A neural network-based 3D rendering LOD decision tree adaptive threshold generation method

Abstract

This invention discloses a neural network-based adaptive threshold generation method for LOD (Level of Detail) in 3D rendering. Addressing the pain points of poor adaptability of traditional fixed LOD thresholds and insufficient cross-device compatibility of existing dynamic solutions, this method constructs a multi-dimensional parameter system combining spatial and hardware parameters. It outputs dynamic thresholds through neural network pre-training, combines this with an LOD decision tree to determine the target LOD level, and iterates based on frame rate and image quality as dual indicators, forming a closed-loop optimization. The process is as follows: pre-training a neural network model to output hardware target values ​​and dynamic LOD thresholds; collecting spatial and hardware parameters during 3D rendering; fusing parameters and dynamic thresholds in the LOD decision tree to output the level; detecting dual indicators after rendering; and optimizing the model based on the feedback results. This method achieves a balance between visual quality and performance, increasing the average frame rate by 15%-20%, reducing GPU time by 25%-30%, and reducing stuttering by over 90%. It is compatible with workstations, PCs, and mobile devices, and is suitable for scenarios such as real-world 3D, VR / AR, and digital twins.

Description

Technical Field

[0001] This invention relates to the fields of computer graphics and 3D rendering technology, specifically to a neural network-based adaptive threshold generation method for 3D rendering LOD decision trees. It is applicable to various scenarios such as real-world 3D rendering, VR / AR, digital twins, GIS maps, and games, and can be adapted to different computing power devices such as high-end workstations, PCs, and mobile devices. Background Technology

[0002] In the realm of realistic 3D, Level of Detail (LOD) technology is a core optimization method for balancing rendering performance and visual effects. Its core functions are reflected in four aspects: reducing hardware load, optimizing memory management, maintaining visual continuity, and supporting multi-scene adaptation. The LOD decision tree, as the core logical framework for implementing the LOD strategy, uses progressive, multi-dimensional conditional combinations to match the optimal level of detail for each 3D object in real time, minimizing rendering resource consumption while ensuring a good visual experience.

[0003] Traditional LOD decision trees generally rely on preset fixed thresholds to classify LOD levels, which has significant technical shortcomings: First, poor scene adaptability. The same threshold can lead to unreasonable LOD switching timing in different scenes such as open fields and dense urban areas, causing computing power overload or loss of details. Second, visual and performance imbalance. If the threshold is too strict, the image quality will be reduced. If the threshold is too loose, it will cause unnecessary high-load rendering, causing frame rate fluctuations or even stuttering.

[0004] While existing dynamic LOD technology has achieved dynamic adjustment of thresholds, it still has unresolved core pain points: insufficient cross-device compatibility, with the computing power gap between high-end devices and mobile devices reaching tens of times, and existing solutions lacking a precise adaptation mechanism; weak processing capabilities for complex models, making it difficult to balance simplification efficiency with detail preservation, and low degree of automated adaptation; poor real-time performance in dynamic interactive scenarios, with delayed response to threshold adjustments, easily resulting in visual defects such as stuttering and screen flickering.

[0005] Related research has evolved from fixed thresholds to dynamic multi-feature thresholds, but three major gaps remain: incomplete decision features (some solutions focus only on spatial features, neglecting key dimensions such as screen occupancy and shader complexity); lack of resource adaptation (hardware computing power and storage resources are not incorporated into the threshold generation logic, resulting in poor cross-device adaptability); and non-intelligent optimization logic (relying heavily on manually preset linear rules, lacking data-driven self-learning and self-iterative capabilities). Therefore, there is an urgent need to construct an adaptive LOD threshold generation mechanism that integrates multi-dimensional features, deeply adapts to hardware resources, and is based on machine learning. Summary of the Invention

[0006] This invention aims to solve four key problems existing in current 3D rendering LOD technology:

[0007] 1. Traditional LOD technology uses distance as the sole decision-making dimension, ignoring core factors such as device performance and scene complexity. The threshold parameters are fixed, lacking dynamic adaptation capabilities and having poor rule reusability, resulting in an imbalance between vision and performance.

[0008] 2. Existing dynamic LOD technology lacks cross-device compatibility. The differences in hardware resources between high-end devices and mobile devices cannot be accurately adapted, which can easily lead to problems such as excessive video memory usage and loading delays.

[0009] 3. When processing complex models, it is difficult to balance simplification efficiency with detail preservation. In dynamic interactive scenarios, the response to threshold adjustment is lagging, causing fluctuations in computing power and visual defects.

[0010] 4. Existing solutions lack data-driven self-learning and closed-loop optimization mechanisms, rely on manual debugging, and have high adaptation costs and limited accuracy.

[0011] To address the aforementioned technical problems, this invention provides an adaptive threshold generation method for 3D rendering LOD decision trees based on neural networks. The specific technical solution is as follows:

[0012] Neural network model pre-training: A fully connected neural network model is constructed and trained based on historical sample data across devices and scenes. The historical sample data includes multi-dimensional sample parameters, corresponding optimal LOD level, resource utilization, and visual effect score. The multi-dimensional sample parameters cover nine categories of parameters: camera distance, screen ratio, viewing angle, model visibility, real-time frame rate, CPU utilization, GPU utilization, video memory usage, and memory usage. The neural network model takes the nine categories of multi-dimensional sample parameters as input, and after non-linear feature mapping in the hidden layer, outputs four types of hardware utilization target values ​​(CPU utilization target value, GPU utilization target value, video memory usage target value, and memory usage target value) and four types of LOD dynamic thresholds (LOD distance threshold, LOD screen ratio threshold, shader LOD threshold, and LOD global bias adjustment threshold).

[0013] The innovation of this technical solution lies in its breakthrough from the static adjustment mode of traditional LOD technology that relies on fixed experience thresholds, and its innovative construction of a dynamic intelligent decision-making system based on deep learning.

[0014] In existing technologies, LOD strategies are typically implemented by developers setting fixed switching thresholds based on specific hardware and scenarios. This approach not only requires extensive manual tuning but also cannot adapt to dynamic changes in device performance, user perspective, and scene complexity. This often results in an inability to achieve the optimal balance between ensuring visual quality and controlling resource consumption, easily leading to issues such as screen stuttering or resource waste.

[0015] The core innovation of this solution lies in its proposal of a novel "data-driven" dynamic LOD adjustment paradigm. Specifically, instead of simply classifying and predicting LOD levels, it uses a carefully designed neural network model to directly learn and output core decision parameters to guide LOD adjustment: four types of hardware utilization target values ​​and four types of dynamic LOD thresholds. This process deeply integrates and intelligently models previously isolated resource monitoring data with LOD adjustment rules.

[0016] By pre-training the model with massive amounts of historical sample data across devices and scenarios, the model can learn the potential mapping relationship between resource consumption and visual effects in complex scenarios. During actual operation, the model can dynamically calculate the optimal hardware resource target and LOD switching threshold based on the multi-dimensional state parameters collected in real time. This dynamic generation rather than static preset mechanism enables the system to make optimal decisions for each unique moment of operation, thereby achieving a global dynamic balance between visual quality and performance overhead.

[0017] This technical solution, by introducing deep learning technology, transforms LOD (Level of Rendering) from a static process that relies on human experience into a dynamic intelligent process that can learn autonomously and make decisions in real time. This significantly improves the adaptability and overall performance of the 3D rendering system in complex and ever-changing environments, demonstrating outstanding substantive features and remarkable progress.

[0018] The neural network model has 32 neurons in its hidden layer and uses the ReLU activation function to achieve nonlinear feature mapping. The neural network model also includes a weight matrix W1 from the input layer to the hidden layer, a weight matrix W2 from the hidden layer to the output layer, a hidden layer bias vector b1, and an output layer bias vector b2.

[0019] The training process uses a weighted mixed loss function L=α⋅L hardware +β⋅L fps +γ⋅L lod The weighting ratio α:β:γ = 0.5:0.3:0.2, where L hardware L represents the hardware utilization loss (using the mean square error between the predicted value and the preset target value). fps For frame rate loss (only penalizing cases where the frame rate is below the target value), L lod The reasonableness loss of the LOD threshold (the constraint threshold is within the preset boundary).

[0020] Real-time acquisition of decision parameters: During the 3D rendering process, the spatial feature parameters and hardware performance parameters of the current scene are acquired in real time to construct a multi-dimensional real-time decision parameter set.

[0021] LOD Level Determination: The LOD decision tree adopts a hierarchical structure with conditional judgment nodes and output nodes, introducing a dynamic LOD threshold output by a neural network model, and integrating multi-dimensional real-time decision parameters for comprehensive judgment. Judgment Logic: First, the relationship between hardware performance parameters and the threshold value is checked. If the hardware performance parameter is greater than or equal to the threshold value, LOD level downgrading is directly triggered; if the hardware performance parameter is less than the threshold value, the level determination is completed by combining spatial feature parameters and the dynamic LOD threshold, outputting the target LOD level (from high-modulus LOD0 to simplified-modulus LOD). n (or no rendering state); the threshold value is a preset hardware performance safety boundary, including GPU utilization of 85%, CPU utilization of 75%, video memory utilization of 75%, memory utilization of 70%, and frame rate of 30FPS.

[0022] Rendering and Dual-Indicator Evaluation: Image rendering is completed according to the target LOD level, and the frame rate and image quality indicators after rendering are simultaneously detected. The preset smoothness threshold for the frame rate indicator is 30FPS or 60FPS. If it is lower than this preset smoothness threshold, it is judged as performance failure. The image quality indicator is evaluated using preset visual evaluation standards, which include model detail retention, texture sharpness, and no LOD switching flicker. If core objects show loss of detail, it is judged as quality failure.

[0023] Model Iterative Optimization: The optimization results are fed back to the neural network model as feedback data. If the rendering results simultaneously meet the frame rate and quality requirements, the learning pattern of the current LOD threshold is retained and solidified. If any indicator fails to meet the requirements, the model weight parameters are adjusted using gradient descent to optimize the generation logic of the dynamic LOD threshold, forming a closed-loop self-optimization process of "pre-training - real-time acquisition - decision-making - rendering - optimization - iteration". The initial value of the learning rate η is set to 0.0001, and the learning rate η can be adjusted within the range of 0.00001 to 0.001 to control the parameter update step size. The weight matrices W1 and W2 and the bias vectors b1 and b2 are iteratively updated by minimizing the weighted mixed loss function.

[0024] The target values ​​for hardware utilization, the preset boundaries of the LOD dynamic threshold, and the frame rate target are as follows: CPU utilization 0.7, GPU utilization 0.8, VRAM usage 0.75, and system memory usage 0.65; LOD distance threshold 5m~200m, LOD screen ratio threshold 0.005~0.1, shader LOD threshold 10~200, and LOD global bias threshold -1.0~1.5; the frame rate target is FPS. target =30.

[0025] The beneficial effects of this invention are as follows:

[0026] 1. Achieve precise multi-dimensional adaptation: Through a dual-dimensional decision parameter system of "spatial features + hardware performance", it breaks through the limitations of traditional single distance dimension decision-making, so that the LOD threshold can not only match visual needs, but also fit the hardware carrying capacity, significantly improving cross-device adaptability and simultaneously meeting the rendering needs of high-end workstations and mobile devices.

[0027] 2. Optimize rendering performance and visual balance: The dynamic threshold generation mechanism driven by neural networks increases the average frame rate from 25-28 FPS to 30-32 FPS (reducing fluctuation by 80%), reduces GPU rendering time by 25%-30%, reduces memory and video memory usage by 25%-30% and 15%-25% respectively, and reduces model switching stuttering rate from 5%-8% to 0.5%-1%, completely solving the problem of coexistence of visual distortion and stuttering.

[0028] 3. Reduce manual debugging costs: The closed-loop self-optimization mechanism enables the model to continuously iterate through feedback data, eliminating the need to debug thresholds separately for different scenarios and devices, greatly improving rule reusability and reducing development and deployment costs.

[0029] 4. Adaptable to complex scenes and dynamic interactions: For dynamic scenes such as dense urban areas and tunnels, the threshold adjustment response is fast and can accurately match the changes in scene complexity and interaction intensity. It can still maintain stable rendering effects in high-frequency interaction scenes such as multi-person interaction in the metaverse and high-speed flight simulation. Attached Figure Description

[0030] Figure 1 This is a flowchart of an adaptive threshold generation method for 3D rendering LOD decision tree based on neural network according to the present invention.

[0031] Figure 2 This is a graph showing the input and output vectors of the neural network model of this invention.

[0032] Figure 3 This invention includes the activation function, weight matrix, and bias vector graph.

[0033] Figure 4 This is a graph showing the hardware utilization target value and LOD dynamic threshold of the present invention.

[0034] Figure 5 This is a comparison chart of the LOD (Level of Detail) of the dynamic parameter adjustment method of this invention and the traditional LOD. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0036] like Figure 1 As shown, Figure 1The flowchart of the adaptive threshold generation method for 3D rendering LOD decision tree based on neural network of the present invention includes the following steps:

[0037] S1. Neural network model pre-training: Construct a neural network model and train it based on historical sample data across devices and scenarios. The neural network model takes multi-dimensional sample parameters as input, and after nonlinear feature mapping of the hidden layer, it outputs the hardware utilization target value and the LOD dynamic threshold.

[0038] S2. Real-time acquisition of decision parameters: During the 3D rendering process, the spatial feature parameters and hardware performance parameters of the current scene are acquired in real time to construct a multi-dimensional real-time decision parameter set;

[0039] S3, LOD Level Determination: The LOD decision tree adopts a hierarchical structure of condition judgment nodes and output nodes, introduces the dynamic LOD threshold output by the neural network model, integrates multi-dimensional real-time decision parameters for comprehensive judgment, and outputs the target LOD level.

[0040] S4. Rendering and Dual-Indicator Optimization: Complete image rendering based on the target LOD level, and simultaneously detect the frame rate and image quality indicators after rendering to determine whether they meet the preset requirements.

[0041] S5. Model Iteration Optimization: The evaluation results are fed back to the neural network model as feedback data, and the model parameters are adjusted to optimize the generation logic of the LOD dynamic threshold, forming a closed-loop self-optimization.

[0042] like Figure 2 As shown, Figure 2 The input vector and output vector graph of the neural network model of this invention are shown. The historical sample data mentioned in step S1 includes multi-dimensional sample parameters, corresponding optimal LOD level, resource utilization rate and visual effect score. The multi-dimensional sample parameters include camera distance, screen ratio, viewing angle, model visibility, real-time frame rate, current CPU utilization, current GPU utilization, current video memory usage, and current memory usage. The hardware utilization target values ​​include CPU utilization target value, GPU utilization target value, video memory usage target value, and memory usage target value. The LOD dynamic thresholds include LOD distance threshold, LOD screen ratio threshold, shader LOD threshold, and LOD global bias adjustment threshold.

[0043] like Figure 3 As shown, Figure 3 The activation function, weight matrix, and bias vector map of this invention are as follows: In step S1, the hidden layer is set with 32 neurons, and the ReLU activation function is used to realize nonlinear feature mapping; The neural network model also includes a weight matrix W1 from the input layer to the hidden layer, a weight matrix W2 from the hidden layer to the output layer, a hidden layer bias vector b1, and an output layer bias vector b2.

[0044] Preferably, in step S1, the neural network model uses a weighted hybrid loss function L=α⋅L during training. hardware +β⋅L fps +γ⋅L lod The weighting ratio α:β:γ = 0.5:0.3:0.2, where L hardware For hardware utilization loss, L fps For frame rate loss, L lod This represents the reasonableness loss of the LOD threshold.

[0045] like Figure 4 As shown, Figure 4 This invention presents the hardware utilization target values ​​and LOD dynamic threshold diagram; hardware utilization target values: CPU utilization 0.7, GPU utilization 0.8, video memory usage 0.75, and system memory usage 0.65; preset boundaries of the LOD dynamic threshold: LOD distance threshold 5m~200m, LOD screen ratio threshold 0.005~0.1, shader LOD threshold 10~200, and LOD global bias threshold -1.0~1.5; frame rate target: FPS. target =30.

[0046] Preferably, the judgment logic of the LOD decision tree in step S3 is as follows: firstly, check the relationship between the hardware performance parameters and the critical value. If the hardware performance parameters are greater than or equal to the critical value, then directly trigger the LOD level downgrade. If the hardware performance parameters are less than the critical value, then combine the spatial feature parameters and the LOD dynamic threshold to complete the level determination and output the target LOD level (from high-poly LOD0 to simplified LODn or no rendering state). The critical value is a preset hardware performance safety boundary, including GPU utilization of 85%, CPU utilization of 75%, video memory utilization of 75%, memory utilization of 70%, and frame rate of 30 FPS.

[0047] Preferably, the specific criteria for judging the excellence of the two indicators in step S4 are as follows:

[0048] Frame rate metric: The preset smoothness threshold is 30FPS or 60FPS. If it is lower than the preset smoothness threshold, the performance is judged to be substandard.

[0049] Image quality metrics: Based on preset visual evaluation standards, determine whether the visual effect of the image meets the requirements of the scene. If the core object loses details, it is judged as substandard. The preset visual evaluation standards include model detail retention, texture clarity, and no LOD switching flicker.

[0050] Preferably, the specific logic for model iterative optimization in step S5 is as follows:

[0051] If the rendering result meets both the frame rate and image quality requirements, the learning pattern of the LOD threshold can be preserved and solidified. If the frame rate or image quality fails to meet the requirements, the neural network will adjust the weights based on the feedback data and re-optimize the generation logic of the LOD threshold.

[0052] Preferably, in step S5, during model iterative optimization, the model weight parameters are adjusted using gradient descent to optimize the generation logic of the LOD dynamic threshold; the weight matrices W1 and W2 and the bias vectors b1 and b2 are iteratively updated, and the parameter update formula is:

[0053]

[0054] Where L is a weighted mixed loss function, the initial value of the learning rate η is set to 0.0001, and the learning rate η can be adjusted in the range of 0.00001 to 0.001 to control the step size of parameter update.

[0055] Preferably, the method is applied to real-world 3D rendering, VR / AR, digital twins, GIS maps, or game scenes, and is compatible with high-end workstations, PCs, and mobile devices.

[0056] This design achieves self-learning and dynamic optimization of LOD core parameters by constructing a fully connected neural network model. The model takes multi-dimensional scene / hardware features as input and outputs a hardware utilization target value and a dynamic LOD threshold that are adapted to the current state. The training direction of the model is constrained by a weighted hybrid loss function, and finally the triple goal of "stable hardware utilization, satisfactory frame rate, and reasonable threshold" is achieved.

[0057] 1. Multi-objective collaboration

[0058] By using weighted loss to balance the three goals of "efficient use of hardware resources", "stable frame rate" and "reasonable threshold", performance / visual imbalance caused by single-dimensional optimization is avoided.

[0059] 2. Boundary Constraints

[0060] Preset upper and lower thresholds to prevent the model from outputting invalid values ​​(such as DT<5m or SST>0.1).

[0061] 3. Dynamic adaptation

[0062] The model input includes real-time hardware status (such as current GPU utilization), and the output LOD threshold can be dynamically adjusted according to hardware load to adapt to different devices and scenarios.

[0063] This invention primarily uses neural networks to optimize the dynamic parameters of the LOD decision tree. Under conditions of high device performance and abundant resources, it prioritizes rendering quality to achieve the best rendering quality. When resources are insufficient, it effectively ensures that rendering quality is not excessively sacrificed in order to maintain the frame rate.

[0064] Figure 5 This is a comparison chart of the performance metrics of the dynamically adjusted LOD and the traditional LOD of this invention. The chart compares the performance of the traditional LOD and the dynamically adjusted LOD. The results show that the dynamically adjusted LOD achieves comprehensive optimization in terms of resource consumption and smooth operation: the traditional LOD has the problem of preloading a lot of resources and large performance fluctuations during switching, while the dynamically adjusted LOD reduces CPU main thread time by 20%-30%, memory consumption by 25%-30%, and VRAM consumption by 15%-25% by loading on demand and dynamically unloading low-priority resources; especially in terms of the stuttering rate of model switching, which is crucial to the user experience, the stuttering rate of the traditional LOD reaches 5%-8%, while the dynamic solution reduces it to 0.5%-1%, an optimization of over 90%, which not only effectively saves hardware resources but also greatly improves the smoothness of operation.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for adaptive threshold generation of 3D rendering LOD decision trees based on neural networks, characterized in that, Includes the following steps: S1. Neural network model pre-training: Construct a neural network model and train it based on historical sample data across devices and scenarios. The neural network model takes multi-dimensional sample parameters as input, and after nonlinear feature mapping of the hidden layer, it outputs the hardware utilization target value and the LOD dynamic threshold. S2. Real-time acquisition of decision parameters: During the 3D rendering process, the spatial feature parameters and hardware performance parameters of the current scene are acquired in real time to construct a multi-dimensional real-time decision parameter set; S3, LOD Level Determination: The LOD decision tree adopts a hierarchical structure of condition judgment nodes and output nodes, introduces the dynamic LOD threshold output by the neural network model, integrates multi-dimensional real-time decision parameters for comprehensive judgment, and outputs the target LOD level. The decision logic of the LOD decision tree is as follows: first, check the relationship between the hardware performance parameters and the critical value. If the hardware performance parameters are greater than or equal to the critical value, then directly trigger the LOD level downgrade. If the hardware performance parameters are less than the critical value, the hierarchical determination is completed by combining the spatial feature parameters and the LOD dynamic threshold. The critical value is a preset hardware performance safety boundary, including GPU utilization of 85%, CPU utilization of 75%, video memory utilization of 75%, memory utilization of 70%, and frame rate of 30FPS. S4. Rendering and Dual-Indicator Optimization: Complete image rendering based on the target LOD level, and simultaneously detect the frame rate and image quality indicators after rendering to determine whether they meet the preset requirements. S5. Model Iteration Optimization: The evaluation results are fed back to the neural network model as feedback data, and the model parameters are adjusted to optimize the generation logic of the LOD dynamic threshold, forming a closed-loop self-optimization.

2. The adaptive threshold generation method for 3D rendering LOD decision trees based on neural networks according to claim 1, characterized in that, The historical sample data mentioned in step S1 includes multi-dimensional sample parameters, corresponding optimal LOD level, resource utilization rate, and visual effect score; the multi-dimensional sample parameters include camera distance, screen ratio, viewing angle, model visibility, real-time frame rate, current CPU utilization, current GPU utilization, current VRAM usage, and current memory usage; the hardware utilization target values ​​include CPU utilization target value, GPU utilization target value, VRAM usage target value, and memory usage target value; the LOD dynamic thresholds include LOD distance threshold, LOD screen ratio threshold, shader LOD threshold, and LOD global bias adjustment threshold.

3. The method for adaptive threshold generation of 3D rendering LOD decision tree based on neural network according to claim 1, characterized in that, In step S1, the hidden layer is set with 32 neurons, and the ReLU activation function is used to realize non-linear feature mapping. The neural network model also includes the weight matrix W1 from the input layer to the hidden layer, the weight matrix W2 from the hidden layer to the output layer, the hidden layer bias vector b1, and the output layer bias vector b2.

4. The method for adaptive threshold generation of 3D rendering LOD decision tree based on neural network according to claim 1, characterized in that, In step S1, during the training process of the neural network model, a weighted mixture loss function L=α⋅L is used. hardware +β⋅L fps +γ⋅L lod The weighting ratio α:β:γ = 0.5:0.3:0.2, where L hardware For hardware utilization loss, L fps For frame rate loss, L lod This represents the reasonableness loss of the LOD threshold.

5. The method for adaptive threshold generation of 3D rendering LOD decision tree based on neural network according to claim 4, characterized in that, Hardware utilization Target values: CPU utilization 0.7, GPU utilization 0.8, VRAM usage 0.75, RAM usage 0.65; Preset boundaries for LOD dynamic thresholds: LOD distance threshold 5m~200m, LOD screen ratio threshold 0.005~0.1, shader LOD threshold 10~200, LOD global bias threshold -1.0~1.5; Frame rate target: FPS target =30.

6. The method for adaptive threshold generation of 3D rendering LOD decision tree based on neural network according to claim 1, characterized in that, The specific criteria for judging excellence using the two indicators in step S4 are as follows: Frame rate metric: The preset smoothness threshold is 30FPS or 60FPS. If it is lower than the preset smoothness threshold, the performance is judged to be substandard. Image quality metrics: Based on preset visual evaluation standards, determine whether the visual effect of the image meets the requirements of the scene. If the core object loses details, it is judged as substandard. The preset visual evaluation standards include model detail retention, texture clarity, and no LOD switching flicker.

7. The method for adaptive threshold generation of 3D rendering LOD decision tree based on neural network according to claim 1, characterized in that, The specific logic of model iterative optimization in step S5 is as follows: If the rendering result meets both the frame rate and image quality requirements, the learning pattern of the LOD threshold can be preserved and solidified. If the frame rate or image quality fails to meet the requirements, the neural network will adjust the weights based on the feedback data and re-optimize the generation logic of the LOD threshold.

8. The adaptive threshold generation method for 3D rendering LOD decision trees based on neural networks according to claim 1, characterized in that, In step S5, during model iterative optimization, the model weight parameters are adjusted using gradient descent to optimize the generation logic of the LOD dynamic threshold; the weight matrices W1 and W2 and the bias vectors b1 and b2 are iteratively updated, and the parameter update formula is as follows: Where L is a weighted mixed loss function, the initial value of the learning rate η is set to 0.0001, and the learning rate η can be adjusted in the range of 0.00001 to 0.001 to control the parameter update step size.

9. A method for adaptive threshold generation of 3D rendering LOD decision tree based on neural networks according to any one of claims 1-8, characterized in that, The method is applicable to real-world 3D rendering, VR / AR, digital twins, GIS maps, or game scenes, and is compatible with high-end workstations, PCs, and mobile devices.

Images