor, describing the changing trend of environmental parameters with depth, ∈ n (z) is the non-linear optimization term, used to further correct the accuracy of the model;

[0049] S3. Finally, through the intelligent sensor data fusion mechanism, different sensor data are deeply fused, and future environmental changes are predicted through time series analysis; the core formula for data fusion and prediction is:

[0050]

[0051] Among them, P(t) represents the environmental change prediction parameter at time t, λ(t) is the fusion weight of different sensor data, which determines the influence of each sensor on the prediction result, κ is the frequency factor of the time series, reflecting the periodicity and trend of environmental changes, and θ(t) is the prediction correction term used to correct the prediction result.

[0052] Furthermore, the construction method steps adopted by the intelligent correction algorithm include:

[0053] S1. Through the multi-dimensional matching of environmental parameters and image features, identify the local distortion areas caused by environmental changes; and based on the intelligent distortion detection model, distinguish and locate the distortion areas;

[0054] S2. Use the non-linear optimization method to gradually correct the distortion areas; at the same time, combine the multi-scale optimization technology to perform hierarchical processing on the entire image;

[0055] S3. Use the convolutional neural network CNN to extract and analyze the features of the distortion areas, and then realize the high-quality reconstruction of the image through the generative adversarial network GAN.

[0056] Furthermore, the construction process of the identification method for the local distortion areas includes:

[0057] S1. By comprehensively matching the environmental parameters of light, turbidity, and temperature with the image features of brightness, texture, and edges, form a high-dimensional feature space to identify the local distortion areas caused by environmental changes; by analyzing the non-linear relationship between environmental parameters and image features, generate a multi-dimensional matching matrix, and the mathematical representation is as follows:

[0058]

[0059] Among them, F(x, y, t) represents the matching function value at the spatial position (x, y) and time t, representing the comprehensive influence between environmental parameters and image features; the parameter α i (t) is the weight of the environmental parameters including light, turbidity, and temperature, controlling the contribution degree of each environmental factor to image distortion, and g i (x, y, t) is the non-linear response function of image features, used to capture the influence of environmental changes on image quality; the correction term β i (t) is used to adjust the model to improve the accuracy of distortion identification;

[0060] S2. Based on the deep learning algorithm, the intelligent distortion detection model learns different types of distortion features through the training of environmental data and image samples, and establishes a multi-level distortion identification system; the adaptive learning process is represented by the following formula:

[0061]

[0062] Among them, S(t) represents the distortion detection result at time t, reflecting the comprehensive judgment of the current environment and image features; the parameter γ j (t) is the weight factor of the distortion feature, used to measure the importance of various distortion features in detection, h j (t) is the corresponding feature response function, describing the dynamic behavior of the distortion feature; the correction term δ j (t) is then used to adjust the sensitivity and adaptability of the model;

[0063] S3. Finally, through the collaborative work of multi-dimensional matching and the intelligent detection model, the synchronous analysis and processing of environmental parameters and image features are achieved; the collaborative detection mechanism combines the joint evaluation of the matching matrix and the intelligent model to deeply analyze and locate the local distortion area; the formula for collaborative detection is as follows:

[0064]

[0065] Among them, C(t) represents the collaborative detection result at time t, which is a comprehensive evaluation of the coupling relationship between environmental parameters and image features; the parameter λ(t) is the weight of collaborative detection, measuring the interaction between the environment and image features in distortion detection, ω is the frequency factor of the time series, used to capture the periodic changes of the environment and image features, and the correction term θt is used to dynamically adjust the detection result.

[0066] Furthermore, the gradual correction of the distortion area is carried out by combining the dynamic feedback of environmental parameters and image features, and multi-round iterative correction is performed on the distortion area; the expression of the optimization process is as follows:

[0067]

[0068] Among them, L(t) represents the non-linear loss function at time t, α i is the weight factor of environmental parameters and image features, Φ i (x, y, t) is the non-linear response function of a specific distortion area, describing the impact of environmental changes on the image, and β(t) is the correction term for optimization;

[0069] The hierarchical processing of the image is to divide the image into multiple scale levels and perform optimization processing at different scales; the expression of the multi-scale optimization process is as follows:

[0070]

[0071] Among them, G(x, y, z) represents the global optimization function at the spatial position (x, y, z), γ j is the optimization parameter at different scales, δ jis a scale-related attenuation factor that describes the optimization effects at different levels, ∈ j Correction term (z) is used for optimizing to balance the global structure and local details.

[0072] Furthermore, for the feature extraction and analysis, a convolutional neural network (CNN) is introduced. Through multi-layer convolutional and pooling operations, high-level features in the image are gradually extracted, including non-linear distortion features caused by environmental changes. The expression for feature extraction is as follows:

[0073]

[0074] where F(x, y, t) represents the feature map value at spatial position (x, y) and time t, and α i are the weight coefficients of different convolutional layers; the function σ is the activation function, which captures the feature relationships by introducing non-linear mapping; the product of the convolutional kernel weight W i and the input feature g(x, y, t) represents the weighted sum of features in the convolutional operation, and b i is the bias term used to adjust the distribution of features;

[0075] The generator of the generative adversarial network (GAN) uses the features extracted by the CNN to generate high-quality images, while the discriminator evaluates the authenticity of the generated images. The expression for the loss function is as follows:

[0076]

[0077] where L(G, D) represents the adversarial loss function between the generator G and the discriminator D; and is the expectation of the real image x. The discriminator D is optimized by judging whether the input image comes from real data or the generator; and evaluates the image output by the generator. The generator tries to deceive the discriminator by generating images.

[0078] Advantages of the present invention:

[0079] By introducing the dynamic monitoring and adjustment of multi-dimensional environmental parameters (such as light, turbidity, temperature), and combining deep learning technologies such as convolutional neural network (CNN) and generative adversarial network (GAN), this method can effectively address the problems of light scattering, absorption, and other non-linear distortions in complex underwater environments, and significantly improve the accuracy of 3D reconstruction.

[0080] Through multi-level optimization and dynamic feedback mechanisms, in different underwater environments, especially in cases of high turbidity and high light fluctuations, it can still maintain high image quality and reconstruction stability. This makes the method highly adaptable in various underwater operations and capable of coping with changing environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is the flowchart of the three-dimensional reconstruction method for the underwater environment of the present invention.

[0082] Figure 2 This is the flowchart of the construction process of the spectral matching algorithm of the present invention.

[0083] Figure 3 This is the flowchart of the adopted solution of the depth compensation technology of the present invention. Detailed implementation manner

[0084] The following will make a detailed description of the specific implementation manner of the present invention with reference to the accompanying drawings.

[0085] Combined with the attached Figure 1 , for the three-dimensional reconstruction method of the underwater environment of the present invention, in the first step, an optimized spectral pattern is projected onto the target surface, and then the reflected light is captured by a camera. The three-dimensional structure of the object surface is deduced by the deformation of light. First, multi-band structured light patterns are projected in the underwater environment, and at the same time, a multi-spectral camera is used to capture images. This process is to project specifically designed spectral patterns in the underwater environment. These patterns cover multiple spectral bands to ensure sufficient information is captured under different optical conditions. The multi-spectral camera can capture the light reflected from different bands, providing multi-dimensional data support for subsequent image fusion and depth extraction. Then, different-band images are fused through a spectral matching algorithm to extract depth information and generate the basic data for three-dimensional reconstruction. In this step, the system fuses the images of each band captured by the multi-spectral camera, uses the spectral matching algorithm to compare and match the features of these images, and calculates the depth information of the object based on the reflection and absorption characteristics of each band for light. The spectral matching algorithm uses the spectral response differences under different bands to accurately calculate the depth value of each pixel point and form preliminary three-dimensional point cloud data, providing a basis for three-dimensional reconstruction. Finally, a depth compensation technology is used to correct the depth error caused by underwater scattering and absorption. In the underwater environment, light is scattered and absorbed by the water body during propagation, resulting in certain errors in the depth information in the captured images. To solve this problem, the depth compensation technology dynamically corrects these errors by constructing a refined optical transmission model, analyzing the propagation path of light underwater, and combining real-time monitored environmental parameters such as the turbidity of the water body and the attenuation characteristics of light. Through depth compensation, the system can eliminate or significantly reduce the depth measurement errors caused by light scattering and absorption, and finally generate accurate three-dimensional reconstruction data.

[0086] In the second step, the image correction strategy is dynamically adjusted by real-time monitoring of underwater environmental parameters. First, underwater sensors, including devices for light, turbidity, temperature, etc., are used to monitor environmental parameters in real time, and an environmental feature model is generated based on these data. This model reflects the complex characteristics of the current underwater environment by integrating multiple environmental parameters, providing accurate environmental background data for subsequent image processing. Then, according to the generated environmental feature model, the image acquisition and processing parameters, including exposure, contrast, and color balance, are adaptively adjusted. This adjustment process is dynamic. The system analyzes the environmental feature model to identify the impact of the current environment on image acquisition, and then adjusts the parameters of the imaging device in real time to adapt to the changing underwater conditions, ensuring the quality and consistency of the images. Finally, through an intelligent correction algorithm, the distorted areas in the image caused by environmental changes are detected, and local correction and overall optimization are performed. The intelligent correction algorithm uses multi-dimensional matching of the environmental feature model and image features to accurately identify the distorted areas in the image caused by environmental factors such as light changes, turbidity fluctuations, or temperature changes. For these distorted areas, the system does not simply perform a global adjustment, but applies a local correction strategy to gradually correct the distorted parts of the image. At the same time, to maintain the overall consistency and visual effect of the image, the system also performs overall optimization to ensure that the corrected image maintains a high degree of coherence and accuracy both in local details and global structure.

[0087] In the third step, by constructing a multi-level image pyramid model, the overall structure is captured at a large scale first, and then the texture details are gradually enhanced at a small scale. First, a multi-scale image pyramid is generated, and the structural information of the three-dimensional image is decomposed layer by layer from coarse to fine. By decomposing the original image into multi-level structures with different resolutions, each layer represents a different scale, from the coarsest overall structure to the most delicate texture details, forming a multi-level image pyramid model. This process helps to balance the global structure and local details of the image during the reconstruction process, ensuring the integrity and fineness of the reconstruction. Then, a generative adversarial network (GAN) is used to enhance the details and restore the texture on each layer. The GAN consists of a generator and a discriminator. Through adversarial training, the generator enhances the details and restores the texture of the image layer by layer, while the discriminator evaluates the output effect of the generator, thus prompting the generator to continuously improve its generation ability. This process is repeated at each level, making the image structure refined step by step from coarse to fine, and the texture information is fully restored, ensuring that even in a complex underwater environment, the details and texture of the image can still be clearly presented. Finally, through an adaptive detail enhancement technique, the key areas are refined to generate an underwater three-dimensional image. The adaptive detail enhancement technique can dynamically adjust the enhancement strategy according to the characteristics of the current image and focus on processing the key areas in the image. These key areas often contain important structural or texture information and are the focus of the image detail performance. Through the adaptive enhancement strategy, the system can precisely control the intensity and range of enhancement, making the details of the key areas clearer. The enhanced image not only maintains coherence as a whole but also has extremely high quality in terms of details. Finally, by combining the multi-scale structure of the pyramid model, the layer-by-layer refinement of the GAN, and the fine processing of the adaptive detail enhancement, a high-quality underwater three-dimensional image is generated, achieving high-precision three-dimensional reconstruction in a complex underwater environment.

[0088] Example 1:

[0089] If Figure 2 As shown, in this embodiment, a three-dimensional reconstruction is performed on an underwater structure to restore the three-dimensional structure. In this task, due to the complex underwater environment, the scattering and absorption of light make it difficult for traditional single-band imaging techniques to obtain clear depth information. Therefore, a multi-spectral imaging technique is adopted, which is achieved by integrating various spectral information such as visible light, infrared light, and ultraviolet light.

[0090] First, an image data of different bands is obtained by using a multi-spectral camera system. The multi-spectral camera can capture the reflected light of visible light (400 - 700nm), near-infrared (700 - 1000nm), and ultraviolet light (200 - 400nm), and these data are recorded as spectral responses S of different bands by the sensor. i(λ). In this example, it is set that the multispectral camera captures spectral information in n = 5 bands, where visible light, infrared light, and ultraviolet light respectively correspond to five bands with wavelengths of 450 nm, 550 nm, 700 nm, 800 nm, and 300 nm.

[0091] The spectral matching algorithm used calculates the depth value of each pixel point based on the following formula:

[0092]

[0093] In the formula, D(x, y) represents the depth value of the pixel point (x, y), λ is the spectral wavelength, and S i (λ) is the spectral response of the i-th band. It is set that the wavelength range is from 200 nm to 1000 nm, covering the ultraviolet, visible light, and near-infrared ranges. The coefficient α i is used to weigh the contribution degree of each band to the depth information, and its value range can be set between 0.1 and 1, and is adjusted according to the importance of different bands; for example, in this example, the coefficients α of visible light and infrared light i are respectively set to 0.7 and 0.8, and the ultraviolet light is set to 0.5 because the attenuation of ultraviolet light in water is greater and the influence is smaller. The absorption characteristic β i (λ) is used to describe the attenuation characteristic of light underwater, and its value range is set to 0.01 to 0.1 to reflect the attenuation degree of different wavelengths in water. For example, β i (λ) can be set to 0.05 for infrared light and 0.1 for ultraviolet light, indicating that the attenuation of ultraviolet light is more severe.

[0094] Substitute these parameters and calculate for a certain pixel point (x = 100, y = 150). Set the spectral response S 1 (450 nm) = 0.8, S 2 (550 nm) = 0.9, S 3 (700 nm) = 0.85, S 4 (800 nm) = 0.7, S 5 (300 nm) = 0.6. Calculate the depth value D(100, 150):

[0095]

[0096] Through numerical integration, the depth value of this pixel point can be obtained as approximately D(100, 150) ≈ 2.5 meters. This indicates that under the fusion of multi-band spectral data, the three-dimensional shape of the underwater structure can be accurately deduced. Such calculations are performed on the entire image to generate complete three-dimensional point cloud data, thus achieving high-precision three-dimensional reconstruction of the underwater environment.

[0097] Then, to ensure the accuracy of the images, the parameters of the spectral matching algorithm can be adjusted according to the real-time changes in the underwater environment. By real-time monitoring of environmental parameters such as underwater light and turbidity, the spectral matching function is dynamically updated to ensure clear images and accurate depth data in a changing environment. Specifically, this dynamic correction is achieved through the adaptive adjustment function Φ(t), and its formula is:

[0098]

[0099] In the formula, Φ(t) represents the dynamically adjusted spectral matching function, which synthesizes the impact of real-time changes in environmental parameters on the images. The parameter γ(t) is used to represent the influence of environmental parameters on the correction function. These environmental parameters include light intensity and water turbidity, etc., and the value range is generally between 0.5 and 1.5 to reflect the variation of the influence of the environment on light propagation; while κ is the adjustment frequency factor, and the value range is usually set from 0.1 to 1.0 to control the response speed of the spectral matching function to environmental changes; Γ(t) is the composite function of light intensity and turbidity, which comprehensively evaluates the changes in environmental light and water transparency, and the value range is set from 100 to 1000 to cover a wide range of scenarios from low light, low transparency to high light, high transparency.

[0100] Through real-time measurement by underwater sensors, it is found that the current environmental parameters are as follows: the light intensity is 800, and the turbidity index is 0.8. In this case, set γ(t) = 1.2, κ = 0.5, and Γ(t) = 800+(0.8×100)=880. Substitute these parameters into the formula for calculation:

[0101]

[0102] Set the time interval T = 10 seconds and consider (because the changes in environmental parameters can be approximated as constants in a short period of time), then:

[0103] Φ(t)=∫ 0 10 [1.2·cos(0.5·t)]dt

[0104] The calculation result is:

[0105]

[0106] The adjusted spectral matching function Φ(t) ≈ 2.3, and this value will be used to update the parameters in the spectral matching algorithm, especially the weight coefficients and attenuation characteristics in the depth calculation formula, so as to adjust the fusion effect of multi-band data in real time. Through this dynamic adjustment process, the system can still maintain efficient correction of the distorted area and accurate depth reconstruction in the complex and changeable underwater environment, ensuring a high degree of consistency in the overall and details of the reconstructed three-dimensional image.

[0107] Finally, to further improve the accuracy of the image, a non-linear fusion model is used to comprehensively form a multi-dimensional feature space from the spectral information of different bands, and multi-level feature mapping and depth information extraction are carried out through a deep learning network. Using the visible light, infrared and ultraviolet spectral data collected by multi-spectral imaging technology, these spectral data are integrated into a high-dimensional feature space. In this process, the formulas for non-linear fusion and depth extraction are used:

[0108]

[0109] Among them, Z(x,y) represents the depth information after non-linear fusion at the pixel point (x,y), Λ is the spectral range, and δ j is the weight coefficient of different bands, reflecting the contribution degree of each band in depth information extraction, and the value range is set from 0.1 to 1.0; ω j is the frequency factor, controlling the frequency characteristics of the spectral response, and the value range is set from 0.05 to 0.5; while T j (λ) is the non-linear spectral response function of the j-th band, describing the spectral response characteristics of each band.

[0110] It is set to use m = 4 spectral bands, corresponding to the visible light, near-infrared and ultraviolet ranges respectively, and the wavelength range Λ is set from 300nm to 900nm to cover the spectral data from ultraviolet to near-infrared. Specifically, the weight coefficients δ 1 = 0.8, δ 2 = 0.7, δ 3 = 0.6, δ 4 = 0.5, and these values are determined according to the underwater penetration and image contribution degree of each band, reflecting the influence size of different bands on depth information. The frequency factor ω j is set to ω 1 = 0.1, ω 2 = 0.2, ω 3 = 0.3, ω 4 = 0.4, and these values are used to adjust the response frequency of each band in the fusion process to adapt to the performance of different spectral characteristics in the underwater environment.

[0111] Substitute these parameters into the formula to calculate the depth information for a certain pixel point (x = 50, y = 100):

[0112]

[0113] Set the spectral response functions of each band to have responses of T at wavelengths of 300 nm, 500 nm, 700 nm, and 900 nm respectively 1 (300nm) = 0.9, T 2 (500nm) = 0.85, T 3 (700nm) = 0.8, T 4 (900nm) = 0.75. Substitute these data into the formula and calculate term by term to obtain:

[0114] Z(50, 100) = (0.8·0.9·sin(0.1·300)) + (0.7·0.85·sin(0.2·500)) + (0.6·0.8·sin(0.3·700)) + (0.5·0.75·sin(0.4·900))

[0115] = (0.8·0.9·sin(30)) + (0.7·0.85·sin(100)) + (0.6·0.8·sin(210)) + (0.5·0.75·sin(360))

[0116] ≈ (0.72·0.5) + (0.595·0.984) + (0.48· -0.5) + (0.375·0)

[0117] = 0.36 + 0.58548 - 0.24 + 0 = 0.70548

[0118] Through this calculation process, the depth information after non - linear fusion of the pixel point (50, 100) is obtained as 0.70548 meters. This non - linear fusion and depth extraction method significantly improves the depth accuracy of underwater three - dimensional reconstruction by performing multi - dimensional processing on spectral data of different bands.

[0119] Example 2:

[0120] Refer to the appendix Figure 3 , in this example, during the three - dimensional reconstruction of underwater structures, in order to improve the accuracy of depth measurement, a depth compensation technique is adopted. By establishing an underwater optical transmission model, the scattering path and absorption coefficient of light during underwater propagation are analyzed, and the non - linear effects of scattering and absorption on depth measurement are calculated dynamically. This process uses the path integral formula to simulate the propagation characteristics of light, and its formula is:

[0121] I(x, y, z) = ∫ 0 L [η(λ)·e -α(λ)·s·cos(θ(s))]ds

[0122] In this formula, I(x, y, z) represents the light intensity at the point (x, y, z), reflecting the energy transmission of light in the underwater environment; λ is the wavelength of light, and α(λ) is the absorption coefficient, indicating the attenuation degree of light at a specific wavelength, usually with a value range between 0.01 and 0.1 to reflect the attenuation characteristics of light with different wavelengths in water. For example, the absorption coefficient of blue light can be set to 0.02, while for red light, due to its faster attenuation in water, the absorption coefficient can be set to 0.08; the spectral response function η(λ) of the scattering path describes the performance of light with different wavelengths after scattering in water, and its value range is usually set between 0.5 and 1.5 to reflect different scattering degrees of light underwater. Set the response of blue light to 1.2 and that of red light to 0.7; θ(s) represents the angular change of the scattering path, and its value can be set between 0 and 90 degrees according to the flow of the water body and the environment, while the integration path L represents the propagation path length of light in the water body, usually with a value range of 1 to 10 meters, specifically determined according to the depth and clarity of the underwater environment.

[0123] Suppose a pixel point (x = 100, y = 150, z = 3) is located on the surface of the target underwater structure. The wavelength λ of the light is 500 nm (blue light), the absorption coefficient α(500 nm) = 0.02, the scattering response η(500 nm) = 1.2, and the propagation path of the light L = 5 meters and the scattering angle change θ(s) = 30 degrees are set. Substitute these parameters into the formula for calculation:

[0124] I(100, 150, 3) = ∫ 0 5 [1.2·e -0.02·s ·cos(30°)]ds

[0125] Simplify the calculation:

[0126]

[0127] From this, the light intensity at this point is approximately 4.946, reflecting the attenuation degree of the light after propagating and scattering for 5 meters underwater. This light intensity value will be used for subsequent calculations of depth compensation. By adjusting the absorption coefficient and response function, the measurement accuracy can be further optimized.

[0128] To address the impact of different depths and water turbidity on light propagation, the parameters of the calibration model are dynamically adjusted based on real-time monitored underwater environmental parameters, and a multi-level calibration mechanism is adopted to optimize depth measurement. In the shallow layer, due to the low turbidity of the water body, the scattering and absorption of light are small, and linear calibration of the light scattering coefficient is used. In the deep or high-turbidity area, the non-linear scattering and absorption of light are more significant, so depth compensation based on the non-linear scattering path is required to correct the depth measurement error. This process is achieved through the following formula:

[0129]

[0130] In this formula, D(z) represents the calibrated depth value at depth z, which is dynamically calculated through a multi-level calibration mechanism. The part represents the linear calibration in the shallow layer, where γ n is the calibration coefficient, reflecting the calibration intensity, usually with a value range of 0.5 to 1.5; β n is the attenuation factor related to depth, used to describe the degree of light attenuation in the shallow layer, with a value range of 0.01 to 0.1. The non-linear calibration term is used to handle the non-linear depth compensation in a complex water environment, where δ n (z) represents the non-linear calibration function, reflecting the change in the light scattering path under deep or high-turbidity conditions, usually set as a function that is sensitive to depth changes and can capture non-linear characteristics.

[0131] Suppose the current water environment monitoring shows that the light scattering coefficient in the shallow layer (within 1 meter) is low, and γ 1 is set to 1.0, β 1 is set to 0.02; while in the deep layer (at 5 meters), due to the increase in water turbidity and significant non-linear effects, γ 2 is set to 1.2, β 2 is set to 0.05, and the non-linear calibration term δ 2 (z) = 0.1z 2 . It is necessary to calculate the calibrated depth value D(5) at depth z = 5 meters.

[0132] Substitute these parameters into the formula:

[0133]

[0134] Expand the calculation:

[0135] For shallow layer calibration (n = 1):

[0136]

[0137] For deep layer calibration (n = 2), including the non-linear compensation term:

[0138]

[0139] Add the results of the two parts:

[0140] D(5) = 4.76 + 5.3088 = 10.0688

[0141] Therefore, the corrected depth value at a depth of 5 meters is 10.0688 meters. This dynamic calculation through a multi-level correction mechanism can accurately compensate for the non-linear effects of different depths and water turbidity on light propagation.

[0142] To ensure the real-time performance and accuracy of depth compensation and reconstruction, a parallel processing architecture is adopted to perform scattering compensation and depth reconstruction in real time. Through a depth error feedback mechanism, the system can dynamically adjust the compensation parameters to ensure a high degree of coordination in the compensation and reconstruction processes. This process of synchronous compensation and reconstruction is achieved through the following formula:

[0143]

[0144] where E(t) represents the depth error after real-time compensation, reflecting the error correction effect during the entire compensation and reconstruction process. The ξ(t) in the formula is the scattering compensation coefficient, used to adjust the compensation intensity in real time, and its value range is generally from 0.5 to 1.5 to meet the scattering compensation requirements in different environments; ω is the synchronous frequency factor, which determines the coordination frequency between compensation and reconstruction, and its value range is set from 0.1 to 1.0 Hz, is the reconstruction feedback adjustment term, and ζ(t) reflects the error feedback during the reconstruction process. By performing a differential operation on the real-time feedback during the reconstruction process, the compensation effect is further optimized.

[0145] Suppose that the underwater light and turbidity monitored in the current environment show certain fluctuations, resulting in the need for real-time adjustment of scattering compensation. Set the scattering compensation coefficient ξ(t) = 1.2, the frequency factor ω = 0.5, and the reconstruction feedback function ζ(t) = 0.1t 2 , where the form of ζ(t) is selected to capture the gradually accumulating errors during the compensation process and dynamically adjust the compensation intensity through its derivative.

[0146] Substitute these parameters into the formula to calculate the depth error compensation effect at time T = 10 seconds:

[0147]

[0148] Expand the calculation:

[0149]

[0150] Therefore, the depth error after real-time compensation within 10 seconds is 13.08088 meters. This real-time compensation and reconstruction synchronization mechanism can continuously optimize the compensation parameters through a parallel processing architecture and error feedback adjustment, ensuring the accuracy of depth reconstruction when the environment changes.

[0151] Example 3:

[0152] In this embodiment, when continuing with the 3D reconstruction of underwater structures, in order to accurately capture the changes in the underwater environment and dynamically adjust the 3D reconstruction parameters, a scheme for generating an environmental feature model is used. First, in the underwater environment, multi-dimensional parameters such as light intensity, turbidity, and temperature are monitored in real time through sensors. Using dynamic environmental feature extraction technology, these multi-source data are subjected to non-linear analysis and fusion to form a preliminary environmental feature model. To adapt to the dynamic changes in the environment during the 3D reconstruction process, the system uses an adaptive algorithm to adjust the model based on real-time data, and dynamically calculates the response parameters of the environmental feature model through the following formula:

[0153]

[0154] Where M(t) represents the dynamic response parameter of the environmental feature model at time t, reflecting the impact of the environment on the 3D reconstruction process. The parameter α(t) represents the weight coefficients of different environmental parameters including light intensity, turbidity, and temperature, and these weights are dynamically adjusted according to the real-time monitored data, usually with a value range of 0.5 to 1.5 to reflect the importance of each environmental parameter under different conditions; ω is the frequency factor of the model response, usually set to 0.1 to 1.0 Hz, representing the periodic frequency of environmental changes; β(t) is a characteristic function describing the non-linear changes of environmental parameters, and its value can dynamically reflect the change trend of the underwater environment.

[0155] Set in a specific area of the research site, the current light intensity measured by the sensor is 800 Lux, the turbidity is 0.7 NTU, and the water temperature is 15 degrees Celsius. According to these data, set the weight coefficient α of light 光照 = 1.2, the weight coefficient α of turbidity 浑浊度 = 1.0, the weight coefficient α of temperature 温度 = 0.8. Set ω = 0.5 Hz, β(t) = 0.05t 2 , where the form of β(t) is used to capture the non-linear characteristics of environmental changes, such as the fluctuations of light intensity over time and the gradual changes of turbidity.

[0156] Substitute these data into the formula to calculate the response of the environmental feature model at time T = 10 seconds:

[0157]

[0158] Expand the calculation:

[0159]

[0160] Therefore, the response of the environmental feature model within 10 seconds is 7.5674. This result reflects the dynamic adjustment effect of the combined influence of light, turbidity, and temperature on the 3D reconstruction process under the current environmental conditions.

[0161] Further improve the construction of the environmental feature model by locally modeling light, turbidity, and temperature respectively to capture the features at different depths and regions, and integrating these local models into a global environmental feature model. For this purpose, a multi-level optimization algorithm is used for dynamic adjustment to ensure that the environmental feature model can reflect the current underwater environmental changes in real time. This optimization process is achieved through the following formula:

[0162]

[0163] where G(x, y, z) is the global environmental feature model at the spatial position (x, y, z), and γ n is the local environmental feature parameter, representing the initial value of the environmental feature in a specific region or depth, usually with a value range of 0.5 to 1.5 to reflect the local feature influence under different environmental conditions; δ n is the depth-related environmental influence factor, describing the variation trend of environmental parameters with depth, with a value range of 0.01 to 0.1 to reflect the attenuation or enhancement of environmental parameters at different depths; ∈ n (z) is the non-linear optimization term, used to further correct the accuracy of the model, describing the non-linear variation characteristics of environmental parameters.

[0164] Monitor and model the light, turbidity, and temperature at different depths and regions at the underwater structure research site. Set at a depth Z = 3 meters, the measured light intensity, turbidity, and temperature are 700 Lux, 0.9 NTU, and 14 degrees Celsius respectively. According to these data, set the local feature parameter γ 光照 of light to be 1.1, the local feature parameter γ 浑浊度 of turbidity to be 0.9, and the local feature parameter γ 温度 of temperature to be 1.3. The depth-related influence factors are respectively set as δ 光照 = 0.02, δ 浑浊度 = 0.05, and δ 温度 = 0.03, and the non-linear optimization term ∈ n (z) = 0.1z 2 , used to capture the non-linear variation trends of each environmental parameter.

[0165] Substitute these data into the formula to calculate the response of the global environmental feature model at the point (x = 100, y = 150, z = 3):

[0166]

[0167] Expand the calculation for illumination:

[0168]

[0169] For turbidity:

[0170]

[0171] For temperature:

[0172]

[0173] Add the results of each part:

[0174] G(100, 150, 3) = 3.219 + 2.518 + 4.631 = 10.368

[0175] The response of the global environmental feature model at a depth of 3 meters is 10.368. This indicates that by local modeling and integration into a global model, and applying a multi-level optimization algorithm, the impact of environmental features at different depths and regions can be dynamically captured and adjusted.

[0176] To improve the system's ability to respond to environmental changes, an intelligent sensor data fusion mechanism is adopted. The illumination, turbidity, and temperature data from different sensors are deeply fused, and future environmental changes are predicted through time series analysis. The core of this step is to provide anticipatory adjustments for the reconstruction process through real-time data fusion and a prediction model of environmental changes. The core formula for data fusion and prediction is:

[0177]

[0178] Where P(t) represents the environmental change prediction parameter at time t, reflecting the future environmental trend; λ(t) is the fusion weight of different sensor data, determining the influence of each sensor in the prediction result, with a value range of 0.5 to 1.5 to flexibly adjust the importance of each sensor; κ is the frequency factor of the time series, usually set to 0.1 to 1.0 Hz, reflecting the periodicity and trend of environmental changes; θ(t) is the prediction correction term, used to correct the prediction result by analyzing the historical trend of sensor data to ensure accurate prediction of future environmental changes.

[0179] Set the data measured in real-time by the sensors used as follows: the current illumination intensity is 800 Lux, the turbidity is 0.8 NTU, and the water temperature is 15 degrees Celsius. Set the fusion weight of illumination as λ 光照 = 1.2, and the fusion weight of turbidity as λ 浑浊度= 1.0, and the fusion weight of temperature is λ 温度 = 0.9. The frequency factor κ of the time series is 0.5, and the prediction correction term θ(t) = 0.05t 2 , which is used to capture non-linear change trends, such as gradually increasing turbidity and light fluctuations.

[0180] Substitute these data into the formula to calculate the environmental change prediction parameters at time T = 10 seconds:

[0181]

[0182] Expand the calculation:

[0183]

[0184] Therefore, the environmental change prediction parameter within 10 seconds is 3.0822. This result indicates that the system can dynamically adjust the key parameters in the 3D reconstruction process through the fusion of intelligent sensor data and accurate prediction of environmental changes, ensuring high precision and stability of the reconstruction in future environmental changes.

[0185] Example 4:

[0186] In this embodiment, during the underwater 3D reconstruction of underwater structures, accurately identifying local distortion areas is crucial for improving the reconstruction quality. The adopted method for identifying local distortion areas constructs a high-dimensional feature space by integrating environmental parameters such as light, turbidity, and temperature with the brightness, texture, and edge features of the image, thereby identifying local distortion areas caused by environmental changes. This method generates a multi-dimensional matching matrix by analyzing the non-linear relationship between environmental parameters and image features to achieve precise distortion identification. The specific mathematical representation is as follows:

[0187]

[0188] Among them, F(x, y, t) represents the matching function value at the spatial position (x, y) and time t, representing the comprehensive influence between environmental parameters and image features. The parameter α i (t) is the weight of environmental parameters such as light, turbidity, and temperature, which controls the contribution degree of each environmental factor to image distortion, and its value range is from 0.5 to 1.5 to flexibly adjust the influence of environmental factors; g i (x, y, t) is the non-linear response function of image features, which is used to capture the influence of environmental changes on image quality and reflects the changes in image brightness, texture, and edges; the correction term β i (t) is used to adjust the model to improve the accuracy of distortion identification and capture the change trend in time.

[0189] In the underwater environment of the research site, the current light intensity measured by the sensor is 750 Lux, the turbidity is 1.0 NTU, and the temperature is 16 degrees Celsius. Based on these data, the weight α of light is set 光照 = 1.1, the weight α of turbidity 浑浊度 = 1.2, and the weight α of temperature 温度 = 0.9. The image feature response function g i (x, y, t) is set as a non-linear function according to the changes in brightness, texture, and edges. For example, it is set as g 1 (x, y, t) = 0.8x, g 2 (x, y, t) = 0.7y 2 and g 3 (x, y, t) = 0.6t. The correction term β i (t) = 0.05t 2 , which is used to further improve the accuracy of the model at different time points.

[0190] Substitute these data to calculate the matching function value at time T = 10 seconds and spatial position (x = 50, y = 75):

[0191]

[0192] Expand the calculation:

[0193]

[0194]

[0195] Therefore, the matching function value at spatial position (50, 75) within 10 seconds is 47749. This result accurately reflects the comprehensive influence of light, turbidity, and temperature on image distortion through the comprehensive matching of environmental parameters and image features, thereby identifying the local distortion area.

[0196] As the identification of the local distortion area progresses step by step, the accuracy and adaptability of distortion identification are further improved through an intelligent distortion detection model based on deep learning algorithms. This model is trained on environmental data and a large number of image samples to learn different types of distortion features, thereby establishing a multi-level distortion identification system. The adaptive learning process is represented by the following formula:

[0197]

[0198] Among them, S(t) represents the distortion detection result at time t, reflecting the comprehensive judgment of the current environment and image features. The parameter γ j(t) is the weight factor of the distortion feature, which measures the importance of each type of distortion feature in detection. The value range is generally from 0.5 to 1.5 to adjust the influence of different distortion features on the model; h j (t) is the corresponding feature response function, which is used to describe the dynamic behavior of the distortion feature and reflects the influence of environmental changes on image distortion; the correction term δ j (t) is used to adjust the sensitivity and adaptability of the model to further improve the accuracy of distortion recognition.

[0199] At the research site of underwater structures, a three-dimensional reconstruction of the underwater environment is carried out. The intelligent distortion detection model learns and adjusts through real-time data monitored by sensors and historical image samples. Set the currently measured environmental parameters as: light intensity 800 Lux, turbidity 1.2 NTU, water temperature 17 degrees Celsius. Set the distortion weight γ of light 光照 = 1.0, the distortion weight γ of turbidity 浑浊度 = 1.3, the distortion weight γ of temperature 温度 = 0.8. The feature response function is set as: h 1 (t) = 0.9sin(0.2t) is used to describe the light distortion feature, h 2 (t) = 1.1cos(0.3t) is used to describe the turbidity distortion feature, h 3 (t) = 0.7sin(0.1t) is used to describe the temperature-related distortion feature. The correction term δ j (t) = 0.1t 2 is used to adjust the model to reflect the changes in non-linear distortion features in the time series.

[0200] Substitute these data to calculate the distortion detection result at time T = 10 seconds:

[0201]

[0202] Expand the calculation:

[0203]

[0204]

[0205] Therefore, the distortion detection result within 10 seconds is 15.5713. Through this adaptive deep learning detection model, the system can adjust the comprehensive judgment of the environment and image features in real time, accurately identify various types of distortion features, and improve the sensitivity and accuracy of distortion recognition through continuous learning and optimization.

[0206] To further improve the synchronous analysis and processing of environmental parameters and image features, a collaborative working mechanism of multi-dimensional matching and intelligent detection model is adopted. This collaborative detection mechanism combines the joint evaluation of the matching matrix and the intelligent model, deeply analyzes and locates the locally distorted areas, and realizes high-precision distortion identification and correction. The collaborative detection process is carried out through the following formula:

[0207]

[0208] Among them, C(t) represents the collaborative detection result at time t, which is a comprehensive evaluation of the coupling relationship between environmental parameters and image features. The parameter λ(t) is the weight of collaborative detection, which measures the interaction between the environment and image features in distortion detection. Its usual value range is from 0.5 to 1.5 to adjust the influence of different factors on the detection result; ω is the frequency factor of the time series, and its value range is from 0.1 to 1.0 Hz, which is used to capture the periodic changes of the environment and image features; the correction term θ(t) is used to dynamically adjust the detection result, reflecting the adaptability of the system to real-time environmental changes.

[0209] Continue to use the sensor to monitor the environmental data near the underwater structure. The currently measured light intensity is 850 Lux, the turbidity is 1.1 NTU, and the water temperature is 16.5 degrees Celsius. According to these data, set the collaborative detection weight λ of light 光照 =1.2, the collaborative detection weight λ of turbidity 浑浊度 =1.0, the collaborative detection weight λ of temperature 温度 =0.9. The frequency factor ω = 0.4 Hz, which reflects the periodic changes of the environment and image features. The correction term θ(t) = 0.07t 2 is used to dynamically adjust the detection result to accurately reflect the change trend over time.

[0210] Substitute these parameters into the formula to calculate the collaborative detection result at time T = 10 seconds:

[0211]

[0212] Expand the calculation:

[0213] C(10) = ∫ 0 10 [1.0·cos(0.4t) + 0.14t]dt

[0214]

[0215] Therefore, the collaborative detection result within 10 seconds is 5.108. This indicates that through the collaborative work of the multi-dimensional matching and intelligent detection model, the system can effectively synthesize the changes of environmental parameters and image features, adjust the distortion detection strategy in real time, and thus accurately identify and locate the distorted areas.

[0216] Example 5:

[0217] In this example, during the underwater three-dimensional reconstruction of underwater structures, in order to ensure the high quality and accuracy of the images, a scheme for gradually correcting the distorted areas was adopted. This scheme performs multiple rounds of iterative corrections on the distorted areas through the dynamic feedback of combining environmental parameters and image features to continuously optimize the reconstruction effect. The mathematical expression of the optimization process is:

[0218]

[0219] where L(t) represents the non-linear loss function at time t, reflecting the error magnitude in the current correction process. The parameter α i is the weight factor of the environmental parameters and image features, used to measure the influence of each factor on distortion, with a value range of 0.5 to 1.5 to dynamically adjust the weights of different factors; Φ i (x, y, t) is the non-linear response function of a specific distorted area, describing the influence of environmental changes on the image, such as the different responses of illumination, turbidity, etc. on image distortion; β(t) is the correction term for optimization, which realizes the dynamic adjustment of errors through differential operations to improve the accuracy of correction.

[0220] During the reconstruction of underwater structures, it is set that the currently monitored environmental parameters are illumination intensity of 820 Lux, turbidity of 1.3 NTU, and water temperature of 18 degrees Celsius. According to these environmental data, the weight of illumination α 光照 is set to 1.0, the weight of turbidity α 浑浊度 is set to 1.2, and the weight of temperature α 温度 is set to 0.8. For the non-linear response function of a specific distorted area, it is set as Φ 1 (x, y, t) = 0.9sin(0.3t), Φ 2 (x, y, t) = 1.1cos(0.2t), Φ 3 (x, y, t) = 0.7sin(0.1t), respectively used to describe the influence of illumination, turbidity, and temperature on the image. The correction term β(t) = 0.1t 2 , used to optimize the process of distortion correction.

[0221] Substitute these parameters into the formula to calculate the non-linear loss function at time T = 10 seconds:

[0222]

[0223] Expand the calculation:

[0224]

[0225] Therefore, the non - linear loss function within 10 seconds is 21.9954. This result indicates that the system can effectively reduce image distortion and improve the accuracy of 3D reconstruction by gradually correcting the distorted areas and combining the dynamic feedback of environmental parameters and image features.

[0226] To optimize the reconstruction quality and accuracy of the image, a scheme of hierarchical processing of the image is adopted. This scheme divides the image into multiple scale levels and performs optimization processing separately at different scales to balance the global structure and local details and ensure high - precision reconstruction in complex environments. The mathematical expression of the multi - scale optimization process is:

[0227]

[0228] where \(G(x,y,z)\) represents the global optimization function at the spatial position \((x,y,z)\), reflecting the optimization effects at different levels. The parameter \(\gamma\) j is the optimization parameter at different scales, usually with a value range of 0.5 to 1.5 to adjust the optimization intensity of each scale; \(\delta\) j is the scale - related attenuation factor, describing the change of optimization effects at different levels with depth, with a value range of 0.01 to 0.1, indicating the optimization attenuation from global to local; \(\epsilon(z)\) j is the optimization correction term, used to further balance the global structure and local details and achieve dynamic adjustment of each level through differential operations.

[0229] In the 3D reconstruction of underwater structures, multi - scale hierarchical processing of the image is carried out. The set depth level is \(Z = 5\) meters, and the illumination, turbidity, and temperature at different scale levels are optimized hierarchically. According to these parameters, the optimization parameter \(\gamma\) of illumination 光照 is set to 1.1, the optimization parameter \(\gamma\) of turbidity 浑浊度 is set to 1.3, and the optimization parameter \(\gamma\) of temperature 温度 is set to 0.9. For the depth - related attenuation factor, \(\delta\) 光照 is set to 0.02, \(\delta\) 浑浊度 is set to 0.05, \(\delta\) 温度 is set to 0.03. The correction term \(\epsilon(z)=0.1z\) j , which is used to dynamically balance the optimization effects of each level. 2

[0230] Substitute these data into the formula to calculate the global optimization function at the depth \(Z = 5\) meters:

[0231]

[0232] Expand the calculation:

[0233] For illumination optimization:

[0234]

[0235] For turbidity optimization:

[0236]

[0237] For temperature optimization:

[0238]

[0239] Add the results of each part:

[0240] G(100, 150, 5) = 5.264 + 5.765 + 6.679 = 17.708

[0241] Therefore, the global optimization function at a depth of 5 meters is 17.708. This result shows that through hierarchical processing and combining optimizations at different scales, the system can effectively balance the global structure and local details, ensuring high precision and quality in the 3D reconstruction process.

[0242] Example 6:

[0243] In this example, during the underwater 3D reconstruction of underwater structures, in order to accurately extract high-level features in the images, especially the non-linear distortion features caused by environmental changes, a convolutional neural network (CNN) is introduced. Through multi-layer convolution and pooling operations, complex features in the images are gradually extracted and analyzed. This process is expressed by the following mathematical formula for feature extraction:

[0244]

[0245] In this formula, F(x, y, t) represents the feature map value at spatial position (x, y) and time t, which reflects the feature changes of the image under different environmental conditions. The parameter α i is the weight coefficient of different convolutional layers, with a value range of 0.5 to 1.5, used to adjust the influence of each convolutional layer on feature extraction; σ is the activation function, usually ReLU or Sigmoid, which captures complex feature relationships through non-linear mapping; W i is the weight of the convolutional kernel, · represents the weighted sum of the convolutional operation on the input feature g(x, y, t), and b i is the bias term, usually with a value range of -0.5 to 0.5, used to adjust and correct the distribution of features.

[0246] Apply CNN for feature extraction in underwater images at the research site of underwater structures. Set the current environmental data as light intensity 800 Lux, turbidity 1.2 NTU, water temperature 17 degrees Celsius, and the input function of image features as g(x,y,t) = 0.8x + 0.6y - 0.2t. Set the number of convolutional layers as N = 3, and the convolutional kernel weights are W 1 = 0.9, W 2 = 1.1, W 3 = 0.8, and the corresponding weight coefficients are α 1 = 1.2, α 2 = 0.8, α 3 = 1.0. The bias terms are set as b 1 = -0.1, b 2 = 0.2, b 3 = 0.0. Select ReLU as the activation function, that is, σ(x) = max(0,x).

[0247] Substitute these parameters into the formula to calculate the feature map value at time T = 10 seconds and spatial position (x = 50, y = 75):

[0248]

[0249] Expand the calculation:

[0250] For convolutional layer 1:

[0251]

[0252] After activation:

[0253]

[0254] For convolutional layer 2:

[0255]

[0256] For convolutional layer 3:

[0257]

[0258]

[0259] Add the results of each layer:

[0260] F(50,75,10) = 906 + 740.8 + 672 = 2318.8

[0261] Therefore, within 10 seconds of time, the feature mapping value at spatial position 50, 75 is 2318.8. This result indicates that through the multi-layer convolution and pooling operations of the CNN, the system can gradually extract high-level features in the image, especially those non-linear distortion features caused by environmental changes.

[0262] To further optimize the quality and authenticity of the image, a Generative Adversarial Network (GAN) architecture is introduced. The GAN consists of two parts: a generator and a discriminator. The generator uses the features extracted by the Convolutional Neural Network (CNN) to generate high-quality images, while the discriminator evaluates the authenticity of the generated images. The adversarial loss function of the GAN is used to optimize the performance of the generator and the discriminator, and its expression is as follows:

[0263]

[0264] Where L(G,D) represents the adversarial loss function between the generator G and the discriminator D, aiming to optimize the performance of the two networks during adversarial training. represents the expectation of the real image x. The discriminator D is optimized by judging whether the input image comes from real data; while evaluates the image output by the generator. The generator G deceives the discriminator by generating realistic images, thereby continuously improving its own generation ability.

[0265] In actual operation, the generator learns to generate high-quality images from the features extracted by the CNN before. Assuming the current feature mapping value is F(50, 75, 10) = 2318.8, the generator uses this as input for training. The settings of the convolutional layer include sampling the initial noise distribution z from a normal distribution with a mean of 0 and a standard deviation of 1. The initial parameters of the discriminator are set with weights between 0.8 and 1.2 and a learning rate of 0.001.

[0266] Substituting these parameters for adversarial training, first, the probability that the discriminator evaluates the real image is:

[0267]

[0268] Then, the generator generates an image based on the input noise z and is evaluated by the discriminator:

[0269]

[0270] Adding the two expected values together, the value of the adversarial loss function is obtained:

[0271] L(G,D) = -0.105 + (-0.511) = -0.616

[0272] During the adversarial training process, the generator continuously adjusts the weights and bias terms to optimize the quality of the generated images, making them more similar to real images; the discriminator optimizes its judgment ability by continuously distinguishing between real and generated images. Through multiple rounds of iteration in this process, the confrontation between the generator and the discriminator gradually tends to balance, and the images output by the generator become realistic.

[0273] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A method for three-dimensional reconstruction of an underwater environment, characterized in that The following steps are involved: S1, projecting the optimized spectral pattern onto the target surface, capturing the reflected light with a camera, and inferring the three-dimensional structure of the object surface through the deformation of light; in: S1.1, projecting multi-band structured light patterns in an underwater environment while capturing images using a multispectral camera; S1.2, through the spectrum matching algorithm, the images of different bands are fused, the depth information is extracted and the basic data for 3D reconstruction is generated; S1.

3. Use depth compensation technology to correct the depth error caused by underwater scattering and absorption; S2, dynamically adjust the image correction strategy by real-time monitoring of underwater environmental parameters; in: S2.

1. Use underwater sensors to monitor environmental parameters including light, turbidity, and temperature in real time and generate environmental characteristic models; S2.2, adaptively adjust image acquisition and processing parameters including exposure, contrast and color balance according to the environmental feature model; S2.3, through intelligent correction algorithm, detect the distorted areas in the image caused by environmental changes, and perform local correction and overall optimization; S3, by constructing a multi-level image pyramid model, first capture the overall structure at a large scale, and then gradually enhance the texture details at a small scale; in: S3.

1. Generate a multi-scale image pyramid to decompose the structural information of the three-dimensional image layer by layer from coarse to fine; S3.2, using a generative adversarial network to enhance details and restore textures at each layer; S3.3, by using adaptive detail enhancement technology, the key areas are refined to generate underwater three-dimensional images; The process of generating the environmental feature model comprises: S1. First, in the underwater environment, real-time monitoring of multi-dimensional parameters such as light, turbidity, and temperature is performed, and dynamic environmental feature extraction technology is used to perform nonlinear analysis and fusion of multi-source data; S2, then locally model the illumination, turbidity and temperature to capture the characteristics of different depths and regions; integrate the local models into a global environmental feature model and dynamically adjust it through a multi-level optimization algorithm; S3. Finally, the data from different sensors are deeply integrated through the intelligent sensor data fusion mechanism, and future environmental changes are predicted through time series analysis. The core formula for data fusion and prediction is: Among them, P(t) represents the environmental change prediction parameter at time t, λ(t) is the fusion weight of different sensor data, which determines the influence of each sensor in the prediction result, κ is the frequency factor of the time series, which reflects the periodicity and trend of environmental changes, and θ(t) is the prediction correction term, which is used to correct the prediction result.

2. The method for three-dimensional reconstruction of an underwater environment according to claim 1, characterized in that The spectrum matching algorithm construction process includes: S1. Using multispectral imaging technology, multiple spectral information of visible light, infrared light and ultraviolet light is integrated into a multidimensional spectral feature matrix. The depth information of each pixel is calculated through the spectral response difference of each band. The calculation formula is expressed as: Where D(x,y) represents the depth value at the pixel point (x,y), λ is the spectral wavelength, S i (λ) is the spectral response of the i-th band, α i and β i (λ) is the coefficient and absorption characteristics used to weigh the contributions of different bands; S2. Adjust the parameters of the spectral matching algorithm according to the real-time changes of the underwater environment; dynamically update the matching function to ensure the accuracy of the image by monitoring the illumination and turbidity environmental parameters; the correction process is achieved through the adaptive adjustment function: Among them, Φ(t) is the spectral matching function after dynamic adjustment, γ(t) represents the influence of environmental parameters on the correction function, κ is the adjustment frequency factor, and Γ(t) is the composite function of ambient light intensity and turbidity; S3. Utilize nonlinear fusion models to integrate spectral information of different bands into a multi-dimensional feature space, and perform multi-level feature mapping and deep information extraction through a deep learning network.

3. The method for three-dimensional reconstruction of an underwater environment according to claim 1, characterized in that The depth compensation technology adopts the following scheme: S1. By establishing an underwater optical transmission model, the scattering path and absorption coefficient of light during underwater propagation are analyzed, and the nonlinear effects of scattering and absorption on depth measurement are dynamically calculated; simulation is performed using the path integral formula, which is expressed as follows: I(x,y,z)=∫0 L [η(λ)·e -α(λ)·s ·cos(θ(s))]ds Among them, I(x,y,z) represents the light intensity at the point (x,y,z), which describes the energy of light propagating in an underwater environment; λ is the wavelength of light, and α(λ) is the absorption coefficient, which represents the degree of attenuation of light at a specific wavelength; S2. Dynamically adjust the parameters of the correction model according to the underwater environmental parameters monitored in real time; through a multi-level correction mechanism, linear correction of the light scattering coefficient is used in shallow areas, and depth compensation based on nonlinear scattering paths is used in deep and high turbidity areas; S3. Through the parallel processing architecture, scatter compensation and depth reconstruction are performed in real time, and the compensation parameters are dynamically adjusted using the depth error feedback mechanism.

4. The method for three-dimensional reconstruction of an underwater environment according to claim 1, characterized in that The construction method steps adopted by the intelligent correction algorithm include: S1. Identify local distortion areas caused by environmental changes through multi-dimensional matching of environmental parameters and image features; and distinguish and locate distortion areas based on an intelligent distortion detection model; S2, use nonlinear optimization method to gradually correct the distorted area; at the same time, combine multi-scale optimization technology to perform layered processing on the entire image; S3. Use the convolutional neural network (CNN) to extract and analyze the features of the distorted area, and then reconstruct the image through the generative adversarial network (GAN).

5. The method for three-dimensional reconstruction of an underwater environment according to claim 4, characterized in that The construction process of the local distortion area identification method includes: S1, by comprehensively matching the environmental parameters of illumination, turbidity, and temperature with the image features of brightness, texture, and edge, a high-dimensional feature space is formed to identify the local distortion areas caused by environmental changes; S2. Based on the deep learning algorithm, the intelligent distortion detection model learns different types of distortion features through training of environmental data and image samples, and establishes a multi-level distortion recognition system; S3. Finally, through the collaborative work of multi-dimensional matching and intelligent detection model, the environmental parameters and image features are analyzed and processed synchronously; and the collaborative detection mechanism combines the joint evaluation of the matching matrix and the intelligent model to perform in-depth analysis and positioning of the local distortion area; and the collaborative detection formula is as follows: Among them, C(t) represents the collaborative detection result at time t, which is a comprehensive evaluation of the coupling relationship between environmental parameters and image features; the parameter λ(t) is the weight of collaborative detection, which measures the interaction between environmental and image features in distortion detection; ω is the frequency factor of the time series, which is used to capture the periodic changes of environmental and image features; the correction term θt is used to dynamically adjust the detection results.

6. The method for three-dimensional reconstruction of an underwater environment according to claim 4, characterized in that The stepwise correction of the distorted area is to perform multiple rounds of iterative correction on the distorted area by combining dynamic feedback of environmental parameters and image features; The hierarchical processing of the image is to divide the image into multiple scale levels and perform optimization processing at different scales; the expression of the multi-scale optimization process is as follows: Among them, G(x,y,z) represents the global optimization function at the spatial position (x,y,z), γ j is the optimization parameter at different scales, δ j is a scale-dependent attenuation factor, describing the optimization effects at different levels, ∈ j (z) is the optimization correction term, which is used to balance the global structure and local details.

7. The method for three-dimensional reconstruction of an underwater environment according to claim 4, characterized in that The feature extraction and analysis described above introduces a convolutional neural network (CNN) to gradually extract high-level features in the image through multi-layer convolution and pooling operations, including nonlinear distortion features caused by environmental changes; The generator of the generative adversarial network GAN generates high-quality images using the features extracted by CNN, while the discriminator evaluates the authenticity of the generated images; the expression of the loss function is as follows: Among them, L(G,D) represents the adversarial loss function between the generator G and the discriminator D; is the expectation of the real image x, and the discriminator D is optimized by judging whether the input image comes from real data or the generator; It is to evaluate the images output by the generator, which deceives the discriminator by generating images.

Images