A virtual simulation lightweight neural network image processing method for network games

Abstract

The application provides a virtual simulation lightweight neural network image processing method for network games, and belongs to the technical field of image processing.The application replaces the conventional convolution with a depth separable convolution module, and combines 8bit quantization perception training, so that the model parameter quantity is reduced from 128.6M of the conventional network to 24.3M, the mobile terminal inference time consumption is reduced from 47.2ms to 21.5ms, and the real-time processing requirement of more than 30 frames per second is met; the inference time consumption dynamic regulation logic is constructed, the feature extraction channel quantity is automatically adjusted according to the real-time occupation rate of CPU / GPU, the response speed is lower than 10ms, the problem that the fixed parameter model cannot adapt to the dynamic switching of the game scene is effectively solved, and the frame rate stability is improved to more than 98%.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a lightweight neural network image processing method for virtual simulation in online games. Background Technology

[0002] Virtual simulation scenarios in online games place stringent demands on the real-time performance of image processing and the adaptability of edge computing power. Conventional neural network models have a large number of parameters and long inference time, making them difficult to adapt to various types of game devices such as mobile devices and low-end PCs. Virtual simulation screens contain diverse image features such as character models, scene textures, and skill effects. The image processing stage needs to balance feature extraction accuracy and computational efficiency to avoid image quality loss and frame rate fluctuations.

[0003] While general convolutional neural networks such as VGG and ResNet have high feature extraction accuracy, their parameter count is usually in the millions or more, and the inference time per frame exceeds 50ms, which cannot meet the real-time requirements of 30-60 frames per second for online games. Lightweight networks such as MobileNet and ShuffleNet reduce the amount of computation to some extent, but they are designed for general natural images and do not fully consider the dynamic, dense, and hierarchical characteristics of virtual simulation images in online games. When directly applied to game scenes, they have problems such as feature omission, image quality degradation, and unstable frame rate.

[0004] Furthermore, existing solutions lack an adaptation mechanism for dynamic fluctuations in computing power on the end side. During the operation of online games, scene switching, special effects release, and changes in the number of players can cause significant fluctuations in CPU and GPU computing power utilization. Neural network models with fixed parameters cannot adapt to such fluctuations, and are prone to excessive inference time, resulting in frame drops and stuttering. Therefore, there is an urgent need for an image processing method that is both lightweight and dynamically adaptable for virtual simulation scenarios in online games to solve the above problems. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a lightweight neural network image processing method for virtual simulation in online games. It aims to resolve the contradiction between insufficient computing power on the edge and image processing quality, effectively reduce the number of model parameters and inference time while controlling image quality loss, and achieve dynamic adjustment based on real-time computing power usage to ensure the smoothness of rendering and visual fidelity of virtual simulation scenes in online games.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A lightweight neural network image processing method for virtual simulation in online games includes the following steps:

[0008] S1. Based on the feature distribution of virtual simulation images in online games, a lightweight depthwise separable convolutional module is reconstructed to reduce the computational load of the model.

[0009] S2. Using 8-bit integer quantization, the trained neural network model is subjected to quantization perception training of weights and activation values. During the quantization process, the feature extraction weights of character models, scene textures and key effects are retained, while minor weight parameters that have no substantial impact on game graphics are removed, and the accuracy loss of the quantized model weights is controlled within a preset threshold.

[0010] S3. Preprocess the virtual simulation image. The preprocessing removes complex color space conversion, multi-scale repeated scaling and repeated noise filtering, and only retains size normalization and pixel value normalization.

[0011] S4. Based on the real-time computing power usage of the network game during runtime, build a dynamic adjustment logic for inference time: read the CPU and / or GPU computing power usage of the game device in real time. When the computing power usage is higher than the first threshold, automatically reduce some redundant feature extraction channels of the model and close non-core detail texture extraction channels. When the computing power usage is lower than the second threshold, automatically restore full-channel feature extraction.

[0012] A preferred technical solution of this application: In step S1, the lightweight depth-separable convolution module includes depth convolution and pointwise convolution, wherein the depth convolution is based on single-channel image feature extraction, and the pointwise convolution fuses channel features;

[0013] The depth convolution kernel size of the pointwise convolution is set to 3×3, and the number of output channels of the pointwise convolution matches the number of feature channels of the virtual simulation image.

[0014] A preferred technical solution of this application: In step S1, the computational complexity of the lightweight depthwise separable convolution module is calculated using the following formula:

[0015]

[0016] in, The computational cost of a single-step convolution is... The height of the input image. The width of the input image. The number of input image channels, The kernel size is the convolution kernel size. This determines the number of output feature channels.

[0017] A preferred technical solution of this application: In step S2, the accuracy loss of the quantized model weights is controlled within a preset threshold. That is, by controlling the error value of the quantized model weights, it is ensured that the quantized model can still accurately extract the key features of the virtual simulation image of the online game, maintaining a balance between image quality and processing efficiency. The error of the quantized model weights is calculated using the following formula:

[0018]

[0019] in, The weighted average error, This represents the total number of model weights. The original weights before quantization. These are the quantized weights.

[0020] A preferred technical solution of this application is as follows: In step S2, the quantization perception training is carried out based on a self-made online game virtual simulation image dataset. The dataset covers static outdoor scenes, dense urban scenes, and highly dynamic combat scenes. During the training process, the quantization amplitude is adjusted according to the feature complexity of different scenes.

[0021] A preferred technical solution of this application is as follows: In step S3, the size normalization adjusts the game screens of different resolutions to a fixed input resolution, and the pixel value normalization maps the image pixel values ​​to the range of 0 to 1.

[0022] A preferred technical solution of this application is as follows: In step S3, the fixed input resolution is 512×512px or 256×256px, and the preprocessing time is controlled within 5ms.

[0023] In a preferred embodiment of this application, in step S4, the first threshold is a computing power utilization rate higher than 70%, and the second threshold is a computing power utilization rate lower than 40%; the response speed of the dynamic control logic for inference time consumption is lower than 10ms.

[0024] A preferred technical solution of this application is as follows: In step S4, when the computing power utilization rate is higher than the first threshold, the non-core detail texture extraction channel is turned off; when the computing power utilization rate is lower than the second threshold, the capture of detail features is strengthened to improve the image processing accuracy and image quality performance.

[0025] A preferred technical solution of this application is as follows: the method is compatible with mobile devices, low-end PC devices or mid-range PC devices, the processing time for a single frame image is less than 33ms, the peak signal-to-noise ratio of the image is maintained above 30dB, and the game frame rate stability is not less than 98%.

[0026] Compared with the prior art, the technical solution provided by this invention has the following advantages:

[0027] 1. This invention replaces conventional convolution with a depthwise separable convolution module and combines it with 8-bit quantized perceptual training, reducing the number of model parameters from 128.6M in conventional networks to 24.3M. The inference time on mobile devices is reduced from 47.2ms to 21.5ms, meeting the real-time processing requirements of over 30 frames per second. It also constructs a dynamic adjustment logic for inference time, automatically adjusting the number of feature extraction channels based on the real-time CPU / GPU utilization, with a response speed of less than 10ms. This effectively solves the problem that fixed parameter models cannot adapt to dynamic switching in game scenarios, and improves frame rate stability to over 98%.

[0028] 2. During the lightweight transformation process, this invention maintains the peak signal-to-noise ratio of the image above 30dB through error control formulas and targeted training strategies, with image quality loss imperceptible to the naked eye, fully meeting the visual presentation requirements of virtual simulation scenes in online games. After eliminating redundant operations, the preprocessing time is controlled within 5ms and is executed in parallel with the game rendering module, without occupying additional running memory or interfering with the core game logic. Moreover, the algorithm model can be seamlessly embedded into existing game rendering pipelines, adapting to various device types such as mobile devices, low-end PCs, and mid-range PCs, and possessing strong generalization ability and scene universality. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The present invention will be further described below with reference to the embodiments.

[0031] Please see Figure 1 The embodiments of this application provide a lightweight neural network image processing method for virtual simulation in online games, including the following steps:

[0032] S1. Based on the feature distribution of virtual simulation images in online games, a lightweight depthwise separable convolutional module is reconstructed to reduce the computational load of the model; by replacing the conventional convolutional layer with the reconstructed lightweight depthwise separable convolutional module, invalid channel operations are reduced.

[0033] In step S1, the lightweight depthwise separable convolution module includes depthwise convolution and pointwise convolution. Depthwise convolution extracts features based on single-channel images, and pointwise convolution fuses channel features. Specifically, the depthwise separable convolution is divided into two steps: depthwise convolution and pointwise convolution. First, single-channel image features are extracted through depthwise convolution, and then channel features are fused through pointwise convolution, avoiding redundant inter-channel operations of conventional convolution.

[0034] The depthwise convolution kernel size of the pointwise convolution is set to 3×3, and the number of output channels of the pointwise convolution matches the number of feature channels of the virtual simulation image.

[0035] The computational complexity of a lightweight depthwise separable convolution module is calculated using the following formula:

[0036]

[0037] in, The computational cost of a single-step convolution is... The height of the input image. The width of the input image. The number of input image channels, The kernel size is the convolution kernel size. This formula is used to calculate the number of output feature channels. It can accurately calculate the computational load of the lightweight convolution module, providing data support for model parameter tuning and adapting to virtual simulation images of online games at different resolutions.

[0038] The reconstructed module is adapted to the multi-layer feature extraction requirements of game images. The depth convolution kernel size is set to 3×3, which fits the small-scale feature extraction scenarios such as character edges and texture details. The number of point-by-point convolution channels matches the number of feature channels of virtual simulation images, reducing invalid channel operations. This module does not require modification of the original feature extraction logic, only optimizes the convolution operation method, and can be directly embedded into the network game image processing flow, adapting to the existing game rendering pipeline.

[0039] S2. Using 8-bit integer quantization, the trained neural network model is subjected to quantization perception training of weights and activation values. During the quantization process, the feature extraction weights of character models, scene textures and key effects are retained, while minor weight parameters that have no substantial impact on game graphics are removed, and the accuracy loss of the quantized model weights is controlled within a preset threshold.

[0040] In step S2, the accuracy loss of the quantized model weights is controlled within a preset threshold. That is, by controlling the error value of the quantized model weights, it is ensured that the quantized model can still accurately extract the key features of the virtual simulation image of the online game, maintaining a balance between image quality and processing efficiency. The error of the quantized model weights is calculated using the following formula:

[0041]

[0042] in, The weighted average error, This represents the total number of model weights. The original weights before quantization. These are the quantized weights;

[0043] The quantization perception training was conducted based on a self-made virtual simulation image dataset from online games. The dataset covers static outdoor scenes, dense urban scenes, and highly dynamic combat scenes. During the training process, the quantization amplitude was adjusted according to the feature complexity of different scenes.

[0044] The training neural network model is quantized using 8-bit integer quantization, which further reduces the model's storage size and inference computation, making it suitable for deployment on low-computing gaming devices such as mobile devices and low-spec PCs. Since the storage space and running memory of online game client devices are limited, large-volume models not only occupy extra storage space but also squeeze the memory resources of the game's core program. Quantization modification can effectively alleviate this problem.

[0045] The quantization process retains the core feature extraction weights of the model, focusing on retaining the feature extraction weights of character models, scene textures, and key special effects, while removing minor weight parameters that have no substantial impact on game graphics quality. This reduces model storage usage and inference time. After quantization, the model weight accuracy loss is controlled within a preset threshold, which will not significantly affect the processing effect and image quality of virtual simulation images.

[0046] The entire quantization perception training process was conducted based on a self-made virtual simulation image dataset from online games. The dataset selected game images with different lighting intensities, scene complexities, and dynamic effect densities, covering three typical game scenarios: static outdoor scenes, densely populated urban scenes, and highly dynamic combat scenes. This ensured that the quantized model was adaptable to diverse game scenarios and avoided issues such as image processing failures and blurry images in specific scenarios. During the training process, the quantization threshold was adjusted synchronously, and the quantization amplitude was adjusted according to the feature complexity of different game scenarios. The quantization amplitude was appropriately increased to improve efficiency in simple scenarios, while the quantization amplitude was slightly increased to ensure accuracy in complex scenarios, which met the stability requirements of real-time processing of game images and avoided image distortion caused by excessive quantization amplitude.

[0047] S3. Preprocess the virtual simulation image to remove complex color space conversion, multi-scale repeated scaling and repeated noise filtering, and retain only size normalization and pixel value normalization.

[0048] In step S3, size normalization adjusts game screens of different resolutions to a fixed input resolution, and pixel value normalization maps image pixel values ​​to the range of 0 to 1.

[0049] The input resolution is fixed at 512×512px or 256×256px, and the preprocessing time is controlled within 5ms.

[0050] The preprocessing stage of virtual simulation images in online games needs to simplify redundant operations, match the inference rhythm of lightweight neural networks, shorten the overall image processing time, and avoid the preprocessing stage becoming a bottleneck for real-time rendering. Conventional image preprocessing processes include redundant steps such as complex color space conversion, multi-scale repeated scaling, and repeated noise filtering. These operations are designed for general natural images and are not suitable for virtual simulation images in online games. Instead, they will increase the computation time.

[0051] The optimized preprocessing workflow directly eliminates redundant steps such as complex color space conversion and multi-scale repeated scaling, retaining only two core operations: size normalization and pixel value normalization. Both operations are consistent with the rendering characteristics of online game screens and have no additional invalid calculations. Size normalization uniformly adjusts game screens of different resolutions to 512×512px or 256×256px, adapting to the fixed input size requirements of lightweight models and avoiding abnormal inference time and feature extraction failures due to image size fluctuations. Pixel value normalization maps image pixel values ​​to the 0~1 range, simplifying the calculation steps of neural network activation functions, reducing the computational power consumption of single-step calculations, and speeding up inference.

[0052] The preprocessing operation is fully integrated with the real-time rendering process of the game screen and is executed in parallel with the rendering module of the game client. It does not occupy additional game running memory, does not interfere with the operation of the core game logic, interactive commands, and screen synchronization, and the preprocessing time is controlled within 5ms, so as not to slow down the overall image processing speed.

[0053] S4. Based on the real-time computing power usage of the network game during runtime, build a dynamic adjustment logic for inference time: read the CPU and / or GPU computing power usage of the game device in real time. When the computing power usage is higher than the first threshold, automatically reduce some redundant feature extraction channels of the model and close non-core detail texture extraction channels. When the computing power usage is lower than the second threshold, automatically restore full channel feature extraction.

[0054] In step S4, when the computing power utilization rate is higher than the first threshold, the non-core detail texture extraction channel is turned off; when the computing power utilization rate is lower than the second threshold, the capture of detail features is strengthened to improve the image processing accuracy and image quality.

[0055] By combining the real-time computing power usage of the client side during the operation of online games, a dynamic adjustment logic for inference time is built to ensure that the image processing time always matches the fixed frame rate requirements of the game and adapts to the computing power fluctuations caused by dynamic switching of game scenes. During the operation of online games, scene switching, special effect release, and changes in the number of players can cause significant fluctuations in the computing power usage of the client-side CPU and GPU. Neural network models with fixed parameters cannot adapt to such fluctuations and are prone to frame drops due to excessive inference time.

[0056] The algorithm has a built-in computing power monitoring module that reads the CPU and GPU computing power utilization of the game device in real time. When the device is running large open scenes or intensive battle scenes and the computing power utilization is higher than 70%, it automatically reduces some redundant feature extraction channels of the model and closes non-core detail texture extraction channels to shorten the inference time per frame. When the device is running static scenes or single-player scenes and the computing power is sufficient and the utilization is lower than 40%, it automatically restores full-channel feature extraction, strengthens the capture of detail features, and improves image processing accuracy and image quality.

[0057] This control logic requires no manual intervention and is executed automatically throughout the entire process. With a response time of less than 10ms, it can quickly adapt to the dynamic switching characteristics of online game scenarios, taking into account the smoothness and image quality of the screen under different operating states. It completely solves the industry pain point that fixed parameter models cannot adapt to variable computing power environments, and further improves the adaptability of the algorithm in actual game operation.

[0058] This invention provides a lightweight neural network image processing method for virtual simulation in online games, which is compatible with mobile devices, low-end PCs, or mid-range PCs. The processing time for a single frame image is less than 33ms, the peak signal-to-noise ratio of the image is maintained above 30dB, and the game frame rate stability is not less than 98%.

[0059] The analysis of the virtual simulation image features and lightweight requirements of this invention for online games is as follows:

[0060] I. Core Features of Virtual Simulation Images in Online Games

[0061] Virtual simulation images in online games possess three core characteristics: dynamism, density, and hierarchy. These characteristics directly determine the design direction of lightweight neural networks. Character movement, scene transitions, and special effects within virtual simulation scenes continuously alter image frame content. Effective features in a single frame have a short dwell time, requiring image processing algorithms to adapt to the game's 30-60 frame / s refresh rate to avoid feature omissions and processing delays. Game screens simultaneously contain multiple layers of features, including character model edges, scene texture details, and areas of gradual light and shadow transitions. Conventional convolutional networks easily extract redundant features, increasing the computational load on the device. Low-end game devices have a fixed computing power limit, unable to support the parallel computation of large-volume neural networks. Lightweight modifications must precisely eliminate ineffective convolutional layers while retaining key feature extraction channels for virtual simulation images, balancing computational efficiency and image fidelity.

[0062] II. Core Constraints of Lightweight Neural Network Image Processing

[0063] Lightweight neural network design must meet three constraints for online game operation: edge computing power threshold, single-frame image processing time limit, and image quality loss threshold. Regarding edge computing power threshold, the peak computing power of mobile game devices is mostly below 8 TOPS, while the peak computing power of low-end PCs is around 15-20 TOPS. The number of model parameters must be controlled within a reasonable range to avoid exceeding the device's computing power capacity. The single-frame image processing time limit must match the game's frame rate requirements, with a single-frame processing time of less than 33ms to ensure smooth rendering without stuttering or frame drops. Regarding image quality loss threshold, the texture clarity and edge integrity of the virtual simulation scene must not be significantly degraded, and the peak signal-to-noise ratio of the image must be maintained within a acceptable range to prevent affecting the player's visual experience and game interaction feedback. This constraint directly limits the extent of lightweight neural network modification; blindly compressing the model size should not lead to a decline in image processing accuracy.

[0064] The algorithm testing and performance analysis of this invention are as follows:

[0065] I. The test environment and test plan are designed as follows:

[0066] The test environment selected three types of mainstream online game running devices, covering mobile terminals, low-end PC terminals, and mid-range PC terminals, to simulate real game running scenarios. The core parameters of the test devices are detailed in Table 1. The test samples selected virtual simulation screens of three mainstream online games, including three typical scenarios: outdoor scenes, urban scenes, and battle scenes. 500 frames of images were selected as test samples for each scenario. The test indicators covered four core aspects: model parameter quantity, average inference time, peak signal-to-noise ratio of images, and frame rate stability. The differences in various indicators between conventional neural networks and lightweight neural networks were compared to verify the practicality and adaptability of the algorithm.

[0067] Table 1 Core parameters of the algorithm testing equipment

[0068]

[0069] II. The model size and inference efficiency test analysis are as follows:

[0070] The model size and inference efficiency tests focused on two indicators: parameter count and average inference time. The test data of the lightweight neural network in this paper were compared with those of the conventional convolutional neural network. The results are detailed in Table 2. The test process uniformly input a 512×512px virtual simulation image of a network game. The test was repeated 10 times for each type of device, and the average value was taken as the final result to exclude random errors. The test data shows that the lightweight neural network has a significantly reduced parameter count compared with the conventional network. The average inference time under the three types of devices is less than 33ms, which meets the real-time processing requirements of network games of more than 30 frames / s. The inference efficiency of mobile devices is particularly improved, which is suitable for the computing power constraints of low-end devices.

[0071] Table 2. Test Results of Model Size and Inference Efficiency

[0072]

[0073] III. The image processing quality and game compatibility test analysis are as follows:

[0074] Image processing quality was evaluated using peak signal-to-noise ratio (PSNR) as the core indicator. A higher PSNR indicates lower image quality loss and more complete preservation of virtual simulation scene features. Game compatibility was determined by the stability of the game's frame rate. Smaller frame rate fluctuations indicate less interference from the algorithm on the game's operation. The test results for both indicators are detailed in Table 3. The test scenario was a uniformly selected online game battle scene, which has dense dynamic features and can better verify the algorithm's actual adaptability. The test data shows that the PSNR of the image processed by the lightweight neural network is slightly lower than that of the conventional network, but the difference is within a reasonable range, and the image quality loss is imperceptible to the naked eye. The game frame rate stability is maintained above 98%, with no frame drops or stuttering, fully adapting to the real-time operation requirements of the online game virtual simulation scene.

[0075] Table 3. Results of Image Processing Quality and Game Compatibility Tests

[0076]

[0077] This invention addresses the pain points of image processing in virtual simulation scenarios of online games by designing a lightweight neural network image processing algorithm. Through reconstructing a lightweight convolutional module, quantizing perceptual training, optimizing preprocessing, and controlling inference time, it achieves model size compression and improved inference efficiency. The algorithm is designed to meet the multi-terminal operation requirements of online games, adapting to different computing power devices and diverse virtual simulation scenarios. While controlling image quality loss, it effectively reduces the computational load on the client side, ensuring real-time rendering and stable frame rates. Test results verify that the algorithm outperforms conventional neural networks in terms of parameter quantity, inference time, and image processing quality, demonstrating strong engineering application value. Further optimization can be performed on highly dynamic combat scenarios in online games to further reduce the image quality difference after lightweight processing, expanding the algorithm's adaptability in VR and AR online game virtual simulation visual processing, and improving the algorithm's scenario universality and technical practicality.

[0078] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

[0079] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only independent technical solutions. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A lightweight neural network image processing method for virtual simulation in online games, characterized in that, Includes the following steps: S1. Based on the feature distribution of virtual simulation images in online games, a lightweight depthwise separable convolutional module is reconstructed to reduce the computational load of the model. S2. Using 8-bit integer quantization, the trained neural network model is subjected to quantization perception training of weights and activation values. During the quantization process, the feature extraction weights of character models, scene textures and key effects are retained, while minor weight parameters that have no substantial impact on game graphics are removed, and the accuracy loss of the quantized model weights is controlled within a preset threshold. S3. Preprocess the virtual simulation image. The preprocessing removes complex color space conversion, multi-scale repeated scaling and repeated noise filtering, and only retains size normalization and pixel value normalization. S4. Based on the real-time computing power usage of the network game during runtime, build a dynamic adjustment logic for inference time: read the CPU and / or GPU computing power usage of the game device in real time. When the computing power usage is higher than the first threshold, automatically reduce some redundant feature extraction channels of the model and close non-core detail texture extraction channels. When the computing power usage is lower than the second threshold, automatically restore full-channel feature extraction.

2. The lightweight neural network image processing method for virtual simulation in online games according to claim 1, characterized in that, In step S1, the lightweight depthwise separable convolution module includes depthwise convolution and pointwise convolution. The depthwise convolution is based on features extracted from a single-channel image, and the pointwise convolution fuses channel features. The depth convolution kernel size of the pointwise convolution is set to 3×3, and the number of output channels of the pointwise convolution matches the number of feature channels of the virtual simulation image.

3. The lightweight neural network image processing method for virtual simulation in online games according to claim 2, characterized in that, In step S1, the computational complexity of the lightweight depthwise separable convolution module is calculated using the following formula: ； in, The computational cost of a single-step convolution is... The height of the input image. The width of the input image. The number of input image channels, The kernel size is the convolution kernel size. This determines the number of output feature channels.

4. The lightweight neural network image processing method for virtual simulation in online games according to claim 1, characterized in that, In step S2, the accuracy loss of the quantized model weights is controlled within a preset threshold. That is, by controlling the error value of the quantized model weights, it is ensured that the quantized model can still accurately extract the key features of the virtual simulation image of the online game, maintaining a balance between image quality and processing efficiency. The error of the quantized model weights is calculated using the following formula: ； in, The weighted average error, This represents the total number of model weights. The original weights before quantization. These are the quantized weights.

5. The lightweight neural network image processing method for virtual simulation in online games according to claim 4, characterized in that, In step S2, the quantization perception training is carried out based on a self-made online game virtual simulation image dataset. The dataset covers static outdoor scenes, dense urban scenes, and highly dynamic combat scenes. During the training process, the quantization amplitude is adjusted according to the feature complexity of different scenes.

6. The lightweight neural network image processing method for virtual simulation in online games according to claim 1, characterized in that, In step S3, the size normalization adjusts the game screens of different resolutions to a fixed input resolution, and the pixel value normalization maps the image pixel values ​​to the range of 0 to 1.

7. The lightweight neural network image processing method for virtual simulation in online games according to claim 6, characterized in that, In step S3, the fixed input resolution is 512×512px or 256×256px, and the preprocessing time is controlled within 5ms.

8. The lightweight neural network image processing method for virtual simulation in online games according to claim 1, characterized in that, In step S4, the first threshold is a computing power utilization rate higher than 70%, and the second threshold is a computing power utilization rate lower than 40%; the response speed of the dynamic control logic for inference time consumption is lower than 10ms.

9. The lightweight neural network image processing method for virtual simulation in online games according to claim 8, characterized in that, In step S4, when the computing power utilization rate is higher than the first threshold, the non-core detail texture extraction channel is turned off; when the computing power utilization rate is lower than the second threshold, the capture of detail features is strengthened to improve the image processing accuracy and image quality.

10. The lightweight neural network image processing method for virtual simulation of online games according to any one of claims 1 to 9, characterized in that, This method is compatible with mobile devices, low-end PCs, or mid-range PCs. The processing time for a single frame image is less than 33ms, the peak signal-to-noise ratio of the image is maintained above 30dB, and the game frame rate stability is not less than 98%.

Images