An underwater net cloth segmentation and embedded deployment method and system based on an improved YOLO architecture

Abstract

The application discloses a kind of underwater net clothes segmentation and embedded deployment method and system based on improved YOLO architecture. Method first constructs underwater net clothes multi-source topological data set;Second, deploy dynamic snake convolution in the neck network of improved YOLO architecture, extract net clothes slender manifold features;While embedding boundary perception loss in detection head to strengthen the accurate alignment of intersection and edge;In the training stage, using mask generation feature reconstruction mechanism, through teacher model guide lightweight student network to carry out structure performance distillation;Finally, the trained lightweight model is quantized and compiled, and deployed on an embedded edge platform. The application solves the technical pain point that thin grid structure is easy to break in complex underwater environment, and realizes high-fidelity, high-frame-rate real-time underwater visual perception on embedded devices with limited computing power.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and embedded system applications, specifically relating to an underwater mesh segmentation and embedded deployment method and system based on an improved YOLO architecture, which is particularly suitable for high frame rate and high-fidelity topology extraction of mesh-like targets in complex underwater environments. Background Technology

[0002] Underwater visual perception is a core technology for achieving autonomous monitoring and intervention in the marine environment, and it is particularly crucial in large-scale aquaculture systems, where the structural integrity of aquaculture facilities directly impacts ecological security and economic sustainability. Net cages, as the core infrastructure of aquaculture, are important physical boundaries for controlling biomass enclosure and water exchange. Therefore, reliable detection and segmentation of net cage structures are fundamental for tasks such as net cage damage assessment, robotic inspection, and long-term environmental monitoring. However, the underwater environment suffers from severe optical degradation, including light attenuation and scattering, as well as dynamic visual interference caused by suspended particles and biological activity, making stable underwater visual perception inherently challenging.

[0003] Beyond environmental challenges, the inherent geometry of net cages presents additional difficulties for traditional perception algorithms. Unlike conventional targets with compact regions and rich internal textures, net cages consist of thin, repeating mesh patterns, with their semantic information primarily concentrated on elongated boundaries and their connected regions. In this context, small local errors in boundary localization can spread into widespread structural inconsistencies, affecting both visual interpretation and the performance of downstream geometric inference tasks. This problem is further exacerbated in underwater environments, where noise, occlusion, and deformation can easily disrupt the already fragile structural continuity of net cages. This issue becomes even more critical when segmentation results serve as structural prior information for downstream tasks such as stereo reconstruction, geometric perception simultaneous localization, and map building.

[0004] The rapid development of deep learning architectures has significantly improved the performance of object detection and segmentation, driving their widespread application in multiple fields. Representative instance segmentation frameworks such as PointRend and Mask2Former have demonstrated excellent modeling capabilities for complex visual scenes; meanwhile, lightweight architectures derived from the YOLO series provide efficient solutions for real-time visual perception. Although the above methods have achieved significant results, they are all based on a region-centric learning paradigm, where feature aggregation and optimization are dominated by the spatial proportion of the target region. This assumption is poorly adapted to mesh-like structures, where semantic information is sparsely distributed on thin, high-curvature boundaries, while the internal regions have almost no discriminative features. Therefore, even if existing models can achieve high region-level accuracy, they are still prone to problems such as smooth boundaries, structural fragmentation, and topological inconsistencies. Directly applying existing segmentation frameworks to underwater cage scenarios often leads to unstable prediction results, especially under conditions of occlusion, low visibility, and geometric deformation. Furthermore, traditional high-precision models have extremely high inference latency on embedded devices, making it difficult to meet the real-time requirements of active high-frequency control for underwater robots. Summary of the Invention

[0005] To address the problems existing in the background technology, such as structural breakage, boundary distortion, and difficulty in real-time deployment on edge devices in underwater mesh segmentation, this invention proposes a topology-aware visual perception method. This method is used to understand and model structure-centric visual patterns, thereby solving the problems of poor topological sensitivity of traditional region-centered segmentation models to thin mesh structures and the difficulty of achieving high frame rate inference of high-performance segmentation algorithms on underwater edge computing platforms with limited computing power.

[0006] The technical solution adopted in this invention is: I. A Method for Underwater Mesh Segmentation and Embedded Deployment Based on an Improved YOLO Architecture Includes the following steps: S1: Acquire underwater net images and preprocess the underwater net images to construct a multi-source topology dataset, which is used as the training set. S2: Construct a topology-aware segmentation model using a multi-source topology dataset, that is, input the multi-source topology dataset as a training set into the topology-aware segmentation model to train the topology-aware segmentation model. The topology-aware segmentation model includes a neck network, a detection head, and a backbone network. The backbone network is connected to the neck network and the detection head. S3: By optimizing the neck network, the neck network extracts the slender manifold features related to the mesh in the multi-source topology dataset, thus obtaining a preprocessed model; S4: In the output of the detection head of the preprocessing model, a total segmentation loss function is constructed, and the boundary-aware loss is embedded into the total segmentation loss function as an independent constraint term. The penalty weight is increased to improve the segmentation prediction accuracy of the mesh edge, thus obtaining the teacher model and the student model. S5: A mask-generated feature reconstruction mechanism is adopted, and the structural representation distillation of the feature level is performed using the teacher model and student model to obtain a lightweight student model. The lightweight student model is the final topology-aware segmentation model based on the improved YOLO architecture. S6: Perform the embedded deployment step, quantize and accelerate the lightweight student model through the tensor inference engine, and then compile it into a hardware-optimized model inference execution program. The model inference execution program is finally deployed in the embedded platform to perform real-time underwater netting segmentation prediction.

[0007] Specifically, in the topology-aware segmentation model, the original image passes through the backbone network to extract multiple layers of basic features, transforming the original pixel image into a multi-scale feature map. The multi-scale feature map output by the backbone network is then fused with shallow and deep features by the neck network. Finally, the multi-scale feature map fused by the neck network passes through the detection head to perform the final calculation on the feature map, outputting two types of results: one is the segmentation mask and the target bounding box, and the other is the class confidence. The final output is the segmentation mask, restoring the feature map into the final segmentation + detection result map.

[0008] The improvements to the YOLO architecture are as follows: a dynamic snake-shaped convolution module is introduced into the neck network; and a boundary-aware loss is introduced into the total segmentation loss function.

[0009] In step S1, underwater net images of different resolutions are acquired using an underwater high-definition camera. The underwater net images are preprocessed by using an adaptive black border filling method that keeps the aspect ratio unchanged and unifies it to the standard input dimension of the network, thereby constructing a multi-source topology dataset.

[0010] The underwater netting image is preprocessed as follows: The underwater netting image is preprocessed using adaptive contrast enhancement and multi-scale noise reduction.

[0011] In step S3, the elongated manifold feature adopts the elongated, curved tubular structure feature of the mesh; by optimizing the neck network, the neck network extracts the elongated manifold features of the mesh in the training set as follows: By modifying the C3k2 modules in layers 13 and 16 of the neck network to dynamic serpentine convolutional modules, the neck network can extract the elongated manifold features of the mesh in the training set. The C3k2 module is a feature extractor in the YOLO architecture used to extract the desired elongated manifold features. The dynamic serpentine convolutional module adaptively adjusts the shape of the convolutional kernel to accurately capture elongated, curved tubular structural features, i.e., elongated manifold features. In step S3, the dynamic serpentine convolution module determines the sampling point coordinates of the mesh through a cumulative offset mechanism and captures the slender manifold features of the mesh, thereby obtaining the deformed coordinates of the sampling point coordinates of the mesh. The specific topology of the dynamic serpentine convolution module adopts a 3×3 spatial convolution layer, and each convolution layer is followed by a ReLU nonlinear activation function, the formula of which is as follows: f(x)=max(0,x); Where f(x) represents the output feature value of the ReLU nonlinear activation function, max represents the maximum value operation, and x represents the input feature value of the output of the previous 3×3 spatial convolutional layer. The calculation logic of the ReLU nonlinear activation function is as follows: when x>0, f(x)=x; when x<=0, f(x)=0. The aforementioned cumulative offset mechanism specifically abstracts the slender manifold features of the mesh into a continuous linear geometric trajectory in the spatial dimension. By establishing a first-order spatial continuity constraint between the sampling points of the convolution kernel, the offset range of each extended sampling point is limited when generating the deformation position of the convolution kernel. This requires fine-tuning based on the coordinates of the immediately preceding sampling point, thereby ensuring that all sampling points are connected in space as a continuous polygonal line. This makes the spatial coordinates of all extended sampling points, except for the center point, no longer independent of each other. Instead, they are chained together using the actual deformation position of their preceding adjacent sampling point as a reference point. This directly converts the geometric topology of the mesh into a sequence of absolute spatial coordinate changes of the sampling points.

[0012] The deformed coordinates of the sampling points of the mesh are: for multiple elongated manifold features of the mesh, the coordinates of the center point of the sampling points of the mesh are set. k i =( x i , y i The center point is along the horizontal x Deformation coordinates of axis extension ( x i±c y i+c ) is represented as: x i±c = x i ±c ; y i+c = y i +∑ △y j ; Where i represents the selected sampling point, c is the extension step size along the topological axis, and x i y represents the horizontal position of sampling point i. i This represents the vertical position of sampling point i. △y j Let j be the vertical coordinate offset learned through the lightweight convolutional layer, where j takes the value [i, i+c]. x i+c , y i+c These represent the horizontal and vertical coordinates of the actual sampling points after network learning and cumulative offset, respectively.

[0013] The coordinates of the center point are along the vertical... y Deformation coordinates of axis extension ( x i+c , y i+c ) is represented as: x i+c = x i +∑ △x j ; y i±c = y i ± c ; in, △x j This represents the horizontal coordinate offset learned through the lightweight convolutional layer. The dynamic serpentine convolution module uses the extension direction of the slender manifold feature of the mesh as a reference, and accumulates the offset of the previous position in a chain along the extension direction, that is, performs chain accumulation, and performs adaptive offset in the normal direction perpendicular to the extension direction, so that the sampling points of the convolution kernel adaptively fit the slender manifold feature in a serpentine shape.

[0014] In step S4, the boundary-aware loss is embedded as an independent constraint term into the total segmentation loss function, specifically as follows: Total segmentation loss function L seg Depend on L Dice and LBCE Two types of basic region masking loss and boundary-aware loss L boundary Composed of, the aforementioned L Dice and L BCE The two basic region masking losses are the Dice coefficient loss and the binary cross-entropy loss. The Dice loss focuses on region overlap, while the binary cross-entropy loss determines segmentation accuracy pixel-by-pixel. The total segmentation loss function is described above. L seg The formula is as follows: L seg = L BCE + L Dice +λ·L boundary ; in, λ The weighting coefficients are used to balance the penalty weights for adjusting boundary perception errors.

[0015] Calculate the boundary sensing loss L boundary At that time, the prediction mask is extracted using morphological operators. P and real label mask g boundary area B Then, the boundary intersection-union ratio (BIoU(g,P)) within the boundary region is calculated, and the formula for the boundary intersection-union ratio BIoU(g,P) is: BIoU(g,P)=|(g∩B)∩(P∩B)| / |(g∩B)∪(P∩B)|; The boundary intersection-union ratio is then converted into a loss value to construct a boundary-aware loss. L boundary .

[0016] In step S4, increasing the penalty weight specifically means: increasing the total segmentation loss function. L seg The value of the balancing weight coefficient λ in the model is used to increase the boundary sensing loss of the topology-aware segmentation model. L boundary The penalty weight of this independent constraint term is increased by balancing the weight coefficient λ to increase the penalty for errors in the intersections and line edges of the mesh, thereby improving the segmentation prediction accuracy of the mesh edges.

[0017] Step S5 specifically involves: The mask generation and feature reconstruction mechanism calculates the feature reconstruction loss after mask processing. Lmgd The aforementioned feature reconstruction loss L mgd The calculation formula is: in, C The number of channels in the feature map. H The height of the feature map in space. W The width of the feature map in space. T c,h,w For the full feature tensor of the teacher model, ( S ) c,h,w For the feature tensor of the student model, M h,w A randomly generated spatial pixel mask matrix. f align For linear operations used to align dimensions, G This is a lightweight feature projector composed of at least one convolutional layer and a neural network activation function. The lightweight feature projector employs lightweight components for efficient feature transformation, dimensionality reduction, or fusion. Each layer of the dynamic serpentine convolutional module is followed by a ReLU nonlinear activation function to construct the lightweight feature projector. In the feature forward flow logic, the lightweight feature projector receives the dimensionality-reduced and aligned incomplete student feature matrix as the input tensor. Through sliding aggregation calculations using a 3×3 local receptive field, it infers and reconstructs the features of the mesh line edges and mesh intersections in the occluded region from the remaining unoccluded visible pixels. Finally, the output of the lightweight feature projector and the full feature tensor are compared in terms of the number of channels. C Space height H Space width W The reconstructed feature map is completely consistent in the upper dimension. This projector is used to perform high-dimensional nonlinear mapping and topological manifold completion on the masked spatial incomplete tensor to generate undamaged reconstructed features that approximate the teacher network.

[0018] The mask generation feature reconstruction mechanism obtains the masked area by masking random pixels of student features and minimizes the feature reconstruction error of the masked area, forcing the student model to generate complete global topological features of the mesh and outputting a lightweight student model.

[0019] II. Underwater Mesh Segmentation and Embedded Deployment System Based on Improved YOLO Architecture The underwater netting segmentation and embedded deployment system includes: The data acquisition module acquires underwater netting images, preprocesses the underwater netting images, and constructs a multi-source topology dataset. The topology segmentation module has an embedded topology-aware segmentation model as described in claim 1. After receiving a multi-source topology dataset, it optimizes the topology-aware segmentation model and outputs a lightweight student model. The edge computing processing module accelerates the trained lightweight student model through a tensor inference engine, and then compiles it into a hardware-optimized model inference execution program. The model inference execution program is finally deployed in an embedded platform to perform real-time underwater netting segmentation and prediction.

[0020] The beneficial effects of this invention are: This invention proposes an underwater netting segmentation and embedded deployment method and system based on an improved YOLO architecture, redefining the netting perception task as a topology-aware structural segmentation problem. The method first constructs a multi-source topology dataset of underwater netting through adaptive image enhancement preprocessing. Second, DSConv is deployed in the neck network of the improved YOLO architecture to extract the slender manifold features of the netting in a geometrically adaptive manner. Simultaneously, a boundary-aware loss is embedded in the detection head to enhance the accurate alignment of intersections and edges. During the training phase, the MGD mechanism is used to guide a lightweight student network to perform structural representation distillation through a teacher model. Experimental results show that the proposed method maintains real-time inference efficiency on an embedded platform while significantly improving the structural fidelity and environmental robustness of the segmentation results. This invention establishes a topology-aware research perspective for the visual understanding of thin, sparse, and deformable structures in complex environments. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method steps provided in the embodiments of the present invention; Figure 2 This is a structural diagram of the improved YOLO model used in Example 1; Figure 3 This is a structural diagram of the DSConv module used in Example 1; Figure 4 This is a schematic diagram of the MGD mechanism used in Example 1; Figure 5 These are the results of different mainstream example algorithm models in Example 1 on the segmentation of typical mesh images; Figure 6 This is the detection effect of the improved YOLO network model in Example 1 after being deployed on an embedded device. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture includes the following steps: S1: Acquire underwater net images using a camera, preprocess the underwater net images, and construct a multi-source topology dataset, which is used as the training set; S2: Construct a topology-aware segmentation model based on an improved YOLO architecture using a multi-source topology dataset. This involves using the multi-source topology dataset as a training set to train the topology-aware segmentation model. The topology-aware segmentation model is based on the YOLO architecture and includes a neck network, a detection head, and a backbone network. The backbone network is connected to the detection head via the neck network. S3: By optimizing the neck network, the neck network can extract the slender manifold features related to the mesh in the multi-source topology dataset, thus obtaining the preprocessed model; S4: Optimize the detection head of the preprocessing model. Construct the total segmentation loss function at the output of the detection head of the preprocessing model. Embed the boundary-aware loss as an independent constraint term into the total segmentation loss function and increase the penalty weight to improve the segmentation prediction accuracy of the mesh edge, including the edge of the mesh hole and the broken mesh wire, to obtain the teacher model and student model. S5: In the model training phase, the Mask Generative Feature Reconstruction (MGD) mechanism is adopted, and the high-complexity teacher model is used to guide the topology-aware segmentation model based on the improved YOLO architecture as a lightweight student model to perform feature-level structural representation distillation, so as to obtain the trained lightweight student model. The lightweight student model is the final topology-aware segmentation model based on the improved YOLO architecture. S6: Perform the embedded deployment step, quantize and accelerate the lightweight student model through the tensor inference engine, and then compile it into a hardware-optimized model inference executor. The model inference executor is finally deployed on the embedded platform to perform real-time underwater netting segmentation prediction.

[0025] Specifically, in the topology-aware segmentation model, the original image passes through the backbone network to extract multiple layers of basic features, transforming the original pixel image into a multi-scale feature map. The multi-scale feature map output by the backbone network is then fused with shallow and deep features by the neck network. Finally, the multi-scale feature map fused by the neck network passes through the detection head to perform the final calculation on the feature map, outputting two types of results: one is the segmentation mask and the target bounding box, and the other is the class confidence. The final output is the segmentation mask, restoring the feature map into the final segmentation + detection result map.

[0026] The specific improvements to the YOLO architecture are as follows: a dynamic snake convolution module is introduced into the neck network; and a boundary-aware loss is introduced into the total segmentation loss function.

[0027] In step S1, the underwater netting image is preprocessed as follows: Underwater netting images of different resolutions were acquired using an underwater high-definition camera. The underwater netting images were preprocessed by using adaptive black border padding to unify the network standard input dimension while maintaining the aspect ratio, and a multi-source topology dataset was constructed.

[0028] Underwater netting images are preprocessed using adaptive contrast enhancement and multi-scale noise reduction to suppress optical degradation interference caused by the underwater environment.

[0029] In step S3, the elongated manifold features utilize the elongated, curved tubular structure features within the mesh; by optimizing the neck network, the specific elongated manifold features related to the mesh extracted from the training set by the neck network are as follows: By modifying the C3k2 modules in layers 13 and 16 of the neck network to dynamic snake convolution (DSConv) modules, the neck network can extract the slender manifold features of the netting in the training set. The C3k2 module is a feature extractor in the YOLO architecture, used to extract the required slender manifold features. The dynamic snake convolution module accurately captures slender, curved tubular structural features, i.e., slender manifold features, by adaptively adjusting the shape of the convolution kernel. Specifically, it extracts the anisotropic slender manifold features of the netting, which are the slender linear structural features of the underwater netting that are easily occluded and broken.

[0030] The specific topology of the dynamic snake-shaped convolution module adopts a 3×3 spatial convolutional layer, and each convolutional layer is followed by a ReLU nonlinear activation function, the formula of which is as follows: f(x) = max(0,x); Where f(x) represents the output feature value of the ReLU nonlinear activation function, max represents the maximum value operation, and x represents the input feature value of the output of the previous 3×3 spatial convolutional layer. The calculation logic of the ReLU nonlinear activation function is as follows: when x>0, f(x)=x; when x<=0, f(x)=0. In step S3, the dynamic serpentine convolution module determines the coordinates of the sampling points of the mesh through a cumulative offset mechanism and captures the slender manifold features of the mesh, i.e., the slender local structure of the mesh, thereby obtaining the deformation coordinates of the sampling points of the mesh. Specifically, the cumulative offset mechanism abstracts the slender manifold features of the mesh into a continuous linear geometric trajectory in the spatial dimension. By establishing a first-order spatial continuity constraint between the sampling points of the convolution kernel, when generating the deformation position of the convolution kernel, the offset range of each extended sampling point is limited, requiring fine-tuning based on the coordinates of the immediately preceding sampling point. This ensures that all sampling points are connected in space as a continuous polygonal line. This makes the spatial coordinates of all extended sampling points except the center point no longer independent, but rather use the actual deformation position of its preceding adjacent sampling point as a reference point for chain-like accumulation, thus directly converting the geometric topology of the mesh into a sequence of spatial absolute coordinate changes of the sampling points.

[0031] The deformed coordinates of the sampling points of the mesh are: For the multiple slender manifold features of the mesh, the center coordinates of the sampling points of the mesh are set. k i =( x i , y i ), standard convolution direction of center coordinates x Deformation coordinates of axis extension ( x i+c , y i±c ) is represented as: x i±c = x i ± c ; y i+c = y i +∑ △y j ; Where i represents the selected sampling point, c is the extension step size along the topological axis, and x i y represents the horizontal position of sampling point i. i This represents the vertical position of sampling point i. △y jLet j be the vertical coordinate offset learned through the lightweight convolutional layer, where j takes the value [i, i+c]. x i+c , y i+c These represent the horizontal and vertical coordinates of the actual sampling points after network learning and cumulative offset, respectively.

[0032] Standard convolution direction of center coordinates y Deformation coordinates of axis extension ( x i+c , y i±c ) is represented as: x i+c = x i +∑ △x j ; y i±c = y i ± c ; in, △x j This represents the horizontal coordinate offset learned through a lightweight convolutional layer.

[0033] The dynamic serpentine convolution module takes the extension direction of the slender manifold feature of the mesh as the reference, accumulates the offset of the previous position in a chain along the extension direction, and performs adaptive offset in the normal direction perpendicular to the extension direction, that is, accumulates the offset of the previous position in the direction perpendicular to the main growth direction, that is, performs chain accumulation, so that the sampling points of the convolution kernel adapt to the slender manifold feature, that is, the slender lines of the mesh, in a serpentine shape.

[0034] In step S4, the boundary-aware loss is embedded as an independent constraint term into the total segmentation loss function. The specific logic is as follows: Total segmentation loss function L seg Depend on L Dice and L BCE Two types of basic region masking loss and boundary-aware loss L boundary Combining L Dice and L BCE The two types of basic region masking losses are the Dice coefficient loss and the binary cross-entropy loss. The Dice loss focuses on region overlap, while the binary cross-entropy loss assesses segmentation accuracy pixel-by-pixel. The total segmentation loss function...L seg The formula is as follows: L seg = L BCE + L Dice +λ·L boundary ; in, λ The weighting coefficients are used to balance the penalty weights for adjusting boundary perception errors.

[0035] Calculate boundary sensing loss L boundary At that time, the prediction mask is extracted using morphological operators. P and real label mask G boundary area B Then, the boundary intersection-union ratio (BIoU) within the boundary region is calculated. The formula for the boundary intersection-union ratio BIoU(g,P) is: BIoU(g,P)=|(g∩B)∩(P∩B)| / |(g∩B)∪(P∩B)|; The boundary intersection-union ratio is then converted into a loss value to construct a boundary-aware loss function. L boundary The forced model shifts the prediction focus from the overlap of the region centers to the precise alignment of the mesh topology boundaries.

[0036] In step S4, increasing the penalty weight specifically means: by increasing the penalty weight of the independent constraint term of boundary perception loss in the topology-aware segmentation model, that is, the penalty weight of the intersection point and line edge error of the mesh, the topology boundary manifold alignment is achieved, thereby improving the segmentation prediction accuracy of the mesh edge.

[0037] Step S5 is as follows: The mask generation feature reconstruction mechanism calculates the feature reconstruction loss after mask processing. L mgd Feature reconstruction loss L mgd The calculation formula is: in, C The number of channels in the feature map. H The height of the feature map in space. W The width of the feature map in space. T c,h,w For the full feature tensor of the teacher model, ( S ) c,h,w For the feature tensor of the student model, Mh,w A randomly generated spatial pixel mask matrix. f align For linear operations used to align dimensions, G It is a lightweight feature projector composed of at least one convolutional layer and a neural network activation function. This projector is used to perform high-dimensional nonlinear mapping and topological manifold completion on the masked spatial incomplete tensor to generate undamaged reconstructed features that approximate the teacher network.

[0038] The lightweight feature projector employs lightweight components for efficient feature transformation, dimensionality reduction, or fusion. Each layer of the dynamic serpentine convolution module is followed by a ReLU nonlinear activation function to construct the lightweight feature projector. In the feature forward flow logic, the lightweight feature projector receives the dimensionality-reduced and aligned incomplete student feature matrix as the input tensor. Through sliding aggregation calculations using a 3×3 local receptive field, it infers and reconstructs the features of the mesh line edges and mesh intersections in the occluded region from the remaining unoccluded visible pixels. Finally, the output of the lightweight feature projector and the full feature tensor are compared in terms of the number of channels. C Space height H Space width W The reconstructed feature map is completely consistent in the upper dimension.

[0039] The masking feature reconstruction mechanism obtains the masked region by masking random pixels of student features, and uses the generator module to minimize the feature reconstruction error of the masked region, forcing the student model to generate complete global topological features of the net under the guidance of the teacher model, and outputting a lightweight student model.

[0040] The underwater mesh segmentation and embedded deployment system based on the improved YOLO architecture includes: The data acquisition module acquires underwater net cage video streams in real time, thereby acquiring underwater net images, and preprocesses the underwater net images to construct a multi-source topology dataset; The topology segmentation module has an embedded topology-aware segmentation model as described in claim 1. After receiving a multi-source topology dataset, it optimizes the topology-aware segmentation model, outputs a lightweight student model, and outputs a mesh segmentation mask. The edge computing processing module uses the Jetson Orin NX embedded platform with TensorRT acceleration engine. The trained lightweight student model is quantized and accelerated through the TensorRT inference engine, compiled into a computation stream adapted to the edge computing architecture, and deployed in the embedded platform to perform real-time underwater net segmentation prediction. The computation processing module uses FP16 mixed precision when performing inference.

[0041] This invention relates to the field of computer vision technology. First, it constructs a multi-source topology dataset of underwater mesh using adaptive image enhancement preprocessing. Second, it deploys dynamic serpentine convolution (DSConv) in the neck network of an improved YOLO architecture to extract the slender manifold features of the mesh in a geometrically adaptive manner. Simultaneously, it embeds boundary-aware loss in the detection head to enhance the accurate alignment of intersections and edges. During the training phase, it utilizes a mask-generated feature reconstruction (MGD) mechanism, using a teacher model to guide a lightweight student network in structural representation distillation. Finally, the trained lightweight model is quantized, compiled, and deployed on an embedded edge platform. This invention addresses the technical challenge of the fragility of thin mesh structures in complex underwater environments, achieving high-fidelity, high-frame-rate real-time underwater visual perception on computationally limited embedded devices.

[0042] Example 1: Specifically, in combination Figures 1 to 6 A method and system for underwater mesh segmentation and embedded deployment based on an improved YOLO architecture, specifically including the following steps: Step S1: Construction and preprocessing of multi-source topology datasets, specifically including the following sub-steps: Multi-source data acquisition: To capture the diverse structural morphologies of aquaculture cages, this embodiment constructs a multi-source dataset comprising actual underwater field data and controlled laboratory experimental data. Actual scene data was collected in a deep-sea aquaculture area near Dongji Island, Zhoushan. Field data acquisition utilized the FIFISHV-EVO underwater robot, equipped with a customized Hikvision stereo camera (60mm baseline, 110° field of view, IP68 protection standard), capable of full 6-DOF motion and centimeter-level depth control. To cover rare long-tail conditions such as cage damage, a controlled laboratory platform was built. By systematically adjusting lighting and mesh configuration, perturbation scene data conforming to physical laws, such as broken net wires and irregular deformation, were synthesized.

[0043] Adaptive Image Enhancement: To address issues such as reduced image contrast and color distortion caused by wavelength-dependent light absorption and scattering underwater, this embodiment introduces FUnIE-GAN as a preprocessing module. This model learns a nonlinear mapping from degraded underwater observation images to visually restored images, transforming the image enhancement problem into a conditional adversarial transformation problem. This enhancement process significantly improves the color constancy and local contrast of the image.

[0044] Geometric correction: The camera is calibrated underwater, and the video frames extracted from multi-source data acquisition are subjected to binocular stereo correction using the calibrated camera intrinsic and extrinsic parameters to eliminate lens distortion and ensure epipolar consistency.

[0045] Dataset construction: After the geometrically corrected images were labeled using the CVAT platform, a dataset containing 1086 high-resolution labeled images was finally constructed and divided into training and validation sets in an 8:2 ratio.

[0046] Step S2: Construct a geometrically adaptive topology-aware segmentation model. This embodiment uses the advanced real-time object detection model YOLO as the computational backbone network. To break the isotropic spatial distribution assumption brought about by traditional convolutional fixed-grid sampling, this network modifies the neck network of YOLO. The original C3k2 modules are modified at layers 13 and 16, and DSConv modules are deployed instead. The module structure is as follows... Figure 3 As shown.

[0047] For the center coordinates Standard convolution x The deformation coordinates of the axis extension are represented as follows: x i±c = x i ± c ; y i+c = y i +∑ △y j ; Standard convolution y The deformation coordinates of the axis extension are expressed as follows: x i+c = x i +∑ △x j ; y i±c = y i ± c ; in, △x and △y The coordinate offset is learned through a lightweight convolutional layer. c This represents the extension step size along the topological axis. It effectively avoids sudden spatial drift and maintains local geometric consistency. Since cumulative offset may generate fractional sampling coordinates, eigenvalues ​​are obtained through bilinear interpolation, defined as: f ( x i±c , y i+c )=∑ b(x c ,yc ; i,j ) f ( i,j ) ; in, f ( i,j ) is an integer position ( i,j Discrete feature map value at ) b (·) represents a separable bilinear interpolation kernel. Although DSConv was originally proposed for segmenting tubular structures, its formula does not impose global topological assumptions; the continuity constraint only applies locally along the principal axis, thus it is equally applicable to periodic mesh structures. In underwater cage segmentation tasks, this mechanism promotes feature alignment along mesh lines, stabilizes feature aggregation at nodes, and suppresses sampling drift towards dynamic background targets. The DSConv module adaptively fits the slender lines of the mesh in a serpentine pattern by accumulating the offset of the previous position perpendicular to the main growth direction.

[0048] Step S3: Contour-fidelity optimization based on boundary-aware loss. Fish cages have high aspect ratio topological features and are often embedded in complex underwater backgrounds. Traditional region basis loss functions such as Dice and BCE have inherent limitations when applied to such fine-scale structures. When the predicted mask and the true mask have high regional overlap but slight boundary misalignment, the gradient signal of the region basis loss decays rapidly. This phenomenon often leads to satisfactory region-level metrics but boundary erosion or topological fragmentation. To address this imbalance, Boundary IoU is integrated into the composite loss framework. To support end-to-end optimization, this embodiment uses morphological operations that approximate discretization with differentiable operators, allowing boundary extraction to be seamlessly embedded into the computational graph. The final loss formula simultaneously imposes alignment constraints on both spatial proportion and geometric configuration.

[0049] Step S3, which incorporates the multi-constraint formula for boundary intersection-union ratio loss, includes the following calculation process: Morphological expansion is approximated using a local maximum pooling operator in a differentiable space with a specific kernel size. D k ( X ) and corrosion E k ( X The operation extracts the residual between the original representation and its eroded version as a soft boundary. B ( X ); The formula for calculating the boundary intersection over union (BIU) loss is as follows: in P To predict the probability map, gFor discrete real labels, Represents the set of all spatial pixels. This represents a smoothing term to address the gradient decay phenomenon that traditional region-based loss functions are prone to in the segmentation of slender structures.

[0050] Total segmentation loss function L seg Depend on L Dice and L BCE Two types of basic region masking loss and boundary-aware loss Composed of: L seg = L BCE + L Dice +λ·L boundary ; in λ To balance the weighting coefficients as penalty weights for adjusting boundary perception errors, the weighting coefficients were determined through empirical evaluation on a customized underwater cage dataset. λ Setting it to 0.5 achieves the best balance between global shape consistency and fine-grained boundary accuracy without compromising the convergence stability that is primarily based on the region.

[0051] In calculation L boundary At that time, the prediction mask is extracted using morphological operators. P Compared to the real label mask g boundary area B Then, the boundary intersection ratio BIoU(g,P) within this boundary region is calculated using the following formula: BIoU(g,P)=|(g∩B)∩(P∩B)| / |(g∩B)∪(P∩B)|; Convert BIoU into loss value and construct L boundary This allows the model to shift its prediction focus from overlapping regions to precise alignment with the mesh topology boundaries.

[0052] Step S4: Knowledge Distillation Based on Mask Generative Feature Reconstruction. To maintain high-quality structural representation under strict computational constraints, a collaborative knowledge distillation mechanism is further introduced and implemented through mask generative feature reconstruction. This framework uses a larger YOLOm model as the teacher, providing high-fidelity structural priors. The lightweight model improved in steps S2 and S3 is used as the student, and optimization is performed under dual constraints of task-specific supervision and feature-level structure guidance. The MGD mechanism is illustrated below. Figure 4As shown.

[0053] The specific calculation steps of the MGD mechanism in step S4 include: set up T (l) and S (l) These represent the structure-aware teacher network and the lightweight instance segmentation model at the backbone network stage, respectively. l Feature map; With a predefined mask rate Sample a random binary space mask M (l) ; To simulate severe signal attenuation, masked student characteristics Calculations based on spatial masking show that: in This represents element-wise multiplication, utilizing a lightweight projection consisting of a convolutional layer and a ReLU activation function. G (·), reconstructing undamaged teacher characteristics from fragmented observational features containing breaks and omissions. : Span of the selected backbone network stage Within this framework, the distillation objective is to minimize the mean squared reconstruction error between the reconstructed features and the teacher features. The calculation formula is as follows: This forces compact, lightweight mesh instance segmentation models to infer and supplement missing mesh topology connections from partially observed features.

[0054] During the distillation process, a random binary space mask is generated. M h,w Student characteristics S Masking is performed, followed by a lightweight generator module. G Feature reconstruction is performed using two 3×3 convolutional layers with ReLU activation. Reconstruction loss. L mgd Defined as the reconstructed features and the teacher's total features tensor T Mean square reconstruction error between: in, C The number of channels in the feature map. H The height of the feature map in space. W The width of the feature map in space. f alignThis is a 1×1 convolution operation used for channel dimension alignment. This adversarial occlusion and generation process forces the lightweight student model to recover the complete mesh topological representation from partial observations.

[0055] Step S5: Embedded Edge Deployment and Mixed Precision Acceleration. After completing the model improvement, to verify the practical deployment capability of the proposed improved YOLO framework, this embodiment systematically optimizes the trained model and deploys it to an embedded edge environment using the TensorRT inference engine. Initial model development and training were performed in a PyTorch-based workstation environment. Subsequently, the model was exported to the Open Neural Network Exchange (ONNX) format using Opset version 17 to ensure seamless cross-platform interoperability. The resulting ONNX model was migrated to the Jetson Orin NX platform and compiled into a high-performance TensorRT engine using the trtexec tool provided by TensorRT 10.3. Throughout the deployment phase, the input resolution was strictly maintained at 640×640, consistent with the upstream training configuration. To analyze the relationship between numerical precision and inference throughput, this embodiment evaluates the framework in both single-precision (FP32) and half-precision (FP16) execution modes. The FP16 configuration utilizes the native mixed precision acceleration capability of the TensorRT engine to significantly accelerate matrix-intensive operations.

[0056] To ensure the feasibility of this invention and verify its practical application effects at the edge, all the above experiments were conducted in two heterogeneous computing environments: a high-performance computing platform for model training and evaluation, and an embedded edge computing platform for deployment and verification. Specifically: Model development and training environment: The high-performance computing platform is preferably a mobile workstation equipped with an NVIDIA GeForce RTX 5070 Ti graphics processing unit (GPU); the operating system is Windows 11, the deep learning framework is built based on PyTorch 2.8, and CUDA 12.8 is used for underlying hardware acceleration. Preferably, all models are trained for 220 epochs under a unified hyperparameter configuration; the training uses the MuSGD optimizer, with a batch size of 32, a learning rate of 0.01, a momentum of 0.937, and a weight decay of 0.0005; a single worker process is used to load data to ensure deterministic training; to stabilize model convergence, mosaic data augmentation is disabled in the last 20 epochs.

[0057] Edge deployment verification environment: Jetson Orin NX Super is used as the edge device. This device integrates an NVIDIA Ampere architecture graphics card with 1024 CUDA cores and an AI computing power of up to 157 TOPS. The device runs the Ubuntu 22.04 operating system, the development environment is built on PyCharm 2025.2, and the runtime environment is Python 3.9.

[0058] In terms of evaluation metrics, the experiment evaluated performance from two complementary perspectives: recognition accuracy and computational efficiency. Core metrics included: average accuracy (mAP50-95), model parameter count, model size, end-to-end frame rate (FPS), and floating-point operations per second (FLOPs). mAP50-95 measures detection performance by averaging the accuracy over multiple intersection-over-union (IoU) thresholds; compared to the commonly used mAP@0.5 metric, mAP50-95 considers a stricter IoU threshold, providing a more comprehensive assessment of positioning accuracy. The number of parameters and model size directly impact the memory consumption of the embedded hardware platform, which is particularly critical for underwater robots with limited onboard resources. FLOPs quantify the theoretical computational cost of the model and are an important reference for evaluating whether the algorithm can meet the real-time inference requirements of underwater inspection tasks. In this embodiment, the specific calculation formulas for the relevant evaluation metrics and computational costs are as follows: in K w and K h These are the width and height of the convolution kernel, respectively. C in,i For the first i Number of input channels for each convolutional layer C out,i Number of output channels H out and W out The spatial height and width of the output feature map; Latency total The total time T is for processing a single input image. preprocess,i For the first i Preprocessing time per sample, T inference,i T represents the forward propagation time of the network. postprocess,i This refers to post-processing time.

[0059] To comprehensively evaluate the performance of the proposed improved YOLO model and verify its effectiveness in underwater cage detection, this embodiment conducts comparative experiments with several representative instance segmentation frameworks. The selected baseline models include PointRend, Mask2Former, Mask R-CNN, SOLOv2, RTMDet-m, YOLOv8n-seg, and YOLOv12n-seg, covering classic two-stage methods and state-of-the-art lightweight real-time architectures. The comparison results are shown in Table 1, and the performance of different mainstream instance segmentation algorithms on typical cage images is illustrated in the figure. Figure 5 As shown in the figure, the above comparison results confirm that a better Pareto balance is achieved between segmentation accuracy, computational efficiency, and environmental robustness, outperforming mature baseline models and contemporary lightweight architectures. More importantly, the observed performance gains are consistent across both quantitative benchmarks and qualitative evaluations, indicating that the proposed method not only improves the numerical accuracy of the output but also enhances its structural reliability.

[0060] Table 1 shows the comparison results of the improved YOLO model with other mainstream models; To verify the practical application capabilities of the proposed framework, this embodiment conducts deployment experiments on the Jetson Orin NX platform, employing an inference optimization strategy based on TensorRT. To analyze the trade-off between computational efficiency and segmentation accuracy, both FP32 and FP16 accuracy modes are evaluated. Deployment performance is shown in Table 2, and the real-time detection effect on the mesh after deployment is as follows: Figure 6 As shown, the results indicate that after switching to FP16 accuracy, the detection speed was significantly improved to 79.4 FPS, achieving a significant system-level acceleration; at the same time, the model size was reduced to 9.4 MB, reducing the memory consumption of the embedded platform.

[0061] Table 2 shows the performance comparison results of the two precision modes after the embedded device was deployed; This embodiment also provides a system for implementing the above-described underwater netting segmentation and embedded deployment method. The system is functionally divided into a data acquisition module, a topology segmentation module, and an edge computing processing module. The collaborative logic between these modules is as follows: Data Acquisition Module: This module corresponds to step S1 described above. In this embodiment, this module is not only a logical data entry point but also includes physical hardware. Specifically, this module consists of an underwater robot (ROV) with 6 degrees of freedom motion control capabilities and a stereo camera mounted on it. During operation, the ROV carries the camera close to the net cage area. The stereo camera uses its 60mm baseline and 110° field of view to acquire a raw image stream containing the topological features of the net cage at a preset frequency. This module transmits the data to the subsequent processing module in real time via wired or wireless communication links.

[0062] Topology Segmentation Module: This module corresponds to steps S2 to S4 above. It is the core of the system's algorithm, running within the memory space of the edge computing platform. This module processes the input image stream frame by frame by calling a pre-trained and distilled topology-aware segmentation model. Its core logic involves using DSConv in the model's neck network to perform geometrically adaptive feature extraction on slender mesh lines, and utilizing a detection head enhanced with boundary-aware loss during training to output a mesh segmentation mask with high topological consistency.

[0063] Edge computing processing module: This module corresponds to step S5 above. In this embodiment, this module is the hardware computing power carrier of the system, preferably equipped with a Jetson Orin NX computing platform based on the NVIDIA Ampere architecture. This module compiles the deep learning model in the topology segmentation module into an inference engine optimized for the edge instruction set through the embedded TensorRT acceleration engine. In actual execution, this module achieves hardware-level acceleration of complex operators such as DSConv through FP16 mixed-precision computing, thereby ensuring that the system can output perception results at an end-to-end frame rate of higher than 70 FPS, supporting the carrier's real-time navigation and inspection tasks.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for underwater mesh segmentation and embedded deployment based on an improved YOLO architecture, characterized in that, Includes the following steps: S1: Acquire underwater netting images and preprocess the underwater netting images to construct a multi-source topology dataset; S2: Construct a topology-aware segmentation model using a multi-source topology dataset. The topology-aware segmentation model includes a neck network, a detection head, and a backbone network. The backbone network is connected to the neck network and the detection head. S3: By optimizing the neck network, the neck network extracts the slender manifold features related to the mesh in the multi-source topology dataset, thus obtaining a preprocessed model; S4: In the output of the detection head of the preprocessing model, a total segmentation loss function is constructed, and the boundary-aware loss is embedded into the total segmentation loss function as an independent constraint term. The penalty weight is increased to improve the segmentation prediction accuracy of the mesh edge, thus obtaining the teacher model and the student model. S5: A mask-generated feature reconstruction mechanism is adopted, and the structural representation distillation of the feature level is performed using the teacher model and student model to obtain a lightweight student model. S6: Perform the embedded deployment step, quantize and accelerate the lightweight student model through the tensor inference engine, and then compile it into a model inference execution program. The model inference execution program is finally deployed in the embedded platform to perform real-time underwater netting segmentation prediction.

2. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 1, characterized in that, Step S1 is as follows: Underwater netting images are acquired using an underwater camera. These images are then preprocessed using an adaptive black border filling method that maintains the aspect ratio while ensuring a uniform input dimension, thus constructing a multi-source topology dataset.

3. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 1, characterized in that, In step S3, the elongated manifold feature adopts the elongated, curved tubular structure feature of the mesh; by optimizing the neck network, the neck network extracts the elongated manifold features of the mesh in the training set as follows: By modifying the C3k2 modules in layers 13 and 16 of the neck network to dynamic serpentine convolutional modules, the neck network can extract the slender manifold features of the mesh in the training set.

4. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 3, characterized in that, In step S3, the dynamic serpentine convolution module determines the sampling point coordinates of the mesh through an accumulated offset mechanism and captures the slender manifold features of the mesh, thereby obtaining the deformation coordinates of the sampling points of the mesh. The deformed coordinates of the sampling points of the mesh are: For the mesh fabric, the center point coordinates of the sampling points of the mesh fabric are set. k i =( x i , y i The center point is along the horizontal x Deformation coordinates of axis extension ( x i±c y i+c ) is represented as: x i±c = x i ± c ； y i+c = y i +∑ △y j ； Where i represents the selected sampling point, and c is the extension step size along the topological axis. x i y represents the horizontal position of sampling point i. i This represents the vertical position of sampling point i. △y j Let j be the vertical coordinate offset learned through the lightweight convolutional layer, where j takes the value [i, i+c]. x i+c , y i+c These represent the horizontal and vertical coordinates of the actual sampling points after network learning and accumulated offset, respectively; The coordinates of the center point are along the vertical... y Deformation coordinates of axis extension ( x i+c y i±c ) is represented as: x i+c = x i +∑ △x j ； y i±c = y i ± c ； in, △x j This represents the horizontal coordinate offset learned through the lightweight convolutional layer. The dynamic serpentine convolution module uses the extension direction of the slender manifold feature of the mesh as a reference, accumulates the offset of the previous position in a chain along the extension direction, and performs adaptive offset in the normal direction perpendicular to the extension direction, that is, performs chain accumulation, so that the sampling points of the convolution kernel adaptively fit the slender manifold feature in a serpentine shape.

5. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 4, characterized in that, The dynamic snake convolution module uses a 3×3 spatial convolutional layer in its topology, and each convolutional layer is followed by a ReLU nonlinear activation function. The activation function formula is as follows: f(x)=max(0,x); Where f(x) represents the output feature value of the ReLU nonlinear activation function, max represents the maximum value operation, and x represents the input feature value of the output of the previous 3×3 spatial convolutional layer.

6. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 4, characterized in that, The aforementioned cumulative offset mechanism transforms the slender manifold features of the mesh into a linear geometric trajectory in the spatial dimension. By establishing a first-order spatial continuity constraint between the sampling points of the convolution kernel in the dynamic serpentine convolution module, the spatial coordinates of each peripheral sampling point of the linear geometric trajectory of the mesh, excluding the center point, are chained together using the actual deformation position of the previous adjacent sampling point as a reference point. This transforms the linear geometric trajectory of the mesh into a sequence of changes in the spatial absolute coordinates of the sampling points, thereby obtaining the deformation coordinates of the sampling points of the mesh.

7. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 1, characterized in that, In step S4, the total segmentation loss function is constructed, and the boundary-aware loss is embedded as an independent constraint term into the total segmentation loss function, specifically as follows: Total segmentation loss function L seg Mainly composed of L Dice and L BCE Two types of basic region masking loss and boundary-aware loss L boundary Composed of, the aforementioned L Dice and L BCE The two types of basic region masking losses are the Dice coefficient loss and the binary cross-entropy loss, respectively. The total segmentation loss function is... L seg The formula is as follows: L seg = L BCE + L Dice +λ·L boundary ; in, λ To balance the weighting coefficients, which serve as penalty weights for adjusting boundary perception errors; Calculate the boundary sensing loss L boundary At that time, the prediction mask is extracted using morphological operators. P and real label mask g boundary area B Then, the boundary intersection-union ratio (BIoU(g,P)) within the boundary region is calculated, and the formula for the boundary intersection-union ratio BIoU(g,P) is: BIoU(g,P)=|(g∩B)∩(P∩B)| / |(g∩B)∪(P∩B)|; The boundary intersection-union ratio is then converted into a loss value to construct a boundary-aware loss. L boundary .

8. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 1, characterized in that, In step S4, increasing the penalty weight specifically means: increasing the total segmentation loss function. L seg The value of the balancing weight coefficient λ in the model is used to increase the boundary sensing loss of the topology-aware segmentation model. L boundary The penalty weight of this independent constraint term is increased by balancing the weight coefficient λ to increase the penalty for errors in the intersections and line edges of the mesh, thereby improving the segmentation prediction accuracy of the mesh edges.

9. The underwater mesh segmentation and embedded deployment method based on the improved YOLO architecture according to claim 1, characterized in that, Step S5 specifically involves: The mask generation and feature reconstruction mechanism calculates the feature reconstruction loss after mask processing. L mgd The aforementioned feature reconstruction loss L mgd The calculation formula is: in, C The number of channels in the feature map. H The height of the feature map in space. W The width of the feature map in space. T c,h,w For the full feature tensor of the teacher model, ( S ) c,h,w M is the feature tensor of the student model. h,w A randomly generated spatial pixel mask matrix. f align For linear operations used to align dimensions, G It is a lightweight feature projector composed of at least one convolutional layer and a neural network activation function; The mask generation feature reconstruction mechanism obtains the masked region by masking random pixels of student features and minimizes the feature reconstruction error of the masked region, forcing the student model to generate complete global topological features of the mesh and outputting a lightweight student model. The student features are high-dimensional feature tensors output by the lightweight student model that are not compressed by the detection head.

10. An underwater netting segmentation and embedded deployment system based on an improved YOLO architecture, applied to the method described in any one of claims 1-8, characterized in that, The underwater netting segmentation and embedded deployment system includes: The data acquisition module acquires underwater netting images, preprocesses the underwater netting images, and constructs a multi-source topology dataset. The topology segmentation module has an embedded topology-aware segmentation model as described in claim 1. After receiving a multi-source topology dataset, it optimizes the topology-aware segmentation model and outputs a lightweight student model. The edge computing processing module accelerates the trained lightweight student model through a tensor inference engine and then compiles it into a model inference execution program. The model inference execution program is finally deployed in an embedded platform to perform real-time underwater netting segmentation prediction.

Images