Submarine landform image processing method

Abstract

The invention discloses a submarine landform image processing method, and relates to the technical field of computer processing, and the method comprises the following steps: S01, obtaining sonar point cloud data and an optical original image during the cooperative detection of a multi-beam sonar and an optical device; wherein the sonar point cloud data is composed of a space coordinate component, a reflection intensity component and a beam incident angle component; s02, performing dynamic grid division on the sonar point cloud data to generate an initial grid model matched with the optical image resolution; and S03, embedding a water body attenuation compensation factor in the initial grid model, and correcting scattering noise of laser scanning and optical images in the turbid water area to obtain a corrected initial grid model. According to the method, cross-resolution high-fidelity fusion of sonar-laser-optical multi-source data is realized, the problems of scattering noise interference and micro-landform feature loss in turbid water areas are synchronously solved, and a seabed landform image with geometric accuracy and clear texture is output.

Description

Technical Field

[0001] The present invention relates to the field of computer processing technology, and in particular to a method for processing seabed landform images. Background Art

[0002] Seabed topography image processing technology is an important support for marine resource exploration, seabed engineering construction, and marine scientific research. With the rapid development of deep-sea exploration technology, heterogeneous data sources such as multi-beam sonar, laser scanning, and optical imaging have gradually become the core means of seabed topography mapping. The demand for high-precision three-dimensional seabed modeling continues to grow in the fields of oil and gas resource exploration, submarine pipeline laying, deep-sea archaeology, etc. Traditional methods mainly rely on single sensor data (such as multi-beam sonar point clouds) to improve resolution through beam cross-block segmentation and water depth value fusion. A typical example is the transceiver beam cross-sub-block segmentation algorithm. However, as multi-source collaborative detection becomes mainstream, the alignment distortion and resolution fault problems of heterogeneous data are becoming increasingly prominent.

[0003] For example, a multi-beam sonar system uses a cross-block of transmitting and receiving beams as the minimum detection unit. Its actual acquisition area is affected by both the width of the transmitting beam and the depth of the seabed, resulting in a sharp drop in resolution in deep waters. However, new optical equipment such as the beyond-the-horizon underwater camera LRC-20 has achieved micron-level imaging accuracy, which can capture extremely subtle topographic features in close-range detection. When the two types of data are fused, the resolution difference can reach orders of magnitude, and traditional interpolation or downsampling methods will cause texture distortion or loss of details. In combination with the Chinese published patent, publication number CN119399406A, publication date 2025-02-07, a fast three-dimensional modeling method, medium and system based on seabed point cloud data is disclosed. Although the method improves modeling efficiency through point cloud downsampling and hybrid decomposition, it does not consider the dynamic interference of suspended matter in the water body on the optical data. Especially in turbid waters, scattering noise occurs in laser scanning and optical images, which makes it impossible to accurately align the initial mesh model constructed by the sonar with the optical details. Therefore, how to solve the problems of inconsistent resolution and feature distortion in the fusion of multi-source heterogeneous seabed geomorphological data and realize high-fidelity, real-time three-dimensional seabed modeling is expected to be well solved. Summary of the Invention

[0004] In response to the above technical problems, the technical solution adopted by the present invention is a method for processing seabed landform images, which includes the following steps: S01. Acquire sonar point cloud data and optical raw images during collaborative detection using a multi-beam sonar and an optical device; wherein the sonar point cloud data is composed of a spatial coordinate component, a reflection intensity component, and a beam incidence angle component; S02, dynamically meshing the sonar point cloud data to generate an initial mesh model that matches the resolution of the optical image; S03, embedding a water attenuation compensation factor into the initial grid model to correct the scattering noise of the laser scanning and optical image in turbid waters to obtain a corrected initial grid model; S04. Extracting structural contour features of the sonar data based on the modified initial grid model; S05, decomposing the optical original image into a basic brightness layer, a detail texture layer, and a color information layer, and then fusing the structural contour features with the detail texture layer to generate a high-precision texture map; S06. Extract the illumination distribution information in the basic brightness layer and the geometric feature information in the high-precision texture map respectively, and fuse the illumination distribution information and the geometric feature information to obtain a fused image of the seabed topography.

[0005] Preferably, the dynamic meshing of the sonar point cloud data in step S02 includes: S21, using the octree space segmentation algorithm to divide the sonar point cloud data into multiple subspaces; S22. In each subspace, the grid size is adaptively adjusted according to the point cloud density; S23, smoothing the mesh boundaries using a local weighted regression algorithm to generate a continuous initial mesh model, or inputting the sonar point cloud data into a pre-trained dynamic mesh generation model to obtain an initial mesh model that matches the resolution of the optical image; S24, the dynamic grid generation model is trained with a sonar point cloud dataset and corresponding optical image resolution labels; S25. The sonar point cloud dataset is collected by a multi-beam sonar system under different seabed terrains; the optical image resolution label is obtained by a beyond-horizon underwater optical imaging device at the same detection position and is generated through manual annotation.

[0006] Preferably, embedding the water attenuation compensation factor in the initial grid model in step S03 includes: S31. Calculate the attenuation coefficient at different wavelengths based on the real-time monitoring data of suspended solids concentration in water bodies; S32. Introducing a dynamic weight function related to the attenuation coefficient at each node of the initial grid model; S33, generating an anti-interference grid model by correcting the scattering noise of the laser scanning and the optical image through weighted averaging.

[0007] Preferably, embedding the water body attenuation compensation factor in the initial grid model in step S03 further includes: S34. Use Mie scattering theory to establish a mathematical model of suspended matter concentration and attenuation coefficient in water. S35, fitting the measured data by the least squares method to optimize the mathematical model; S36, embedding the optimized mathematical model into each node of the initial grid model to generate a dynamic attenuation compensation field, or inputting the initial grid model into the attenuation compensation model to obtain an anti-interference grid model; The attenuation compensation model is obtained by training the initial grid model data set and the corresponding anti-interference grid model label; The initial grid model data set is obtained by collaborative detection using a multi-beam sonar system and over-the-horizon underwater optical imaging equipment; The anti-interference grid model label is obtained by manually marking the scattered noise distribution in turbid waters and generating a noise-free grid model.

[0008] Preferably, extracting the structural contour features of the sonar data in step S04 includes: S41, extracting structural contours from the modified initial grid model using an edge detection algorithm; S42. Optimize the continuity of contour lines through morphological operations; S43, converting the optimized contour line into a binary image to generate a structural contour feature map, or inputting the modified initial grid model into a structural contour extraction model to obtain a structural contour feature map; The structure contour extraction model is obtained by training the modified initial grid model data set and the corresponding structure contour labels; The modified initial grid model data set is obtained by collaborative detection using a multi-beam sonar system and over-the-horizon underwater optical imaging equipment; The structure contour label is generated by manually marking the structure contour of the seabed landform.

[0009] Preferably, the step S05 of decomposing the optical original image into a basic brightness layer, a detail texture layer, and a color information layer includes: S51. Use Retinex theory to decompose the original optical image into a basic brightness layer and a reflection component layer; S52, extracting a detail texture layer from the reflection component layer using a bilateral filtering algorithm; S53, subtracting the original image from the detail texture layer to generate a color information layer, or inputting the optical original image into an image decomposition model to obtain a basic brightness layer, a detail texture layer, and a color information layer; The image decomposition model is trained using an optical raw image dataset and corresponding hierarchical labels; The optical raw image dataset is collected by over-the-horizon underwater optical imaging equipment in different seabed terrains (such as coral reefs, deep sea plains, and hydrothermal vents); The layered labels are generated by manually annotating the basic brightness layer, the detail texture layer and the color information layer.

[0010] Preferably, generating a high-precision texture map in step S05 includes: S54, performing pixel-level alignment on the structure contour feature map and the detail texture layer; S55, using a weighted fusion algorithm to merge the structural outline and the detail texture; S56, optimizing the edge smoothness of the fusion result by a guided filtering algorithm to generate a high-precision texture map, or inputting the structural contour feature map and the detail texture layer into a texture mapping model to obtain a high-precision texture map; The texture mapping model is trained using a structure contour feature map dataset, a detail texture layer dataset, and corresponding high-precision texture map labels; The structural profile feature map dataset is obtained by collaborative detection using a multi-beam sonar system and over-the-horizon underwater optical imaging equipment. The detailed texture layer data set is collected by over-the-horizon underwater optical imaging equipment; and the high-precision texture map labels are generated by manual annotation.

[0011] Preferably, obtaining the seabed topography fusion image in step S06 includes: S61: Map the basic brightness layer and the high-precision texture map to independent color spaces to obtain a first mapping image corresponding to the basic brightness layer and a second mapping image corresponding to the high-precision texture map. S62, extracting a brightness channel containing illumination distribution information from the first mapping image; S63, extracting a texture channel and a color channel containing geometric feature information from the second mapped image; S64. Fuse the brightness channel, texture channel, and color channel to generate a fused image of the seabed topography.

[0012] The present invention has at least the following beneficial effects: 1. Through dynamic meshing technology, the sonar point cloud data is generated into an initial mesh model that matches the optical image resolution, and the sonar structure contour features and optical detail texture layer are integrated. This effectively solves the resolution fault problem of sonar, laser, and optical equipment, avoids texture blurring or detail loss caused by traditional interpolation or downsampling, and significantly improves the geometric accuracy and texture fidelity of seabed landform images.

[0013] 2. A water attenuation compensation factor is introduced, and a dynamic mathematical model of suspended matter concentration and attenuation coefficient is established based on Mie scattering theory. The scattering noise of laser scanning and optical images is corrected by weighted averaging, and forward / backward scattering compensation mechanisms are distinguished. This significantly improves imaging robustness in turbid waters and reduces feature distortion caused by suspended matter interference.

[0014] 3. Image layer decomposition technology is used to separate the basic brightness layer, detailed texture layer and color layer of the optical image. Combined with the morphological optimization of the sonar structure contour, high-precision texture mapping is generated through pixel-level alignment and weighted fusion to achieve a detailed restoration of the seabed micro-topography.

[0015] 4. The illumination distribution information of the basic brightness layer and the geometric features of the texture map are separated and processed in an independent color space. Through channel-level fusion, the illumination consistency and geometric structure authenticity are preserved, solving the problem of uneven illumination caused by water absorption or equipment differences, and outputting a fused image of the seabed topography with natural colors and clear details. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 A flow chart of a method for processing seabed landform images provided in Example 1 of the present invention; Figure 2 This is a flowchart of S02 provided in the first embodiment of the present invention; Figure 3 This is a flowchart of S03 provided in the first embodiment of the present invention; Figure 4 This is a flowchart of S04 provided in the first embodiment of the present invention; Figure 5 This is a flowchart of S05 provided in the first embodiment of the present invention; Figure 6 This is a flowchart of S06 provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way are interchangeable where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] Example 1

[0021] This embodiment provides a method for processing seabed landform images, including the following steps: Figure 1 As shown: S01. When the multi-beam sonar and optical equipment are used for collaborative detection, sonar point cloud data and optical original images are obtained; wherein the sonar point cloud data is composed of a spatial coordinate component, a reflection intensity component, and a beam incidence angle component; Specifically, this step includes a multi-beam sonar system and optical equipment (laser scanning device and beyond-horizon underwater optical imaging equipment), among which: The multi-beam sonar system uses the cross-block of transmitting and receiving beams as the minimum detection unit, and its acquisition area is limited by the transmit beam width and the seabed depth. Laser scanning devices can obtain high-precision point cloud data, but they are easily disturbed by suspended matter in turbid waters and produce scattered noise; Beyond-the-horizon underwater optical imaging equipment (such as the LRC-20 camera) achieves micron-level imaging accuracy (approximately 5.86 microns) and can capture seabed micro-topography features, but its effective detection range is limited by the water attenuation coefficient; However, in the process of fusion of multi-source data, there are problems of resolution fault (order of magnitude difference) and dynamic interference distortion. Traditional interpolation or downsampling methods will lead to texture blur or detail loss.

[0022] Furthermore, the transmit beam width of the multi-beam sonar system includes an adjustable range of 3°-5°, and the corresponding seabed coverage width increases linearly with increasing depth; The operating band of over-the-horizon underwater optical imaging equipment includes the blue-green light window of 450nm-550nm, which has the smallest attenuation coefficient in water. The spatial coordinate components of sonar point cloud data include X, Y, and Z three-dimensional coordinates; the reflection intensity component is related to the acoustic impedance of the seabed medium; The beam incident angle component affects the spatial distribution density of point cloud data.

[0023] A triangular mesh topology is generated using the sonar point cloud as vertices. The mesh density is inversely constrained based on the optical image resolution (e.g., 0.05 mm / pixel). For example, in shallow water (where sonar resolution is higher), a small mesh is generated to match optical details; in deep water (where sonar resolution is lower), the mesh is denser through interpolation. The larger the angle of incidence, the sparser the point cloud's spatial distribution. The algorithm automatically identifies low-density areas and uses local weighted regression to smooth the boundaries and avoid mesh breaks. It should be noted that the vertex spacing of the sonar mesh (e.g., 1 meter) differs from the optical pixel scale (5.86 μm) by over five orders of magnitude. This matching is achieved through a hierarchical mapping strategy: 1. The sonar mesh defines the macroscopic geometry of the terrain (e.g., ridge orientation). 2. Optical image pixels are mapped to sonar mesh triangles, each carrying tens of thousands of optical pixels, achieving a "coarse framework + fine texture" coupling.

[0024] S02, dynamically meshing the sonar point cloud data to generate an initial mesh model that matches the resolution of the optical image; Specifically, such as Figure 2 As shown in the figure, the dynamic meshing of sonar point cloud data in the above steps includes: S21, using the octree space segmentation algorithm to divide the sonar point cloud data into multiple subspaces; S22. In each subspace, the grid size is adaptively adjusted according to the point cloud density; S23, smoothing the mesh boundaries using a local weighted regression algorithm to generate a continuous initial mesh model, or inputting the sonar point cloud data into a pre-trained dynamic mesh generation model to obtain an initial mesh model that matches the resolution of the optical image; S24, the dynamic mesh generation model is trained with the sonar point cloud dataset and the corresponding optical image resolution labels; S25, sonar point cloud datasets were collected using a multi-beam sonar system under different seabed terrains; optical image resolution labels were obtained using over-the-horizon underwater optical imaging equipment at the same detection location and generated through manual annotation.

[0025] In the above technology, the overall sonar point cloud space is first recursively partitioned into an octree structure, with each subspace representing a cubic region of varying scale. Within each subspace, the point cloud density (i.e., the number of points per unit volume) is calculated in real time: If the density is higher than the threshold (such as in shallow water steep slope areas), the subspace is further divided into smaller levels to generate a fine-grained grid to match the optical high resolution; If the density is lower than the threshold (e.g., deep water flat area), the segmentation is stopped and the coarse-grained grid is retained to avoid redundant calculations.

[0026] After the segmentation is completed, the grid boundaries of adjacent subspaces are smoothly interpolated using a local weighted regression algorithm to eliminate grid cracks caused by density mutations and generate a continuous and seamless initial grid model.

[0027] When faced with complex terrain (such as seamounts and hydrothermal vents), the octree algorithm may lose topographic features due to regular geometric segmentation. In this case, a pre-trained dynamic mesh generation model is enabled. The model input is point cloud data collected by multibeam sonar on different typical terrains (flat seabed, slopes, and seamounts), and the labels are the resolution annotation data of the optical camera (such as LRC-20) at the corresponding location. The model learns the mapping relationship from the spatial distribution characteristics of the point cloud (such as curvature changes and elevation variance) to the target mesh size. When inputting real-time sonar point clouds, the model automatically identifies the terrain category and outputs a non-uniform mesh topology. For example, a high-density triangular mesh is output at the edge of a steep slope (to match optical details) and a sparse quadrilateral mesh is output in flat areas (to improve computational efficiency).

[0028] S03, embedding a water attenuation compensation factor into the initial grid model to correct the scattering noise of the laser scanning and optical image in turbid waters to obtain a corrected initial grid model; Specifically, such as Figure 3 As shown in Figure 2, the water attenuation compensation factor embedded in the initial grid model includes: S31. Calculate the attenuation coefficient at different wavelengths based on the real-time monitoring data of suspended solids concentration in water bodies; S32, introducing a dynamic weight function related to the attenuation coefficient at each node of the initial grid model; S33, generating an anti-interference grid model by correcting the scattering noise of the laser scanning and the optical image through weighted averaging.

[0029] Furthermore, it also includes: S34. Use Mie scattering theory to establish a mathematical model of suspended matter concentration and attenuation coefficient in water. S35, fitting the measured data by the least squares method to optimize the mathematical model; S36, embedding the optimized mathematical model into each node of the initial grid model to generate a dynamic attenuation compensation field, or inputting the initial grid model into the attenuation compensation model to obtain an anti-interference grid model; The attenuation compensation model is trained based on the initial grid model dataset and the corresponding anti-interference grid model labels; The initial grid model dataset is obtained through collaborative detection using a multi-beam sonar system and over-the-horizon underwater optical imaging equipment. The anti-interference grid model label is obtained by manually annotating the scattered noise distribution in turbid waters and generating a noise-free grid model.

[0030] In the above technology, a quantitative relationship is established between the concentration of suspended solids in water and the attenuation of light / sound waves. The higher the suspended solids concentration, the stronger the scattering effect on specific wavelengths (such as 532nm laser and 450-550nm optical blue-green light). A dynamic weight function is embedded in each node of the initial mesh model (such as the vertex of the Delaunay triangle): W(λ,C)=e −β(λ)⋅d⋅C , where: e −β(λ is the wavelength-dependent attenuation coefficient, d is the real-time suspended solids concentration, and C is the propagation path length. Forward scattering compensation (for laser scanning) and backscattering compensation (for optical imaging) are used to correct for signal distortion caused by different scattering mechanisms. Forward scattering diffuses the laser point cloud, while backscattering causes haze in the optical image, requiring separate treatment.

[0031] When the physical model parameters are unknown, a pre-trained attenuation compensation model (deep learning network) is used: It takes as input an initial noisy mesh model (consisting of sonar coordinates and optical texture) and outputs an interference-resistant mesh model. This model is trained on a field-measured dataset: simultaneous sonar and optical data collected in various turbid waters (e.g., the Yellow River Estuary and the Yangtze River Estuary), with noise-free mesh labels manually annotated. The network directly predicts the compensation field by learning the spatial distribution of scattering noise (e.g., the accumulation characteristics of suspended solids near the bottom).

[0032] S04. Extracting structural contour features of the sonar data based on the modified initial grid model; Specifically, such as Figure 4 As shown, the above steps include: S41, extracting structural contours from the modified initial grid model using an edge detection algorithm; S42. Optimize the continuity of contour lines through morphological operations; S43, converting the optimized contour line into a binary image to generate a structural contour feature map, or inputting the modified initial grid model into a structural contour extraction model to obtain a structural contour feature map; The structure contour extraction model is trained with the modified initial grid model dataset and the corresponding structure contour labels; The revised initial grid model dataset is acquired through collaborative detection using a multi-beam sonar system and over-the-horizon underwater optical imaging equipment. Structural contour labels are generated by manually annotating the structural contours of the seabed landforms.

[0033] In the above technology, based on real-time monitoring of suspended solids concentration data in water, Mie scattering theory is used to establish a mathematical model for the attenuation coefficient of light and sound waves at different wavelengths. This model relates parameters such as suspended solids particle size distribution and refractive index to attenuation intensity, and optimizes model parameters by fitting measured data using the least squares method. At each node of the initial mesh model (i.e., the vertex of the Delaunay triangulation mesh), a dynamic weight function is embedded: W(x,y,z)=e −β(λ)⋅d , where e −β(λ) is the attenuation coefficient corresponding to wavelength β(λ), and d is the propagation path length from the node to the sensor. This function acts on the laser point cloud (forward scatter compensation) and the optical image (backscatter compensation): Laser scanning: Compensates for point cloud position drift caused by forward scattering of suspended objects and restores the true terrain by adjusting the point cloud coordinate weights; Optical imaging: Suppress the image fogging effect caused by backscattering and enhance texture contrast.

[0034] Finally, an anti-interference grid model is generated, and its node coordinates and attribute values ​​are corrected by weighted average to eliminate the influence of dynamic noise.

[0035] S05, decomposing the original optical image into a basic brightness layer, a detail texture layer, and a color information layer, and then fusing the structural contour features with the detail texture layer to generate a high-precision texture map; Specifically, such as Figure 5 As shown, the above steps decompose the original optical image into a basic brightness layer, a detail texture layer, and a color information layer, including: S51. Use Retinex theory to decompose the original optical image into a basic brightness layer and a reflection component layer; S52, extracting a detail texture layer from the reflection component layer using a bilateral filtering algorithm; S53, subtracting the original image from the detail texture layer to generate a color information layer, or inputting the optical original image into an image decomposition model to obtain a basic brightness layer, a detail texture layer, and a color information layer; The image decomposition model is trained with optical raw image dataset and corresponding hierarchical labels; The optical raw image dataset was collected by over-the-horizon underwater optical imaging equipment in different seabed terrains (such as coral reefs, abyssal plains, and hydrothermal vents); Hierarchical labels are generated by manually annotating the basic brightness layer, detail texture layer, and color information layer.

[0036] After obtaining an interference-resistant mesh model, this step first uses an edge detection algorithm (such as the Canny operator) to scan the mesh for areas of gradient abrupt changes. Because the modified mesh incorporates water attenuation compensation factors (forward and backscatter compensation), the reflection intensity values ​​of the mesh nodes more closely resemble the actual seafloor properties. The Canny operator uses a dual-threshold hysteresis process (a high threshold captures strong edges, while a low threshold connects weak edges) to identify topographic abrupt changes (such as trench and ridge boundaries) and generate an initial contour. Morphological operations (dilation and erosion) then repair broken contours: dilation fills small gaps caused by noise or resolution differences, while erosion removes isolated noise points, ultimately generating a continuous, closed structural contour. This contour is converted into a binary feature map (contour value = 1, background value = 0) to serve as the geometric skeleton of the sonar data.

[0037] The image is considered the product of an illumination component and a reflectance component. First, a multi-scale Gaussian convolution is performed on the original optical image to simulate the attenuation of light in seawater, isolating the base brightness layer (which includes the light and dark distribution caused by non-uniform illumination and water absorption). Next, a bilateral filter is used to process the reflectance component layer: a spatial Gaussian kernel preserves large-scale smooth regions, while a range Gaussian kernel protects gradient abrupt changes (such as sand ripples and reef edges), thereby extracting a detailed texture layer (containing only high-frequency information of microtopography). Finally, the original image is subtracted from the detailed texture layer to extract the color information layer (containing low-frequency color features such as sediment color and biofouling).

[0038] Furthermore, in the above embodiment, generating a high-precision texture map includes: S54, performing pixel-level alignment on the structure contour feature map and the detail texture layer; S55, using a weighted fusion algorithm to merge the structural outline and the detail texture; S56, optimizing the edge smoothness of the fusion result by a guided filtering algorithm to generate a high-precision texture map, or inputting the structural contour feature map and the detail texture layer into a texture mapping model to obtain a high-precision texture map; The texture mapping model is trained with a data set of structure contour feature maps, a data set of detail texture layers, and corresponding high-precision texture map labels; The structural contour feature map dataset is obtained through collaborative detection using a multi-beam sonar system and over-the-horizon underwater optical imaging equipment. The detailed texture layer dataset is collected by over-the-horizon underwater optical imaging equipment; high-precision texture map labels are generated by manual annotation.

[0039] In this technique, the binary structural contour feature map extracted by sonar (e.g., trench edges) is first aligned with the detailed texture layer of the optical image (e.g., sediment lines). Because the sonar grid model is dynamically partitioned to match the optical resolution, the two have a basis for alignment at the spatial scale. An affine transformation matrix is ​​used to correct for micron-level offsets caused by platform sway or differences in detection angles, ensuring strict overlap between contour lines and texture pixels. An alpha blending algorithm is used to dynamically balance the contribution weights of contour and texture: Give the sonar structure a higher weight (α>0.7) near the contour line (such as the ridge boundary) to strengthen the rigidity of the geometric skeleton; In flat areas (such as deep-sea plains), emphasis is placed on optical texture weight (α<0.3) to preserve the fine layers of sediment.

[0040] After fusion, edge optimization is performed through guided filtering: using the sonar contour as a guide map, the filtering direction of the texture layer is constrained so that the fusion boundary transitions naturally along the terrain and eliminates the jagged effect. When using a pre-trained texture mapping model, the model learns cross-modal feature associations through an encoder-decoder architecture: On the encoder side, a dual-branch CNN is used to extract the topological structural features of the contour map (such as edge curvature) and the local statistical features of the texture layer (such as gradient distribution). Fusion layer: Automatically focuses on key areas (such as the junction of steep and gentle slopes) through the attention mechanism to suppress the false texture of optical images caused by water refraction; On the decoder side: an adversarial loss function (GAN) is introduced during upsampling reconstruction to ensure that the output texture map has both structural accuracy and visual realism.

[0041] S06. Extracting illumination distribution information from the basic brightness layer and geometric feature information from the high-precision texture map respectively, and fusing the illumination distribution information and the geometric feature information to obtain a fused image of the seabed topography.

[0042] Specifically, such as Figure 6 As shown, the above steps include: S61: Map the basic brightness layer and the high-precision texture map to independent color spaces to obtain a first mapping image corresponding to the basic brightness layer and a second mapping image corresponding to the high-precision texture map. S62, extracting a brightness channel containing illumination distribution information from the first mapping image; S63, extracting a texture channel and a color channel containing geometric feature information from the second mapped image; S64. Fuse the brightness channel, texture channel, and color channel to generate a fused image of the seabed topography.

[0043] In this technology, base luminance layer mapping converts the base luminance layer containing global illumination information (such as sunlight transmission and device fill light) into Lab color space, extracting the L channel as the light distribution carrier. This channel removes color interference and purifies it into a grayscale brightness gradient from 0 to 100%, accurately reflecting the absorption and scattering effects of water (especially the attenuation characteristics of blue-green light in the 450-550nm band).

[0044] Texture mapping involves converting a high-precision texture map (including geometric structure and micro-relief texture) to Lab space, separating the a channel (red and green hues) and the b channel (blue and yellow hues). These channels together represent the reflective properties of the seafloor (e.g., the red color of coral and the yellow-brown color of sediment), while texture details are expressed through spatial frequency variations in the a / b channels.

[0045] Furthermore, adaptive histogram equalization is performed on the L channel to compress overexposed areas (bright spots) and brighten dark areas (trench bottom), eliminating illumination discontinuities caused by water attenuation. A turbidity compensation coefficient (based on the Mie scattering model) is also injected to suppress brightness jumps caused by suspended matter. Adaptive histogram equalization is performed on the L channel to compress overexposed areas (bright spots) and brighten dark areas (trench bottom), eliminating illumination discontinuities caused by water attenuation. A turbidity compensation coefficient (based on the Mie scattering model) is also injected to suppress brightness jumps caused by suspended matter.

[0046] The corrected L channel is recombined with the optimized a / b channels to generate a fused image in Lab space. The L channel weight is dynamically adjusted (with a +15% weight in deep water to compensate for light attenuation) to ensure consistent brightness across the entire scene. When converting back to RGB space, a color gamut clipping algorithm is used to project colors outside the display's gamut (such as the fiery orange-red color of deep-sea hydrothermal vents) toward the displayable boundary, preventing color overflow. The human visual response curve is also overlaid to enhance the visual discernibility of details in dark areas.

[0047] This first embodiment uses dynamic meshing technology to generate an initial mesh model that matches the resolution of the optical image from sonar point cloud data, and integrates sonar structural contour features with optical detail texture layers. This effectively solves the resolution gap problem of sonar, laser, and optical equipment, avoids texture blurring or detail loss caused by traditional interpolation or downsampling, and significantly improves the geometric accuracy and texture fidelity of the seabed topography image. Secondly, a water attenuation compensation factor is introduced, and a dynamic mathematical model of suspended matter concentration and attenuation coefficient is established based on Mie scattering theory. The scattering noise of laser scanning and optical images is corrected through weighted averaging, and forward / backward scattering compensation mechanisms are distinguished. This significantly improves imaging robustness in turbid waters and reduces feature distortion caused by suspended matter interference. Furthermore, an image layered decomposition technique is used to separate the basic brightness layer, detail texture layer, and color layer of the optical image. Combined with morphological optimization of the sonar structural contour, a high-precision texture map is generated through pixel-level alignment and weighted fusion, achieving a refined restoration of the seabed microtopography. Furthermore, the illumination distribution information of the basic brightness layer and the geometric features of the texture map are separated and processed in an independent color space. Through channel-level fusion, the illumination consistency and geometric structure authenticity are retained, which solves the problem of uneven illumination caused by water absorption or equipment differences, and outputs a fused image of the seabed topography with natural colors and clear details.

[0048] Example 2

[0049] An embodiment of the present invention provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by a processor to implement the steps: When multi-beam sonar and optical equipment cooperate in detection, sonar point cloud data and optical original images are obtained; the sonar point cloud data consists of spatial coordinate components, reflection intensity components, and beam incidence angle components; Dynamically mesh the sonar point cloud data to generate an initial mesh model that matches the resolution of the optical image; A water attenuation compensation factor is embedded in the initial grid model to correct the scattering noise of laser scanning and optical images in turbid waters to obtain a corrected initial grid model. Extract the structural contour features of the sonar data based on the modified initial grid model; Decompose the original optical image into a basic brightness layer, a detail texture layer, and a color information layer, and then fuse the structural contour features with the detail texture layer to generate a high-precision texture map; The illumination distribution information in the basic brightness layer and the geometric feature information in the high-precision texture map are extracted respectively, and the illumination distribution information and the geometric feature information are fused to obtain a fused image of the seabed topography.

[0050] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0051] Those skilled in the art will clearly understand that for the sake of convenience and brevity in description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0052] Example 3

[0053] An embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to implement the following steps: When multi-beam sonar and optical equipment cooperate in detection, sonar point cloud data and optical original images are obtained; the sonar point cloud data consists of spatial coordinate components, reflection intensity components, and beam incidence angle components; Dynamically mesh the sonar point cloud data to generate an initial mesh model that matches the resolution of the optical image; A water attenuation compensation factor is embedded in the initial grid model to correct the scattering noise of laser scanning and optical images in turbid waters to obtain a corrected initial grid model. Extract the structural contour features of the sonar data based on the modified initial grid model; Decompose the original optical image into a basic brightness layer, a detail texture layer, and a color information layer, and then fuse the structural contour features with the detail texture layer to generate a high-precision texture map; The illumination distribution information in the basic brightness layer and the geometric feature information in the high-precision texture map are extracted respectively, and the illumination distribution information and the geometric feature information are fused to obtain a fused image of the seabed topography.

[0054] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for processing seabed landform images, characterized in that: The method comprises the following steps: S01. Acquire sonar point cloud data and optical raw images during collaborative detection using a multi-beam sonar and an optical device; wherein the sonar point cloud data is composed of a spatial coordinate component, a reflection intensity component, and a beam incidence angle component; S02, dynamically meshing the sonar point cloud data to generate an initial mesh model that matches the resolution of the optical image; S03, embedding a water attenuation compensation factor into the initial grid model to correct the scattering noise of the laser scanning and optical image in turbid waters to obtain a corrected initial grid model; S04. Extracting structural contour features of the sonar data based on the modified initial grid model; S05, decomposing the optical original image into a basic brightness layer, a detail texture layer, and a color information layer, and then fusing the structural contour features with the detail texture layer to generate a high-precision texture map; S06. Extract the illumination distribution information in the basic brightness layer and the geometric feature information in the high-precision texture map respectively, and fuse the illumination distribution information and the geometric feature information to obtain a fused image of the seabed topography.

2. The method for processing seabed landform images according to claim 1, characterized in that: The step S02 of dynamically meshing the sonar point cloud data includes: S21, using the octree space segmentation algorithm to divide the sonar point cloud data into multiple subspaces; S22. In each subspace, the grid size is adaptively adjusted according to the point cloud density; S23, smoothing the mesh boundaries using a local weighted regression algorithm to generate a continuous initial mesh model, or inputting the sonar point cloud data into a pre-trained dynamic mesh generation model to obtain an initial mesh model that matches the resolution of the optical image; S24, the dynamic grid generation model is trained with a sonar point cloud dataset and corresponding optical image resolution labels; S25. The sonar point cloud dataset is collected by a multi-beam sonar system under different seabed terrains; the optical image resolution label is obtained by a beyond-horizon underwater optical imaging device at the same detection position and is generated through manual annotation.

3. The method for processing seabed landform images according to claim 1, characterized in that: Embedding the water body attenuation compensation factor in the initial grid model in step S03 includes: S31. Calculate the attenuation coefficient at different wavelengths based on the real-time monitoring data of suspended solids concentration in water bodies; S32. Introducing a dynamic weight function related to the attenuation coefficient at each node of the initial grid model; S33, generating an anti-interference grid model by correcting the scattering noise of the laser scanning and the optical image through weighted averaging.

4. The method for processing seabed landform images according to claim 3, characterized in that: The water attenuation compensation factor embedded in the initial grid model in step S03 further includes: S34. Use Mie scattering theory to establish a mathematical model of suspended matter concentration and attenuation coefficient in water. S35, fitting the measured data by the least squares method to optimize the mathematical model; S36. Embed the optimized mathematical model into each node of the initial grid model to generate a dynamic attenuation compensation field, or input the initial grid model into the attenuation compensation model to obtain an anti-interference grid model.

5. The method for processing seabed landform images according to claim 1, characterized in that: The extraction of structural contour features of sonar data in step S04 includes: S41, extracting structural contours from the modified initial grid model using an edge detection algorithm; S42. Optimize the continuity of contour lines through morphological operations; S43, converting the optimized contour line into a binary image to generate a structural contour feature map, or inputting the corrected initial grid model into a structural contour extraction model to obtain a structural contour feature map.

6. The method for processing seabed landform images according to claim 1, characterized in that: The step S05 of decomposing the optical original image into a basic brightness layer, a detail texture layer, and a color information layer includes: S51. Use Retinex theory to decompose the original optical image into a basic brightness layer and a reflection component layer; S52, extracting a detail texture layer from the reflection component layer using a bilateral filtering algorithm; S53 , subtracting the original image from the detail texture layer to generate a color information layer, or inputting the optical original image into an image decomposition model to obtain a basic brightness layer, a detail texture layer, and a color information layer.

7. The method for processing seabed landform images according to claim 1, characterized in that: The step S05 of generating a high-precision texture map includes: S54, performing pixel-level alignment on the structure contour feature map and the detail texture layer; S55, using a weighted fusion algorithm to merge the structural outline and the detail texture; S56. Optimizing the edge smoothness of the fusion result through a guided filtering algorithm to generate a high-precision texture map, or inputting the structural contour feature map and the detail texture layer into a texture mapping model to obtain a high-precision texture map.

8. The method for processing seabed landform images according to claim 1, characterized in that: The step S06 of obtaining the seabed topography fusion image includes: S61: Map the basic brightness layer and the high-precision texture map to independent color spaces to obtain a first mapping image corresponding to the basic brightness layer and a second mapping image corresponding to the high-precision texture map. S62, extracting a brightness channel containing illumination distribution information from the first mapping image; S63, extracting a texture channel and a color channel containing geometric feature information from the second mapped image; S64. Fuse the brightness channel, texture channel, and color channel to generate a fused image of the seabed topography.

9. A non-transitory computer-readable storage medium, wherein at least one instruction or at least one program is stored in the non-transitory computer-readable storage medium, characterized in that: The at least one instruction or the at least one program is loaded and executed by the processor to implement the seabed topography image processing method according to any one of claims 1 to 8.

10. An electronic device, characterized in that: It includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the seabed topography image processing method as described in any one of claims 1 to 8.

Images