A marine early warning information broadcast short video AI automatic rapid generation method

Abstract

The application provides a marine early warning information broadcast short video AI automatic rapid generation method, and belongs to the technical field of marine early warning short video generation. The application forms a structured demand parameter set by analyzing user demand through a large language model, calls an external interface to obtain weather forecast data and combines a numerical mode three-dimensional space-time matrix, processes through artificial intelligence, combines a TV-L1 regularization rendering pipeline and an ENO-WENO interpolation format to complete high gradient area anti-oscillation rendering, completes video frame sequence real-time encoding through a layered pipeline parallel architecture, realizes audio-visual frame level alignment through a CTC forced alignment algorithm, and finally generates text video prompt words through a large language model and drives a text video model to output a broadcast short video, thereby solving the technical problem of false oscillation stripes generated during high gradient meteorological field area rendering, which leads to distortion of key physical information such as typhoon intensity.

Description

Technical Field

[0001] This invention belongs to the field of marine early warning short video generation technology, specifically, it relates to an AI-based method for automatically and quickly generating short videos for broadcasting marine early warning information. Background Technology

[0002] The automatic generation of short videos for marine early warning information broadcasting is an important means of disseminating marine disaster prevention and mitigation information. Currently, the production of meteorological and marine early warning videos typically relies on a manual synthesis process, combined with two-dimensional visualization rendering technology of numerical model output fields. Uniform sampling grids are used to interpolate and render vector fields, scalar fields, and contour lines, supplemented by text-to-speech and nonlinear editing tools to complete the audio-visual synthesis. This process can meet basic broadcasting needs under normal sea conditions and has been widely used in operational scenarios such as meteorological stations and marine forecasting centers.

[0003] However, traditional uniform rendering meshes have inherent limitations in high-gradient meteorological field regions such as typhoon eyewalls and strong fronts. When there are strong discontinuities or high gradients in the meteorological field in local areas, the uniform interpolation format introduces Gibbs oscillations in the difference template that crosses the gradient discontinuities, which manifests as false stripe noise in video frames, masking key physical information such as typhoon intensity.

[0004] In current automatic generation of marine early warning videos, uniform rendering grids cannot adaptively allocate rendering resolution based on the local gradient intensity of the meteorological field, and traditional interpolation formats lack anti-oscillation mechanisms for discontinuous regions. This makes it difficult to balance visual accuracy and numerical stability in the rendering results of high-gradient regions. In other words, existing technologies suffer from technical problems such as generating false oscillating stripes and distorting key physical information such as typhoon intensity when rendering high-gradient meteorological field regions. Summary of the Invention

[0005] In view of this, the present invention provides an AI-based method for automatically and rapidly generating short videos of marine early warning information broadcasts, which can solve the technical problems in the prior art where false oscillating stripes are generated when rendering short videos of marine early warning information broadcasts in high-gradient meteorological field areas, leading to distortion of key physical information such as typhoon intensity.

[0006] This invention is implemented as follows: This invention provides an AI-powered method for automatically and rapidly generating short videos for marine early warning information broadcasting, comprising the following steps:

[0007] The system receives video generation requests from users, understands and decomposes these requests using a large language model, and extracts elements such as visual style, scene, broadcast area, date, and anchor image to form a structured set of request parameters.

[0008] Based on the structured requirement parameter set, weather forecast data of the target area is obtained through external interface. Simultaneously, the three-dimensional spatiotemporal matrix of numerical model of operational ocean wave height, period, wave direction and storm surge is read. The ocean scene spatiotemporal feature decoupled perception model is input, and the calculation of ocean wave classification matrix, shore thrust velocity matrix and tide level spatiotemporal field matrix is ​​completed. The spatial relationship of points of interest is integrated and the ocean hydrological feature description set is output.

[0009] The set of marine hydrological features is input into a dynamic meteorological field adaptive rendering mesh algorithm driven by weighted Voronoi topology partitioning. A weighted Voronoi diagram is constructed with the physical salience of cyclone center and extreme point as weights to generate an adaptive non-uniform rendering mesh. Combined with TV-L1 regularized rendering pipeline and ENO-WENO interpolation format, anti-oscillation rendering of high gradient regions is completed in GPU shader.

[0010] Based on a hierarchical pipelined parallel architecture, vector fields, scalar fields, and contour rendering are distributed to independent GPU streaming multiprocessor groups. CUDA graph technology is used to eliminate kernel function startup overhead, and inter-frame motion compensation differential coding is introduced to complete the real-time encoding output of video frame sequences.

[0011] Key physical event time nodes are extracted from the broadcast text, a semantic anchor time axis is constructed, and time nodes are injected as hard constraints into the text-to-speech generation process. The CTC forced alignment algorithm is used to establish bidirectional dynamic programming alignment at the audio frame level and the video frame level, compressing the audio-visual timestamp error to within the error threshold.

[0012] The structured requirement parameter set, marine hydrological feature description set, rendering frame sequence, audio alignment sequence and subtitle content are fused together by a large language model to generate complete text-based video prompts. Inputting these prompts into the text-based video model automatically generates a short broadcast video.

[0013] The structured requirement parameter set refers to a set of structured data that includes elements such as visual style, scene geographic information, broadcast date, anchor image, camera movement requirements, and overall atmosphere, formed after the large language model parses the user's natural language input.

[0014] The marine scene spatiotemporal feature decoupling perception model adopts an encoder-decoder backbone. The encoder consists of three levels of spatiotemporal convolutional blocks. Each level includes a two-dimensional spatial convolutional layer and a one-dimensional convolutional layer along the time axis. The encoder output is fed into a graph attention network branch. The input of each level of the decoder is composed of the encoder skip connection output and the topological importance mask matrix, which are weighted by a gated fusion layer.

[0015] The graph attention network branch constructs a contour map structure using contour extreme points as nodes, contour arc segments as edges, and curvature and arc length as edge attributes. Messages are passed through the graph attention network to calculate the topological importance score of each contour arc segment, forming a topological importance mask matrix.

[0016] The marine scene spatiotemporal feature decoupling perception model is equipped with a memory pressure adjustment function. The memory pressure index is calculated based on the current GPU memory occupancy rate, the current number of input matrix nodes and the current batch size. The number of attention heads in the graph attention network branch is dynamically adjusted according to the memory pressure index. The memory partition size of the graph attention network branch is reduced proportionally as the number of attention heads decreases.

[0017] The weighted Voronoi topology partitioning-driven dynamic meteorological field adaptive rendering grid algorithm constructs a weighted Voronoi map with pressure gradient modulus and absolute vorticity as weights. Cells in high-weight regions are automatically subdivided into high-resolution rendering sub-grids, while low-weight regions retain coarse grids. Incremental Voronoi repartitioning is triggered when the movement of feature point positions exceeds the grid movement threshold.

[0018] The TV-L1 regularized rendering pipeline uses the total variational L1 norm as the regularization term, which suppresses high-frequency oscillation noise while preserving the strong gradient boundaries of the image. A local gradient detection branch is embedded in the GPU shader, and when the local gradient exceeds the gradient detection threshold, the high-precision interpolation path of the ENO-WENO interpolation format is automatically triggered.

[0019] The ENO-WENO interpolation format avoids introducing Gibbs oscillations by selecting the smoothest interpolation template from multiple candidate templates. Gibbs oscillations refer to the spurious high-frequency oscillations that occur near discontinuities when interpolating signals with discontinuities or high gradients.

[0020] The layered pipeline parallel architecture distributes vector fields, scalar fields, and contour rendering to independent GPU stream multiprocessor groups for parallel execution. CUDA graph technology eliminates the scheduling overhead of starting kernel functions frame by frame by pre-recording kernel function call graphs and submitting them all at once.

[0021] The inter-frame motion compensation differential coding borrows from the inter-frame prediction mechanism to perform motion estimation and differential coding on the meteorological field at adjacent time points, and only re-renders the changed areas.

[0022] The semantic anchor timeline extracts key physical event time node sequences from the warning text in advance, including peak wind speed, extreme wave occurrence, and storm surge arrival, and injects them as hard time constraints into the text-to-speech generation process.

[0023] The CTC forced alignment algorithm establishes a bidirectional dynamic programming path between the audio frame sequence and the video frame sequence, calculates the optimal alignment mapping, and compresses the audio-visual timestamp error to within the error threshold, which is 40ms.

[0024] The marine hydrological feature description set is output by the marine scene spatiotemporal feature decoupling perception model, and expresses wave classification information, shore thrust velocity information, tide level information and overtopping shore section information in the form of natural language description.

[0025] The tidal level spatiotemporal field matrix is ​​processed by an independent tidal level calculation submodule. It takes storm surge data and astronomical tide data as input, and outputs a three-dimensional tidal level spatiotemporal field matrix after element-by-element superposition. The matrix is ​​then compared with the warning tidal level values ​​issued by government departments to identify the sections of the embankment that have overflowed.

[0026] Among them, the memory pressure index The calculation method is to divide each of the three data points by its respective mean on the standard test set and then calculate the weighted sum; when Maintain a head count of 4 at the same time. When the attention count is adjusted to 2, The number of heads to focus on is adjusted to 1.

[0027] This invention employs a dynamic meteorological field adaptive rendering mesh algorithm driven by weighted Voronoi topology partitioning, combined with the TV-L1 regularized rendering pipeline and the ENO-WENO interpolation format, to solve the technical problem of generating false oscillating stripes and causing distortion of key physical information such as typhoon intensity during the rendering of high-gradient meteorological field regions.

[0028] This invention constructs a weighted Voronoi diagram using the physical saliency of topological feature points such as cyclone centers and extreme points as weights. This automatically subdivides rendering cells in high-gradient regions into high-resolution submesh, thereby physically concentrating rendering resources on regions with the highest information density. The TV-L1 regularized rendering pipeline uses the total variational L1 norm as the regularization term, suppressing high-frequency oscillation noise while preserving strong gradient boundaries. The ENO-WENO interpolation format fundamentally avoids introducing Gibbs oscillations by selecting the smoothest template from multiple candidate templates for interpolation.

[0029] In summary, this invention solves the technical problem mentioned in the background art of generating false oscillating stripes and causing distortion of key physical information such as typhoon intensity during high-gradient meteorological field regional rendering. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2This is a schematic diagram of the OSTF-Net model structure and feature extraction process.

[0032] Figure 3 A schematic diagram of the adaptive rendering of the cell distribution of the weighted Voronoi topology.

[0033] Figure 4 This is a graph showing the audio-visual timestamp error distribution for the CTC forced alignment algorithm.

[0034] Figure 5 The graph shows the changes in the validation set classification accuracy and speed mean square error as a function of the training rounds during the training of the OSTF-Net model.

[0035] Figure 6 A bar chart showing the distribution of single-frame processing latency for each rendering task during the typhoon process. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0037] like Figure 1 The diagram shown is a flowchart of an AI-powered method for automatically and rapidly generating short videos for marine early warning information broadcasting, provided by this invention. This method includes the following steps:

[0038] S01. Receive the video generation requirements input by the user, understand and decompose the requirements through a large language model, and extract elements such as screen style, scene, broadcast area, date and anchor image to form a structured set of requirements parameters.

[0039] S02. Based on the structured requirement parameter set, obtain weather forecast data of the target area through external interface, simultaneously read the three-dimensional spatiotemporal matrix of the numerical model of operational ocean wave height, period, wave direction and storm surge, input the marine hydrological feature extraction model, complete the calculation of ocean wave classification matrix, shore thrust velocity matrix and tide level spatiotemporal field matrix, integrate the spatial relationship of points of interest, and output the marine hydrological feature description set required for video generation.

[0040] S03. Input the marine hydrological feature description set into the dynamic meteorological field adaptive rendering mesh algorithm driven by weighted Voronoi topology partitioning. Construct a weighted Voronoi diagram with the physical salience of topological feature points such as cyclone center and extreme point as weights to generate an adaptive non-uniform rendering mesh. Combine the TV-L1 regularized rendering pipeline and ENO-WENO interpolation format to complete the anti-oscillation rendering of high gradient regions in the GPU shader.

[0041] S04. Based on a hierarchical pipeline parallel architecture, vector fields, scalar fields, and contour rendering are distributed to independent GPU stream multiprocessor groups. CUDA graph technology is used to eliminate kernel function startup overhead, and inter-frame motion compensation differential coding is introduced to complete the real-time encoding output of video frame sequences.

[0042] S05. Extract key physical event time nodes from the broadcast text, construct a semantic anchor time axis, inject the time nodes as hard constraints into the text-to-speech generation process, and use the CTC forced alignment algorithm to establish bidirectional dynamic programming alignment at the audio frame level and video frame level, compressing the audio-visual timestamp error to within 40ms.

[0043] S06. The structured requirement parameter set, marine hydrological feature description set, rendering frame sequence, audio alignment sequence and subtitle content are fused together by a large language model to generate complete text-based video prompts. The text-based video prompts are then input into the text-based video model to automatically generate a short broadcast video.

[0044] The structured requirement parameter set refers to a set of structured data that includes elements such as visual style, scene geographic information, broadcast date, anchor image, camera movement requirements, and overall atmosphere, formed after the large language model parses the user's natural language input.

[0045] The marine hydrological feature extraction model is a decoupled perception model for spatiotemporal features of marine scenes, named OSTF-Net. The specific structure of the OSTF-Net model is as follows: the model adopts an encoder-decoder backbone. The input is a time-varying three-dimensional spatiotemporal matrix of ocean waves, with a matrix shape of longitude × latitude × time. The longitude and latitude resolution is between 0.05° and 0.25°, and the time step is 1 hour. The encoder consists of three levels of spatiotemporal convolutional blocks. Each level contains one two-dimensional spatial convolutional layer and one one-dimensional convolutional layer along the time axis. The spatial convolutional kernel size is 3×3, the temporal convolutional kernel size is 3, and the stride is 1. The activation function is ReLU, and the number of output channels for each level is 32, 64, and 128, respectively. The encoder output is fed into a graph attention network branch. This branch constructs a contour map structure using contour extreme points as nodes, contour arc segments as edges, and curvature and arc length as edge attributes. Message passing is performed through a four-layer graph attention network, with four attention heads per layer and an output dimension of 64. The topological importance score for each contour arc segment is calculated, forming a topological importance mask matrix. The decoder consists of three levels of deconvolutional blocks. Each level's input is formed by weighting the encoder's corresponding level skip connection output and the topological importance mask matrix through a gated fusion layer. The gated fusion layer uses a sigmoid gating mechanism, with output channels of 64, 32, and 16 respectively. The decoder's final output is mapped to a 9-level wave classification matrix and a wave shore thrust velocity matrix through two parallel linear layers. The tidal level spatiotemporal field matrix is ​​processed by an independent tidal level calculation submodule. This submodule takes storm surge data and astronomical tide data as input, performs element-wise superposition, and outputs a three-dimensional tidal level spatiotemporal field matrix. This matrix is ​​then compared with warning tidal levels issued by government departments to identify sections of the breakwater that have overflowed. In terms of memory allocation, the three spatiotemporal convolutional blocks of the encoder are allocated independent CUDA streams, the graph attention network branches are allocated independent memory partitions, and the gated fusion layer is allocated a shared memory buffer. The number of CUDA streams corresponds one-to-one with the encoder level. The size of the graph attention network memory partition is linearly related to the number of nodes in the topological importance mask matrix. Specifically, the product of the memory allocation and the number of nodes divided by the number of baseline nodes equals the baseline memory allocation. The number of baseline nodes and the baseline memory allocation are determined by taking the average of multiple rounds of experiments on the memory utilization rate on the standard test set, with no less than 20 rounds of experiments.

[0046] The steps for establishing the training dataset for the OSTF-Net model specifically include: collecting historical operational wave numerical model data from 2010 to 2024, including hourly fields of wave height, period, and wave direction, with spatial coverage of the entire globe or the target sea area, totaling no less than 50,000 time steps; extracting refined coastlines for each time step for land-sea labeling and separation, and forming an equidistant regular grid field through interpolation; manually labeling the wave height field according to the wave classification standard to form a 9-level wave classification labeling matrix; extracting isoline extreme points and arc segments, constructing an isoline map structure, and recording node coordinates, curvature, and arc length attributes; and dividing the dataset into a training set, a validation set, and a test set in an 8:1:1 ratio.

[0047] The specific steps for training the OSTF-Net model include: using the Adam optimizer, with an initial learning rate set to... The training consisted of 200 epochs with a batch size of 16. The loss function was a weighted sum of wave graded cross-entropy loss and shore push velocity mean square error loss, with a weight ratio of 3:1. The weight ratio was determined by conducting grid search experiments on the validation set to evaluate classification accuracy and velocity error, with an experimental step size of 0.5. Classification accuracy and velocity mean square error were evaluated on the validation set every 10 epochs. The learning rate decay was triggered when the validation set loss did not decrease for 20 consecutive epochs, with a decay coefficient of 0.5. The final performance was evaluated on the test set after training.

[0048] The OSTF-Net model uses a topological importance mask matrix as the spatial attention mask for the decoder, enabling the model to allocate more pixel generation capacity in high-gradient regions with dense contour lines. This improves the classification accuracy of marine hydrological features and the estimation accuracy of shore push velocity in high-gradient regions without increasing overall computational cost. The graph attention network's topological awareness of the contour map structure allows the model to capture non-local structural changes in the wave field, while the gated fusion layer ensures an adaptive trade-off between pixel-level and topological features. The final output marine hydrological feature description set has a more accurate semantic correspondence with actual sea conditions, providing more physically consistent input for subsequent video generation.

[0049] The weighted Voronoi topology partitioning-driven dynamic meteorological field adaptive rendering grid algorithm constructs a weighted Voronoi diagram using the pressure gradient magnitude and absolute vorticity of topological feature points such as cyclone centers, fronts, and extreme points as weights. The Voronoi cell boundaries serve as the rendering grid boundaries. Cells in high-weight regions are automatically subdivided into high-resolution rendering subgrids, while low-weight regions retain a coarse grid. When the temporal evolution of the meteorological field causes the feature point position to move by more than one grid point, incremental Voronoi repartitioning is triggered, reconstructing only the affected cells to avoid the computational overhead of global repartitioning. The target value of the ratio of rendering resolution in high-gradient regions to rendering resolution in background regions is determined through the following experiments: Pareto front experiments are conducted on 10–50 sets of typical typhoon field data, with rendering visual error and the increase in the total number of grid points as dual objectives. The ratio corresponding to the inflection point that minimizes the increase in the number of grid points while reducing visual error is taken, and the experimental iterations are no less than 30 rounds. The technical advantages of the weighted Voronoi topology partitioning-driven dynamic weather field adaptive rendering mesh algorithm are as follows: through the physical saliency-weighted automatic cell subdivision mechanism, rendering resources are concentrated in high-gradient regions such as the typhoon eyewall, while reducing redundant calculations in low-gradient background regions; the incremental re-partitioning strategy only updates locally affected cells, avoiding repeated calculations caused by global re-partitioning every frame, thereby ensuring rendering accuracy in high-gradient regions while reducing the overall rendering computational load and improving the real-time performance of video frame generation.

[0050] The TV-L1 regularized rendering pipeline refers to an image rendering pipeline that uses the total variation L1 norm as the regularization term. Its principle is to suppress high-frequency oscillation noise while preserving the strong gradient boundaries of the image. In high gradient regions, the rendering pipeline adaptively switches to the ENO-WENO interpolation format. The ENO-WENO interpolation format avoids introducing Gibbs oscillations by selecting the smoothest interpolation template from multiple candidate templates. A local gradient detection branch is embedded in the GPU shader. When the local gradient exceeds the threshold, a high-precision interpolation path is automatically triggered. The threshold is determined experimentally on 50 meteorological field samples containing typhoon eyewalls, with the frequency of oscillation stripes and the interpolation calculation time as dual objectives. The experimental iteration rounds are no less than 20 rounds. The technical effect of combining the TV-L1 regularized rendering pipeline with the ENO-WENO interpolation format is that TV-L1 regularization suppresses spurious oscillations during the interpolation process, while the ENO-WENO interpolation format preserves the physical information of strong gradient regions such as the eyewall of the typhoon. The combination of the two enables the rendering of video frames with high gradient fields to achieve a balance between visual accuracy and numerical stability, avoiding the loss of typhoon intensity information caused by gradient truncation processing.

[0051] The layered pipelined parallel architecture involves allocating vector field, scalar field, and contour rendering tasks to independent GPU stream multiprocessor groups, with each group executing in parallel to eliminate waiting overhead in serial rendering. CUDA graph technology eliminates the scheduling overhead of kernel function startup per frame by pre-recording the kernel function call graph and submitting it all at once. Inter-frame motion compensation differential coding borrows from the H.265 inter-frame prediction mechanism to perform motion estimation and differential coding on meteorological fields at adjacent time points, re-rendering only changed areas to reduce redundant computation. The ratio of the number of GPU stream multiprocessor groups to the computational load of the three types of rendering tasks (vector field, scalar field, and contour) was determined through performance experiments on 10–30 sets of meteorological field data at different resolutions. The experimental metric was single-frame processing latency, and the experiment iterated for no less than 15 rounds. The technical effects of the layered pipelined parallel architecture are: the parallel allocation of multi-layer rendering tasks breaks the serial bottleneck caused by data dependency; CUDA graph technology reduces kernel function scheduling overhead; and inter-frame differential coding reduces redundant rendering computation between adjacent frames. The synergy of these three technologies significantly reduces single-frame processing latency, meeting the frame rate requirements for real-time broadcasting.

[0052] The semantic anchor timeline refers to the sequence of key physical event time nodes pre-extracted from the warning text, including peak wind speed, extreme wave occurrence, and storm surge arrival. This is injected as a hard time constraint into the text-to-speech generation process to ensure that the speech output and video keyframes are aligned on the timeline. The CTC forced alignment algorithm establishes a bidirectional dynamic programming path between the audio and video frame sequences, calculates the optimal alignment mapping, and compresses the audio-visual timestamp error to the target range. The target error threshold is determined experimentally using subjective audio-visual desynchronization scores as an indicator on 20 sets of broadcast samples containing multiple semantic anchors, with at least 15 iterations. The technical effect of the CTC forced alignment algorithm is that by establishing bidirectional dynamic programming alignment at the audio and video frame levels, it overcomes the double error superposition caused by asynchronous sampling of the text-to-speech model output frame rate and video frame rate, ensuring that the broadcast content and video scene accurately correspond at the semantic event time nodes, and eliminating information transmission deviations caused by audio-visual desynchronization.

[0053] The marine hydrological feature description set refers to the collection of wave classification information, shore thrust velocity information, tide level information, and overtopping shoreline information expressed in natural language by the output of the OTF-Net model, which is used for the generation of subsequent text-based video prompts.

[0054] The textual prompts in the video refer to complete natural language descriptions generated by a large language model that conform to the input specifications of the textual video model after integrating elements such as visual style, scene geographic information, marine hydrological feature description set, anchor image, camera movement requirements, subtitle content, and overall atmosphere.

[0055] A memory pressure adjustment function is designed to dynamically adjust the number of attention heads in the graph attention network branches of the OSTF-Net model during inference. This function calculates a memory pressure index based on three factors: current GPU memory utilization, current number of input matrix nodes, and current batch size. The calculation method involves dividing each of the three factors by its mean on the standard test set and then taking a weighted sum. The weights are determined through experiments in 20 different load scenarios, with memory overflow rate and model output accuracy as dual objectives. The experiments are iterated for at least 10 rounds. When the memory pressure index... satisfy At that time, the number of heads to focus on remains at 4; when When, the number of heads to focus on is adjusted to 2; when At this time, the number of attention heads is adjusted to 1, and the size of the memory partition of the graph attention network branch is reduced proportionally as the number of attention heads decreases, to ensure that memory overflow does not occur during the inference process.

[0056] The Gibbs oscillation refers to the spurious high-frequency oscillation phenomenon that occurs near the discontinuity point when interpolating signals with discontinuities or high gradients, which manifests as stripe noise in video frames.

[0057] The ENO-WENO interpolation format, also known as the intrinsically non-oscillating-weighted intrinsically non-oscillating interpolation format, avoids introducing oscillation errors by selecting the template with the smoothest values ​​from multiple candidate difference templates for interpolation, thus avoiding the introduction of oscillation errors by the difference across gradient discontinuity regions.

[0058] The CUDA graph technology refers to the technique of pre-recording a series of GPU kernel function calls and their dependencies as a computation graph, and submitting them to the GPU for execution in one go as a graph unit, which can eliminate the CPU-GPU scheduling overhead of starting kernel functions one by one.

[0059] The CTC forced alignment algorithm, also known as the connectionist temporal classification forced alignment algorithm, establishes a monotonic dynamic programming path between two time series to find the mapping relationship that minimizes the alignment cost. In this scheme, it is used to establish frame-level alignment between audio frame sequences and video frame sequences.

[0060] The weighted Voronoi diagram, also known as the weighted Voronoi power diagram, adjusts the radius of influence of each feature point by the physical salience weight of each feature point, so that the cell area corresponding to the highly physical salience feature point is smaller and the local resolution is higher.

[0061] The term "overflowing embankment section" refers to the shoreline section whose tide level exceeds the warning value after calculating the spatiotemporal field matrix of the tide level and comparing it with the warning tide level value issued by the government.

[0062] The GPU streaming multiprocessor group refers to a group of computing resources in a GPU consisting of multiple streaming multiprocessors. Different groups can independently and concurrently execute different rendering tasks to achieve task-level parallelism.

[0063] Optionally, the present invention also provides a computer-based method for forming an AI-powered automatic rapid generation system for short videos of marine early warning information broadcasting. The computer is equipped with a readable storage medium that stores program instructions, which are used to execute the above-described method when the computer is run.

[0064] The specific implementation of step S01 is as follows: The system receives the video generation request input by the user in natural language and feeds the input text into a large language model for semantic understanding and structured decomposition. The large language model extracts elements such as visual style, scene geographic information, broadcast area range, broadcast date, anchor description, camera movement requirements, and overall atmosphere through contextual semantic parsing, and maps these elements into a set of structured request parameters in key-value pair form. This parameter set serves as the global input for subsequent steps, ensuring that subsequent modules operate within a unified semantic framework. The core function of this step is to transform unstructured natural language input into programmable structured data, providing a consistent semantic anchor for subsequent data acquisition and video generation.

[0065] The specific implementation of step S02 is as follows: Based on the target area information in the structured requirement parameter set, the system obtains weather forecast data for the corresponding area through an external meteorological interface, and simultaneously reads the hourly three-dimensional spatiotemporal matrix of sea wave height, period, wave direction, and storm surge increase from the operational numerical weather prediction system. The spatial resolution of the above matrix is ​​between 0.05° and 0.25°, and the time step is 1 hour. The three-dimensional spatiotemporal matrix is ​​input into the marine scene spatiotemporal feature decoupled perception model (OSTF-Net). OSTF-Net adopts an encoder-decoder backbone structure. The encoder consists of three levels of spatiotemporal convolutional blocks. Each level contains a two-dimensional spatial convolutional layer and a one-dimensional convolutional layer along the time axis, with output channels of 32, 64, and 128 respectively. The encoder output enters the graph attention network branch. This branch constructs a contour map structure with contour extreme points as nodes and contour arcs as edges. The topological importance score of each arc is calculated through a four-layer graph attention network to form a topological importance mask matrix. The decoder's three-stage deconvolutional blocks adaptively weight and fuse the encoder's skip connection features and topological importance mask matrix through a gated fusion layer at each stage. Finally, two parallel linear layers output a 9-level wave classification matrix and a wave shore thrust velocity matrix, respectively. The tidal level spatiotemporal field matrix is ​​processed by an independent tidal level calculation submodule. It element-wise superimposes the storm surge field and astronomical tide data, compares it with the warning tide levels issued by government departments, and identifies sections of the breakwater that have overflowed. During inference, a memory pressure adjustment function monitors the GPU memory usage, the number of input matrix nodes, and the batch size in real time, calculating the memory pressure index. ;when Maintain a head count of 4 at the same time. When adjusted to 2, The time is adjusted to 1, and the memory partition of the graph attention network branch is reduced proportionally with the number of attention heads to prevent memory overflow. After the above outputs are fused with the spatial relationships of interest points, a set of marine hydrological feature descriptions expressed in natural language is formed for subsequent video generation.

[0066] The specific implementation of step S03 is as follows: The marine hydrological feature description set is input into a dynamic meteorological field adaptive rendering grid algorithm driven by weighted Voronoi topology partitioning. The algorithm constructs a weighted Voronoi power graph using the pressure gradient magnitude and absolute vorticity of topological feature points such as cyclone centers, fronts, and extreme points as weights. The radius of the influence domain of each feature point shrinks as the weight increases, making the cell area corresponding to highly physical saliency feature points smaller and the local rendering resolution higher. Cells in high-weight regions are automatically subdivided into high-resolution rendering sub-grids, while low-weight regions retain coarse grids. The ratio of rendering resolution between high-gradient regions and background regions is determined through Pareto front experiments, with at least 30 iterations. When the temporal evolution of the meteorological field causes the position of feature points to move beyond the grid point movement threshold (1 grid point), incremental Voronoi repartitioning is triggered, and only the affected cells are reconstructed. After the adaptive grid is generated, the rendering pipeline uses TV-L1 regularization, with the total variation L1 norm as the regularization term, to suppress high-frequency oscillation noise while preserving strong gradient boundaries. The GPU shader embeds a local gradient detection branch. When the local gradient exceeds the gradient detection threshold, it automatically switches to the ENO-WENO interpolation format. It selects the template with the smallest smoothness indicator from multiple candidate difference templates for interpolation to avoid Gibbs oscillation. The gradient detection threshold is determined through a dual-objective experiment on 50 sets of meteorological field samples containing typhoon eyewalls. The experiment has no less than 20 iterations.

[0067] The specific implementation of step S04 is as follows: A layered pipelined parallel architecture is adopted, and the three types of tasks—vector field rendering, scalar field rendering, and contour rendering—are allocated to independent GPU streaming multiprocessor groups. Each group executes in parallel, eliminating the waiting overhead in serial rendering. CUDA graph technology pre-records kernel function calls and dependencies as a computation graph, submitting it to the GPU in one go as a graph unit, eliminating the CPU-GPU scheduling overhead of starting kernel functions frame by frame. Inter-frame motion compensation differential coding borrows from the inter-frame prediction mechanism, performing motion estimation and differential coding on the meteorological field at adjacent time points, re-rendering only the changed areas, reducing redundant computation. The ratio of the number of GPU streaming multiprocessor groups to the computational load of the three types of rendering tasks is determined through performance experiments on 10–30 sets of meteorological field data at different resolutions. The experimental metric is single-frame processing latency, and the experimental iteration rounds are no less than 15 rounds. The above three mechanisms work together to achieve real-time encoded output of video frame sequences.

[0068] The specific implementation of step S05 is as follows: Key physical event time nodes, such as peak wind speed, extreme wave occurrence, and storm surge arrival, are pre-extracted from the broadcast text to construct a semantic anchor timeline. These time nodes are injected as hard constraints into the text-to-speech generation process to ensure that the speech output aligns with the video keyframes at the corresponding time nodes. The CTC forced alignment algorithm (connectionist temporal classification forced alignment algorithm) establishes a bidirectional dynamic programming path between the audio frame sequence and the video frame sequence, solving for the mapping relationship that minimizes the alignment cost, compressing the audio-visual timestamp error to within 40ms. The error threshold is determined experimentally using subjective audio-visual desynchronization scores as an indicator on 20 sets of broadcast samples containing multiple semantic anchors, with at least 15 iterations.

[0069] The specific implementation of step S06 is as follows: The visual style, scene geographic information, anchor image, camera movement requirements, and overall atmosphere from the structured requirement parameter set, along with the marine hydrological feature description set, rendering frame sequence information, audio alignment sequence information, and subtitle content, are input into the large language model. The large language model performs semantic fusion on the above multi-source structured information to generate a complete natural language description text that conforms to the input specifications of the text-based video model, i.e., text-based video prompt words. After the text-based video prompt words are input into the text-based video model, the model automatically generates a short broadcast video, completing the entire generation process.

[0070] It should be noted that the key technologies of this invention include: the weighted Voronoi topology partitioning adaptive rendering mesh technology, through a physically saliency-weighted cell subdivision mechanism, adaptively increases the interpolation node density in high-gradient regions, thereby reducing the probability of the differential template crossing discontinuities in principle; the ENO-WENO interpolation format, through a template selection mechanism driven by smoothness indicators, eliminates Gibbs oscillations across discontinuities from the source; and the TV-L1 regularized rendering pipeline applies a global penalty to residual high-frequency oscillations with the L1 norm, forming a dual anti-oscillation guarantee with the ENO-WENO interpolation format. The synergistic effect of the three technologies is as follows: the adaptive mesh reduces the probability of inter-discontinuity oscillations, the ENO-WENO interpolation format eliminates existing inter-discontinuity oscillations, and TV-L1 regularization further suppresses high-frequency components in the interpolation residuals. Together, these three technologies achieve a balance between visual accuracy and numerical stability in the rendering of video frames of high-gradient meteorological fields, enabling key physical information such as typhoon intensity to be accurately presented in the broadcast short videos. Meanwhile, the topological importance mask matrix mechanism and memory pressure adjustment function of OSTF-Net ensure the stability of the inference process, and the CTC forced alignment algorithm ensures the accurate temporal correspondence between audio-visual semantic events.

[0071] It should be noted that in the scenario of automated video generation for marine early warning information, when the broadcast video needs to simultaneously cover multiple semantic event time points (such as peak wind speed, extreme wave values, and storm surge arrival), and the text-to-speech model and video rendering module use asynchronous sampling frame rates, a systematic time deviation will occur between the speech output and the video keyframes, causing audio-visual desynchronization between the broadcast content and the video scene at key physical event moments. The reason for this technical problem is that the output frame rate of the text-to-speech model and the frame rate of the video rendering are usually inconsistent, and both have independent sampling errors. When these two errors are superimposed, the audio-visual deviation at key physical event moments will exceed the acceptable range of human hearing and vision. Furthermore, as the number of semantic events increases, the cumulative effect of the deviation further intensifies. Traditional methods based on fixed timestamp alignment cannot adapt to the combined effects of speech rate changes and frame rate jitter. The usual solution to this technical problem is to perform post-processing alignment after video generation, that is, to correct the audio-visual desynchronization through manual or automatic time offset adjustments. However, this method can only apply a fixed offset to the entire video, and cannot perform independent frame-level alignment for each semantic event time point. When the audio-visual deviations of different semantic events are different, adjusting the fixed offset will align some events while increasing the deviation of others, making it difficult to achieve accurate global frame-level alignment. This invention effectively solves this technical problem by pre-extracting key physical event time points from the broadcast text, constructing a semantic anchor time axis, and injecting these time points as hard constraints into the text-to-speech generation process, thus imposing time constraints on key event moments during the speech output generation stage. The CTC forced alignment algorithm establishes a bidirectional dynamic programming path between the audio frame sequence and the video frame sequence, independently solving for the optimal alignment mapping for each semantic anchor point, thereby achieving accurate frame-level alignment for each key physical event moment. This eliminates the cumulative deviation caused by the superposition of double errors from asynchronous sampling, ensuring that the broadcast content and video scene correspond precisely at the semantic event time points.

[0072] Specifically, the principle of this invention is:

[0073] The fundamental reason why this invention can solve the above-mentioned technical problems is that uniformly rendered meshes are not sensitive to gradient intensity, while the cell area of ​​a weighted Voronoi diagram is inversely proportional to the physical saliency weight of feature points. Cells in high-weight regions are smaller, resulting in higher local rendering resolution. This makes interpolation nodes more dense in high-gradient regions, thereby reducing the gradient change amplitude between adjacent nodes and lowering the probability of the difference template crossing discontinuities.

[0074] Traditional uniform interpolation schemes use fixed difference templates in high-gradient regions, inevitably introducing Gibbs oscillations when the template crosses gradient discontinuities. The ENO-WENO interpolation scheme, however, evaluates smoothness indicators among multiple candidate difference templates for each interpolation point, selecting the template with the smoothest numerical value for interpolation, thus eliminating the use of templates that cross discontinuities. This selection process is coupled with local gradient strength, ensuring that in regions with strong discontinuities such as the typhoon eyewall, the interpolation template remains on the same side as the discontinuity, avoiding spurious oscillations.

[0075] The TV-L1 regularization term serves as a global constraint in the rendering pipeline. Its L1 norm is more sensitive to isolated high-frequency oscillations than the L2 norm, and it can penalize oscillatory components without blurring gradient boundaries, thus forming a double anti-oscillation guarantee with the ENO-WENO interpolation format.

[0076] Regarding the temporal evolution of the rendered mesh, when the position of the meteorological field feature point moves beyond a set number of grid points, this invention triggers incremental Voronoi re-division, reconstructing only the affected cells, ensuring the dynamic consistency between the rendered mesh and the physical field, and avoiding the local gradient region from falling back into the low-resolution mesh due to mesh lag.

[0077] Furthermore, this invention embeds a local gradient detection branch into the GPU shader. When a local gradient exceeds a threshold, it automatically switches to a high-precision interpolation path, ensuring that the anti-oscillation mechanism is activated only in necessary regions. This guarantees rendering accuracy in high-gradient areas while maintaining computational efficiency in low-gradient areas. These mechanisms logically form a closed loop: adaptive meshing ensures local resolution, the ENO-WENO interpolation format eliminates intermittent oscillations, and TV-L1 regularization suppresses residual high-frequency noise, collectively ensuring the accurate representation of critical physical information such as typhoon intensity in video frames.

[0078] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0079] The specific implementation method of step S01 is as follows.

[0080] The system receives video generation requests from users in natural language, performs semantic understanding and element decomposition using a large language model, and outputs a structured set of request parameters. This set includes elements such as visual style, scene geographic information, broadcast date, anchor image, camera movement requirements, and overall atmosphere, stored as key-value pairs for later use.

[0081] The specific implementation method of step S02 is as follows.

[0082] Based on a structured set of required parameters, weather forecast data for the target area is acquired through an external meteorological interface, and operational ocean wave numerical model output is read simultaneously to form a three-dimensional spatiotemporal matrix. The input matrix shape is... ,in This represents the number of grid points in the longitude direction. The number of grid points in the latitudinal direction. This represents the total number of time steps, with a spatial resolution between 0.05° and 0.25°, and a time step size of 1 hour. The matrix elements... Including the The first longitude grid point, the first The first latitudinal grid point, the first Wave height at each time step (unit ),cycle (unit ) and wave direction (unit ), and then input into the ocean scene spatiotemporal feature decoupling perception model (abbreviated as ), after being stitched together according to the channel dimension. ).

[0083] The encoder consists of three levels of spatiotemporal convolutional blocks, the first being... The output of the spatial convolution is:

[0084] ;

[0085] In the formula, For the first Level 3×3 spatial convolution kernel weight matrix, For the first Spatial convolution bias scalar, For the first Level output feature map, ,when hour That is, the input matrix Feature map after channel splicing To correct the linear unit activation function, the number of output channels The numbers are 32, 64, and 128 respectively. and This represents the offset index within the spatial convolution kernel. The temporal convolution output is:

[0086] ;

[0087] In the formula, For the first A one-dimensional convolution kernel weight vector along the time axis. For the first Time-limited convolution bias scalar, This is the offset index within the temporal convolution kernel. For the first The final output feature map of the spatiotemporal convolution block is obtained by the first-order spatiotemporal convolution.

[0088] The encoder output is fed into the graph attention network branch. The set of nodes is based on the extreme points of the contour lines. The contour arc segments form the edge set. curvature With arc length For the edge The attributes of the contour map are used to construct the contour map structure. ,in The total number of nodes. The node update formula for a layered graph attention network is:

[0089] ;

[0090] In the formula, For the number of attention heads, For the first Layer nodes The feature vector has a dimension of 64. , For nodes The set of neighboring nodes, For activation function, For the first Layer The learnable weight matrix for each attention head, and the attention coefficients. For the first Layer Each attention node For neighboring nodes The normalized attention weights are determined by the edge attributes. The weighted and normalized values ​​are then calculated. The graph attention network has four layers and ultimately outputs a topological importance mask matrix. Each element takes a value between 0 and 1.

[0091] Decoder The input to the gated fusion layer is the encoder skip connection feature. With topological importance mask matrix The gating fusion formula is:

[0092] ;

[0093] In the formula, For the first Level gate vector, For logical functions, For element-wise multiplication, for Upsampled to the first Feature map corresponding to the spatial resolution of level 1 and The first Learnable weight matrix and bias vector of the gated fusion layer For the first Feature maps after multi-level fusion. The decoder output channels are 64, 32, and 16 respectively, and are finally mapped to a 9-level wave classification matrix through two parallel linear layers. With the shore thrust matrix , The unit is .

[0094] The tidal space-time field matrix is ​​calculated by independent sub-modules, and the superposition formula is:

[0095] ;

[0096] In the formula, For the first The first longitude grid point, the first The first latitudinal grid point, the first Storm surge increase at each time step (units) ), For the corresponding astronomical tide level (unit) ), Synthetic tidal level (unit) (and the government-issued warning tide levels) (unit ) Comparison, when It was identified as a section of the embankment.

[0097] The loss function during the training phase is:

[0098] ;

[0099] In the formula, For wave-level cross-entropy loss, The mean square error loss of the shore thrust velocity, A wave classification and labeling matrix, A matrix of shore thrust velocity labels (units) The weight ratio of 3:1 was determined by performing a grid search on the validation set to evaluate classification accuracy and speed error, with a search step size of 0.5. The Adam optimizer was used, with an initial learning rate of... The training process is conducted for 200 rounds with a batch size of 16. The learning rate is multiplied by a decay factor of 0.5 when the validation set loss does not decrease for 20 consecutive rounds.

[0100] Regarding memory allocation, the memory allocation of the graph attention network branch... (unit ) and number of nodes satisfy:

[0101] ;

[0102] In the formula, Base memory allocation (units) ), The baseline number of nodes is determined by averaging the results of at least 20 rounds of memory utilization experiments on a standard test set.

[0103] In the memory pressure adjustment function, the memory pressure index It is a dimensionless exponent, and the calculation formula is:

[0104] ;

[0105] In the formula, This represents the current GPU memory usage. This represents the average video memory usage on the standard test set. This is the current batch size. The average batch size on the standard test set. , , For the corresponding weight coefficients and satisfying The method was determined through experiments under 20 different load scenarios, with memory overflow rate and model output accuracy as dual objectives, and the number of iterations was no less than 10 rounds. Maintain a head count of 4 at the same time. When adjusted to 2, When the time is adjusted to 1, the size of the GPU memory partition of the graph attention network decreases proportionally with the number of attention heads.

[0106] The specific implementation method of step S03 is as follows.

[0107] Pressure gradient model based on topological features such as cyclone center, front, and extreme points (unit ) and absolute value of vorticity (unit Using weights , construct a weighted Voronoi diagram. Feature points The influence domain weights are:

[0108] ;

[0109] In the formula, Normalized reference value for pressure gradient mode (unit) ), Normalized reference value for absolute vorticity (units) ), and For the coefficients to be determined, satisfying It was determined through Pareto front experiments that The weights are dimensionless. Spatial points in a weighted power graph. Belonging to feature points The conditions for the cell unit are:

[0110] ;

[0111] In the formula, For spatial points With feature points Euclidean distance between them (units) ), For spatial points With feature points Euclidean distance between them (units) ),because It is a dimensionless quantity and the unit of distance squared is , here After dimensional consistency processing, it has the same dimensions as the square of the distance. The processing method is to... Multiply by the square of the reference length (unit The empirical value is the square of the grid spacing, which means that the actual weighting factor used in the comparison is... In high-weight regions, cells are automatically subdivided into high-resolution rendering sub-mesh, while low-weight regions retain a coarse mesh. When the temporal evolution of the meteorological field causes feature point positions to shift by more than one grid point, incremental re-subdivision is triggered, reconstructing only the affected cells. The ratio of rendering resolution between high-gradient regions and background regions is determined through Pareto front experiments on 10 to 50 sets of typical typhoon field data, with rendering visual error and grid point increment as dual objectives. The ratio corresponding to the inflection point is taken, and the iteration rounds are no less than 30. High-gradient regions employ a fully variational L1 regularized rendering pipeline, combined with an essentially non-oscillating weighted essentially non-oscillating (ENO-WENO) interpolation format, embedding a local gradient detection branch in the GPU shader. Exceeding the threshold Automatically triggers high-precision interpolation path, Experiments were conducted on 50 sets of meteorological field samples containing typhoon eyewalls, with the dual objectives of the frequency of oscillating fringes and the interpolation calculation time. The number of iterations was determined to be no less than 20 rounds. This represents the current scalar value of the meteorological field.

[0112] The specific implementation method of step S04 is as follows.

[0113] Based on a hierarchical pipelined parallel architecture, vector fields, scalar fields, and contour rendering are distributed to independent GPU streaming multiprocessor groups for parallel execution. A unified computing device architecture graph technique is used to pre-record kernel function call graphs and submit them all at once, eliminating frame-by-frame scheduling overhead. Inter-frame motion compensation differential coding borrows from the H.265 inter-frame prediction mechanism, estimating motion in the meteorological field at adjacent time points and re-rendering only the changed areas, completing real-time encoding and output of the video frame sequence. The ratio of the number of GPU streaming multiprocessor groups to the computational load of the three types of rendering tasks was determined through performance experiments on 10 to 30 sets of meteorological field data at different resolutions. The experimental metric was single-frame processing latency, with at least 15 iteration rounds.

[0114] The specific implementation method of step S05 is as follows.

[0115] Key physical event time nodes, such as peak wind speed, extreme wave value, and storm surge arrival time, are extracted from the broadcast text to construct a semantic anchor timeline. ,in The total number of semantic anchors. For the first The time nodes of semantic anchor points are injected as hard constraints into the text-to-speech generation process. A connectionist temporal classification forced alignment algorithm is adopted in the audio frame sequence. With video frame sequence Establish a bidirectional dynamic programming alignment path between them, where The total number of audio frames. The total number of video frames. For the first One audio frame, For the first Given a set of video frames, the objective is to find the optimal alignment mapping. Minimize alignment cost:

[0116] ;

[0117] In the formula, For the first The timestamps (in units) corresponding to each semantic anchor point in the audio frame sequence. ), For the first The timestamps (in units) corresponding to each semantic anchor point in the video frame sequence. ), Normalized reference duration (unit) The experience value is 1 second. A monotonic alignment mapping function from video frame timestamps to audio frame timestamps is used to compress the audio-visual timestamp error to within 40ms. The target error threshold is determined experimentally using subjective audio-visual desynchronization scores as an indicator on 20 sets of broadcast samples containing multiple semantic anchors, with at least 15 iterations.

[0118] The specific implementation method of step S06 is as follows.

[0119] The structured requirement parameter set, marine hydrological feature description set, rendering frame sequence, audio alignment sequence and subtitle content are input into the large language model. By integrating elements such as visual style, scene geographic information, anchor image, camera movement requirements and overall atmosphere, a complete natural language description prompt word that conforms to the input specification of the Wensheng video model is generated. Inputting this word into the Wensheng video model automatically generates a short broadcast video.

[0120] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the technical effect of the invention, the technicians selected a typical typhoon process as a test scenario, used the technical solution of the invention to automatically generate short videos of the marine early warning information of the typhoon process, and recorded and analyzed the output of each key module in detail during the generation process.

[0121] Technicians first input video generation requirements using natural language, describing the visual style as realistic weather animation, the scene as a sea area in the Northwest Pacific, the broadcast area as coastal regions affected by the typhoon, the broadcast date as 48 hours before to 12 hours after the typhoon's landfall, the anchor's appearance as a standard news anchor, and the camera movement requirement as gradually advancing towards the typhoon's center. The large language model performs semantic parsing on the above input, extracting a structured set of requirement parameters that includes visual style, scene geographic information, broadcast date, anchor appearance, camera movement requirements, and overall atmosphere, which serves as the global input for subsequent modules.

[0122] The system acquires weather forecast data for the target area through an external meteorological interface and reads hourly three-dimensional spatiotemporal matrices of sea wave height, period, wave direction, and storm surge increase from an operational numerical weather prediction system. The matrices have a resolution of 0.1° and a time step of 1 hour, covering 60 time steps and totaling over 50,000 grid points. These three-dimensional spatiotemporal matrices are then input into the OSTF-Net model, as shown below. Figure 2 As shown, the OSTF-Net encoder extracts spatial and temporal features sequentially using three levels of spatiotemporal convolutional blocks. The graph attention network branch calculates the topological importance mask matrix based on the contour map structure. The decoder outputs a 9-level wave classification matrix and a shore thrust velocity matrix through a gated fusion layer. The tide calculation submodule overlays storm surge field and astronomical tide data element by element to identify sections of the breakwater that have overflowed. In this test, the GPU memory utilization rate was 72% during inference, the number of input matrix nodes was 1.1 times the standard value, the batch size was 16, and the memory pressure index was high. After weighted summation, the values ​​fall within the range of [0.6, 0.85). The number of attention heads in the graph attention network branches is automatically adjusted to 2, and no memory overflow occurs. The model outputs a set of marine hydrological feature descriptions as shown in Table 1.

[0123] Table 1. Example of key parameters in the marine hydrological feature description set.

[0124]

[0125] Adaptive rendering stage, such as Figure 3 As shown, the weighted Voronoi topological partitioning algorithm constructs a weighted Voronoi power graph using the magnitude of the typhoon's central pressure gradient and the absolute value of its vorticity as weights. The cell area of ​​the typhoon eyewall region is approximately 1 / 8 of that of the background region, and the rendered submesh is automatically subdivided at the eyewall. During the typhoon's movement, when a feature point moves more than one grid point, incremental Voronoi repartitioning is triggered, reconstructing only the affected cells, ensuring the rendered mesh remains dynamically consistent with the physical field. When the GPU shader detects that the local gradient in the typhoon eyewall region exceeds the gradient detection threshold (experimentally determined to be approximately 3 times the average gradient of the background region), it automatically activates the ENO-WENO interpolation format, selecting the candidate template with the smallest smoothness indicator to complete the interpolation, effectively avoiding Gibbs oscillations. The TV-L1 regularized rendering pipeline applies a total variation L1 norm constraint to the entire frame, further suppressing residual high-frequency oscillation noise, thus preserving the physical information of strong gradient regions such as the typhoon eyewall.

[0126] During the video frame encoding stage, the layered pipeline parallel architecture distributes the rendering of vector fields, scalar fields, and contour lines to three independent GPU stream multiprocessor groups for parallel execution. CUDA graph technology pre-records the kernel function call graph and submits it all at once. Inter-frame motion compensation differential encoding only re-renders the changed areas in adjacent frames. The single-frame processing latency is significantly lower than that of the serial rendering process, meeting the frame rate requirements for real-time broadcasting. The latency distribution of each rendering task is shown in Table 2.

[0127] Table 2 Distribution of Single-Frame Processing Latency for Each Rendering Task

[0128]

[0129] In the audio-visual alignment stage, three semantic anchor points are extracted from the broadcast text: peak wind speed (6 hours before landfall), extreme wave value (4 hours before landfall), and storm surge arrival (2 hours before landfall). A semantic anchor point timeline is constructed and injected into the text-to-speech generation process. The CTC forced alignment algorithm establishes a bidirectional dynamic programming path between the audio and video frame sequences, independently solving for the optimal alignment mapping for each of the three semantic anchor points, such as... Figure 4 As shown, the distribution of audio-visual timestamp errors indicates that the audio-visual timestamp errors of the three semantic anchors are all compressed to within 40ms, and the broadcast content and video scene accurately correspond at the moment of key physical events.

[0130] In the final generation stage, the large language model integrates the structured requirement parameter set, the marine hydrological feature description set, the rendered frame sequence information, the audio alignment sequence information, and the subtitle content to generate complete text-based video prompts. These prompts are then input into the text-based video model, which automatically outputs a short video for broadcast. The distribution of each element of the text-based video prompts is shown in Table 3.

[0131] Table 3. Composition of Video Prompt Elements

[0132]

[0133] like Figure 5 As shown, the curves of the wave classification accuracy and mean square error of shore push velocity of the OSTF-Net model on the test set with the number of training rounds indicate that the model converges after 200 training rounds, the classification accuracy on the validation set is stable, the mean square error of velocity continues to decrease, and the topological importance mask matrix mechanism makes the classification accuracy of high gradient regions significantly better than the baseline model without graph attention network branches.

[0134] like Figure 6 As shown in the diagram, the cell distribution of the weighted Voronoi rendering mesh in the eyewall region of the typhoon indicates that the cell density in the high gradient region is significantly higher than that in the background region, and the effect of concentrated allocation of rendering resources is intuitive.

[0135] Compared with traditional methods, this invention achieves the following advancements at the technical principle level: Traditional uniform rendering meshes cannot perceive the local gradient intensity of the meteorological field, while the weighted Voronoi topology subdivision mechanism, through physically significant weighted cell subdivision, inherently couples the rendering resolution with the physical field gradient distribution, fundamentally solving the problem of insufficient resolution of uniform meshes in high gradient regions; the smoothness indicator-driven template selection mechanism of the ENO-WENO interpolation format enables the interpolation process to have discontinuity perception capability, eliminating Gibbs oscillations from the source, rather than relying on the passive suppression method of post-processing filtering; the bidirectional dynamic programming alignment mechanism of the CTC forced alignment algorithm enables independent optimization of audio-visual alignment for each semantic anchor point, overcoming the fundamental defect that fixed offset post-processing alignment cannot adapt to independent deviations of multiple events.

[0136] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0137] Table 4. Variable Explanation Table (Part 1)

[0138]

[0139] Table 5. Variable Explanation Table (Part Two)

[0140]

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for AI-automated and rapid generation of short videos for marine early warning information broadcasting, characterized in that, Includes the following steps: The system receives video generation requests from users, understands and decomposes these requests using a large language model, and extracts a structured set of request parameters. Read the three-dimensional spatiotemporal matrix of numerical patterns of operational ocean wave height, period, wave direction and storm surge, input the decoupled perception model of ocean scene spatiotemporal features, complete the calculation of ocean wave classification matrix, shore thrust velocity matrix and tide level spatiotemporal field matrix, and integrate the spatial relationship of points of interest to output a set of ocean hydrological feature descriptions. The set of marine hydrological features is input into a dynamic meteorological field adaptive rendering mesh algorithm driven by weighted Voronoi topology partitioning. A weighted Voronoi diagram is constructed with the physical salience of cyclone center and extreme point as weights to generate an adaptive non-uniform rendering mesh. Combined with TV-L1 regularized rendering pipeline and ENO-WENO interpolation format, anti-oscillation rendering of high gradient regions is completed in GPU shader. Based on a hierarchical pipelined parallel architecture, vector fields, scalar fields, and contour rendering are distributed to independent GPU streaming multiprocessor groups. CUDA graph technology is used to eliminate kernel function startup overhead, and inter-frame motion compensation differential coding is introduced to complete the real-time encoding output of video frame sequences. Key physical event time nodes are extracted from the broadcast text, a semantic anchor time axis is constructed, and time nodes are injected as hard constraints into the text-to-speech generation process. The CTC forced alignment algorithm is used to establish bidirectional dynamic programming alignment at the audio frame level and the video frame level, compressing the audio-visual timestamp error to within the error threshold. The structured requirement parameter set, marine hydrological feature description set, rendering frame sequence, audio alignment sequence and subtitle content are fused together by a large language model to generate complete text-based video prompts. Inputting these prompts into the text-based video model will automatically generate a short video for broadcasting. The marine scene spatiotemporal feature decoupling perception model adopts an encoder-decoder backbone. The encoder consists of three levels of spatiotemporal convolutional blocks. Each level includes a two-dimensional spatial convolutional layer and a one-dimensional convolutional layer along the time axis. The encoder output is fed into a graph attention network branch. The input of each level of the decoder is composed of the encoder skip connection output and the topological importance mask matrix weighted by a gated fusion layer. The graph attention network branch constructs a contour map structure with contour extreme points as nodes, contour arc segments as edges, and curvature and arc length as edge attributes. It passes messages through the graph attention network, calculates the topological importance score of each contour arc segment, and forms a topological importance mask matrix. The weighted Voronoi topology partitioning-driven dynamic meteorological field adaptive rendering grid algorithm constructs a weighted Voronoi map with pressure gradient modulus and absolute vorticity as weights. Cells in high-weight regions are automatically subdivided into high-resolution rendering sub-grids, while low-weight regions retain coarse grids. Incremental Voronoi repartitioning is triggered when the movement of feature point positions exceeds the grid movement threshold. The TV-L1 regularized rendering pipeline uses the total variation L1 norm as the regularization term, which suppresses high-frequency oscillation noise while preserving the strong gradient boundaries of the image. A local gradient detection branch is embedded in the GPU shader, and when the local gradient exceeds the gradient detection threshold, the high-precision interpolation path of the ENO-WENO interpolation format is automatically triggered. The ENO-WENO interpolation format avoids introducing Gibbs oscillations by selecting the smoothest interpolation template from multiple candidate templates. Gibbs oscillations refer to the spurious high-frequency oscillations that occur near discontinuities when interpolating signals with discontinuities or high gradients.

2. The method according to claim 1, characterized in that, The structured requirement parameter set refers to a set of structured data that includes elements such as visual style, scene geographic information, broadcast date, anchor image, camera movement requirements, and overall atmosphere, formed after the large language model parses the user's natural language input.

3. The method according to claim 2, characterized in that, The marine scene spatiotemporal feature decoupling perception model has a memory pressure adjustment function. It calculates the memory pressure index based on the current GPU memory usage, the current number of input matrix nodes, and the current batch size. It dynamically adjusts the number of attention heads in the graph attention network branch according to the memory pressure index. The memory partition size of the graph attention network branch decreases proportionally as the number of attention heads decreases.

4. The method according to claim 3, characterized in that, The hierarchical pipelined parallel architecture distributes vector fields, scalar fields, and contour rendering to independent GPU stream multiprocessor groups for parallel execution. CUDA graph technology eliminates the scheduling overhead of starting kernel functions frame by frame by pre-recording the kernel function call graph and submitting it all at once.

5. The method according to claim 4, characterized in that, The inter-frame motion compensation differential coding borrows from the inter-frame prediction mechanism to perform motion estimation and differential coding on the meteorological field at adjacent time points, and only re-renders the changed areas.

Images