An animation image three-dimensional rendering analysis method based on artificial intelligence

Abstract

The application discloses an animation image three-dimensional rendering analysis method based on artificial intelligence, relates to the technical field of three-dimensional modeling and rendering, and comprises the following steps: optimizing a geometric body in a three-dimensional scene through a deep learning algorithm, enhancing the material of the geometric body by using a generative adversarial network model, and optimizing a rendering light source of the three-dimensional scene through reinforcement learning; a sampling density strategy is configured, Monte Carlo random sampling and a convolutional neural network model are used, the sampling density strategy is adjusted, an effective light path and a high-quality light map are selected, preliminary image rendering is performed by using the optimized geometric body, the enhanced geometric body material and the high-quality light map, image semantic segmentation is performed through U-Net, and post-processing is performed according to a semantic segmentation result and an SRGAN. The three-dimensional scene rendering is optimized through the deep learning algorithm, the generative adversarial network and the reinforcement learning, and efficient and intelligent three-dimensional animation image rendering is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional modeling and rendering, and particularly relates to an animation image three-dimensional rendering analysis method based on artificial intelligence. BACKGROUND

[0002] With the development of computer graphics, the three-dimensional rendering technology of animation images has evolved from simple geometric shapes to complex scene simulation. Early techniques mainly relied on hand-made models and preset light sources, limiting the realism and detail performance of animation images. In recent years, breakthroughs in deep learning algorithms in the field of image processing have brought innovation to three-dimensional rendering. In particular, through deep learning optimization of geometric body structure in three-dimensional scenes and the use of generative adversarial networks to enhance material texture, the quality of animation images has made a qualitative leap. However, existing methods still face high computational cost and lack of realism when dealing with complex scenes, especially in dynamically adjusting sampling density to adapt to different scene requirements.

[0003] Although existing technologies have made certain progress in the three-dimensional rendering of animation images, there are still several key problems that need to be solved. First, in terms of geometric body optimization, traditional methods rely on manual adjustment or simple algorithm implementation, making it difficult to accurately capture complex shapes and detailed features, resulting in insufficient realism and delicacy of rendering effects. Second, although generative adversarial networks have shown great potential in material enhancement, their training process is complex, computationally expensive, and performs poorly in handling multi-material mixed scenes, often resulting in texture distortion. Light source optimization is another challenge. Current techniques mostly use preset light source positions and intensities, lacking dynamic adjustment capabilities, making it difficult to adapt to different scene requirements and affecting the realism and level of detail of the final rendering effect. Finally, regarding sampling density strategy, existing methods usually adopt fixed or simple adaptive strategies, which cannot efficiently handle complex three-dimensional scene changes, resulting in low rendering efficiency and unstable image quality. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an animation image three-dimensional rendering analysis method based on artificial intelligence to solve the problems of low rendering efficiency and unstable image quality.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an animation image three-dimensional rendering analysis method based on artificial intelligence, which comprises: optimizing a geometric body in a three-dimensional scene through a deep learning algorithm, enhancing the material of the geometric body by using a generative adversarial network model, and optimizing a rendering light source of the three-dimensional scene through reinforcement learning; configuring a sampling density strategy, using Monte Carlo random sampling and a convolutional neural network model to dynamically adjust the sampling density strategy, generating and selecting an effective light path and a high-quality light map; using the optimized geometric body, the enhanced material of the geometric body, and the high-quality light map to perform preliminary image rendering to generate a preliminary rendering image result; performing image semantic segmentation through a U-Net, and performing post-processing according to a semantic segmentation result and a SRGAN to generate a high-quality rendering image.

[0008] As a preferred scheme of the animation image three-dimensional rendering analysis method based on artificial intelligence, wherein: the step of optimizing the geometric body in the three-dimensional scene through the deep learning algorithm comprises:

[0009] using PointNet++ to select a representative point as a center point through far point sampling at each layer in the three-dimensional scene;

[0010] using a spherical neighborhood query to find neighbor points around the center point and divide them into local regions;

[0011] using a multi-layer perceptron to extract geometric shape features, geometric normal direction features, and geometric curvature features of the neighbor points in the local regions;

[0012] aggregating the geometric shape features, the geometric normal direction features, and the geometric curvature features into global features through maximum pooling.

[0013] As a preferred scheme of the animation image three-dimensional rendering analysis method based on artificial intelligence, wherein: the step of enhancing the material of the geometric body by using the generative adversarial network model comprises:

[0014] reading color, reflectivity, and transparency from the geometric body as material data;

[0015] inputting the material data into the generative adversarial network model as a training set;

[0016] creating multiple variants to rotate, scale, and adjust the brightness of the material data through data enhancement;

[0017] generating new material data by a generator of the generative adversarial network model according to the trained material data;

[0018] reapplying the properties and textures of the new material data to the geometric body to replace the original material.

[0019] As a preferred scheme of the animation image three-dimensional rendering analysis method based on artificial intelligence, wherein: the step of optimizing the rendering light source of the three-dimensional scene through reinforcement learning is specifically,

[0020] Reading light source data from the three-dimensional scene through a rendering engine API;

[0021] Defining the state, action and reward of the three-dimensional scene environment;

[0022] Selecting a deep Q learning model, generating a frame of illumination image using the light source data, and rendering the illumination image in the three-dimensional scene environment;

[0023] Evaluating the quality of the rendered illumination image through image processing and giving a reward to feed back to the deep Q learning model;

[0024] If the deep Q learning model detects that the shadow area of the illumination image is too large, the light source is increased to fill the shadow area;

[0025] If the deep Q learning model detects that the illumination image is overexposed, the light source intensity is reduced;

[0026] According to the reflectivity and geometric layout of the rendered illumination image, the light source position and direction are adjusted.

[0027] As a preferred scheme of the animation image three-dimensional rendering analysis method based on artificial intelligence, wherein: the step of configuring the sampling density strategy, using Monte Carlo random sampling and a convolutional neural network model, dynamically adjusting the sampling density strategy, generating and selecting effective light paths and high-quality illumination maps, is specifically,

[0028] According to the optimized rendering light source, pre-rendering is performed to generate a low-resolution illumination map as a basis for evaluating the illumination intensity in different areas of the three-dimensional scene;

[0029] Defining the sampling density strategy and setting the illumination intensity threshold;

[0030] In areas with high illumination intensity, when the illumination intensity exceeds the illumination intensity threshold, the sampling density is reduced;

[0031] In areas with low illumination intensity, when the illumination intensity does not exceed the illumination intensity threshold, the sampling density is increased;

[0032] According to the real-time light source data in rendering, the sampling density strategy is adjusted;

[0033] Randomly generating light paths from the light source data through Monte Carlo random sampling, simulating the light paths from the light source, passing through the geometric surface in the three-dimensional scene, and reaching the camera rendering light path;

[0034] The simulated light path is analyzed by a convolutional neural network model, and an effective light path for image rendering is selected;

[0035] The light contribution of the light intensity in the three-dimensional scene is calculated to obtain a high-quality light map.

[0036] As a preferred scheme of the three-dimensional rendering analysis method for animated images based on artificial intelligence, wherein: the simulated light path is analyzed by a convolutional neural network model, and an effective light path for image rendering is selected, and the specific steps are,

[0037] The total path length of the light from the light source to the camera rendering is calculated;

[0038] When simulating the light path, the number of reflections of the light is counted to obtain the number of reflections;

[0039] The material reflectivity of each reflection of the light path is recorded;

[0040] The path length, the number of reflections and the material reflectivity are combined as path features;

[0041] The path features are defined as a training set of the convolutional neural network model and are input into the convolutional neural network model;

[0042] The local features of the light path are extracted by a convolutional layer as a feature map;

[0043] The dimension of the feature map is reduced by a pooling layer;

[0044] The feature map is converted into a light path importance score by a fully connected layer, and an effective light path is selected.

[0045] As a preferred scheme of the three-dimensional rendering analysis method for animated images based on artificial intelligence, wherein: the optimized geometric body, the enhanced geometric body material and the high-quality light map are used for preliminary image rendering to generate a preliminary rendering image result, and the specific steps are,

[0046] The optimized geometric body and the enhanced geometric body material are used for preliminary image rendering;

[0047] The initial image of each geometric body is drawn by a graphics API;

[0048] The high-quality light map is combined with the color, reflectivity and transparency of the material data to calculate the final color value of each preliminary rendering image pixel, and is applied to the initial image of each geometric body;

[0049] The initial images of all geometric bodies are combined into an image frame as a preliminary rendering image result.

[0050] As a preferred scheme of the animation image three-dimensional rendering analysis method based on artificial intelligence, the image semantic segmentation is performed by the U-Net, and the high-quality rendering image is generated by performing post-processing according to the semantic segmentation result and the SRGAN, and the specific steps are as follows,

[0051] The preliminary rendering image is segmented by the U-Net to identify the image semantics of the character region, the building region, the sky region and the background region in the preliminary rendering image, and a segmentation result is generated.

[0052] According to the segmentation result, the character region is sharpened, and the background region is blurred.

[0053] Based on the image semantics of the character region, the building region, the sky region and the background region, the color tone, the saturation and the brightness of the preliminary rendering image are adjusted.

[0054] The depth of field of the camera is simulated by using the segmentation result.

[0055] The generator in the SRGAN is used to convert the preliminary rendering image into a high-resolution image.

[0056] The discriminator in the SRGAN is used to distinguish the high-resolution image from a real high-resolution image, and to enhance the texture details and the clarity of the preliminary rendering image.

[0057] The high-quality rendering image is output by the generator in the SRGAN.

[0058] In a second aspect, the present application provides a computer device, comprising a memory and a processor, and the memory stores a computer program, wherein the computer program is executed by the processor to realize any step of the animation image three-dimensional rendering analysis method based on artificial intelligence according to the first aspect of the present application.

[0059] In a third aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by the processor to realize any step of the animation image three-dimensional rendering analysis method based on artificial intelligence according to the first aspect of the present application.

[0060] The application has the beneficial effects that: the application optimizes three-dimensional scene rendering through a deep learning algorithm, a generative adversarial network and reinforcement learning, realizes efficient and intelligent three-dimensional animation image rendering, first, the deep learning algorithm is used for optimizing geometric bodies, extracting and aggregating shape features and other information, enhancing the model detail expressiveness, providing high-quality data support for subsequent steps, and improving the realism of the rendered image, second, the generative adversarial network is used to enhance the material of the geometric body, create multiple variants, so that the material presents a natural effect under different lighting conditions, greatly enriches the form of expression of the object, improves the quality of the rendered image, and finally, the reinforcement learning is used to dynamically adjust the light source configuration, automatically optimizes the light source position and intensity according to the rendering result, reduces the manual intervention, improves the shadow and overexposure problems, and enhances the natural and harmonious degree of the light and shadow effect. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0062] Figure 1 The flowchart of the animation image three-dimensional rendering analysis method based on artificial intelligence in embodiment 1.

[0063] Figure 2 The flowchart of generating a preliminary image rendering result in embodiment 1. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification.

[0065] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the application, therefore the application is not limited to the specific embodiments disclosed below.

[0066] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0067] Embodiment 1, refer to Figure 1 and Figure 2The embodiment provides an artificial intelligence-based three-dimensional rendering analysis method for an animation image, and comprises the following steps:

[0068] S1, optimizing a geometric body in a three-dimensional scene through a deep learning algorithm, enhancing a material of the geometric body by using a generative adversarial network model, and optimizing a rendering light source of the three-dimensional scene through reinforcement learning.

[0069] Further, PointNet++ is used to select representative points as center points through far point sampling at each layer in the three-dimensional scene;

[0070] Specifically, at each layer of the three-dimensional scene, PointNet++ selects representative points by using a far point sampling method, and the points are selected as center points to ensure uniform distribution of the point cloud in space and reduce redundant points.

[0071] A spherical neighborhood query is used to find neighbor points around the center points and divide the neighbor points into local regions;

[0072] Specifically, for each selected center point, PointNet++ uses a spherical neighborhood query to find all neighbor points within a certain radius around the center point, and the neighbor points form a local region or neighborhood, and in this way, PointNet++ establishes a local point set for each center point.

[0073] A multi-layer perception machine is used to extract geometric shape features, geometric normal direction features and geometric curvature features of the neighbor points in the local region;

[0074] Specifically, in each local region, PointNet++ applies a multi-layer perception machine to process the neighbor points, and the multi-layer perception machine is responsible for extracting key geometric features from the position, relative position and other information of the points.

[0075] For each center point, PointNet++ has obtained a group of neighbor points through the spherical neighborhood query, and each point contains its coordinates (x, y, z) in the three-dimensional space, which can be directly input into the multi-layer perception machine, or additional attributes of the point such as color and normal vector are input;

[0076] The multi-layer perception machine accepts the coordinates of each point as input, and may also add relative position information of the point, that is, the distance or vector of the point to the center point, which can better capture the local structure.

[0077] In order to enhance the feature representation, the multi-layer perception machine embeds the original point coordinates into a higher-dimensional space, for example, a position vector (3-dimensional) of a point can be mapped to a higher-dimensional feature vector (such as 64-dimensional or 128-dimensional), and the process is realized through a linear transformation and a nonlinear activation function (such as ReLU);

[0078] Multilayer perceptrons process these points through multiple (usually 3) fully connected neural network layers, each containing weights and biases. By learning to optimize these parameters, multilayer perceptrons can identify which input features are most useful for geometry shape, geometry normal direction, and geometry curvature.

[0079] Specifically, through deep learning of the network, the multilayer perceptron can capture the overall shape of the point cloud, such as the distribution and density of points, i.e., geometric shape features, which describe the overall distribution and structure of the points.

[0080] Multilayer perceptrons learn patterns of normal vectors to obtain normal information, which reflects the orientation of the surface in space, i.e., the geometric normal direction characteristics.

[0081] By analyzing the relationship between points using a multilayer perceptron, the multilayer perceptron can infer the curvature of a surface and obtain the curvature characteristics of the geometric body, which indicate the degree of surface change, such as whether the surface is flat or curved.

[0082] After processing each or three fully connected neural network layers, the resulting geometric feature vector contains both the local information of each point and the relationship with its neighboring points. The geometric features provide a rich geometric description for each local region.

[0083] For each point in a local region, after processing by a multilayer perceptron, a new feature vector is output, which represents the geometric features of the region.

[0084] Max pooling is used to aggregate geometric shape features, geometric normal direction features, and geometric curvature features into global features;

[0085] Specifically, PointNet++ uses max pooling to aggregate the features of all local regions (geometric shape features, geometric normal direction features, and geometric curvature features). For each feature (shape feature, normal direction feature, curvature feature, etc.), max pooling selects the maximum value in each feature dimension. If a geometric feature vector has N dimensions, then max pooling will select the maximum value from the same dimension of all local regions, ensuring that the most significant geometric features are retained in the final global feature vector.

[0086] The advantage of doing this is that it can generate a fixed-length global feature vector regardless of how the number or distribution of local regions changes, enabling PointNet++ to handle point clouds with a variable number of points.

[0087] Read the color, reflectivity, and transparency of the geometry as material data;

[0088] Color, reflectivity, and transparency are extracted directly because each geometry is equipped with these materials;

[0089] Specifically, color is the color information extracted from the surface of each geometry, which is usually in the form of RGB values ​​or material maps;

[0090] Reflectivity is the reflection coefficient or reflection map of a geometry material, describing how light is reflected on the surface;

[0091] Transparency is defined by reading transparency values ​​or transparency maps to determine the transparency or semi-transparency properties of a material;

[0092] Material data is used as a training set and input into the generative adversarial network model;

[0093] Specifically, the material data is converted into an image format and input into the generative adversarial network model;

[0094] All extracted material data were then compiled into a training dataset, which contained information on the color, reflectivity, and transparency of all geometries.

[0095] Create multiple variations and use data augmentation to rotate, scale, and adjust the brightness of material data;

[0096] Specifically, this involves rotating the material data to generate new perspectives or forms of expression.

[0097] Adjust the size of the material data to create material samples at different resolutions;

[0098] The brightness of the material is changed to simulate different lighting conditions, thus enriching the training material samples;

[0099] The generator of the generative adversarial network model generates new material data based on the trained material data;

[0100] Specifically, the generator learns material features from the trained material data, generates new texture details from the original material data, and adjusts reflection and refraction properties to make the material look more natural and realistic.

[0101] The properties and textures of the new material data are reapplied to the geometry, replacing the original material;

[0102] Light source data (including point light, directional light, and ambient light) can be directly read from the 3D scene via the rendering engine API;

[0103] Define the state, actions, and rewards of the 3D scene environment;

[0104] Specifically, the state is defined to include the current configuration of all light sources (position, direction, intensity), the position of the geometry, and material properties, etc.

[0105] Actions are defined as actions such as moving a light source, changing the direction of a light source, increasing or decreasing the intensity of a light source, or increasing / decreasing the number of light sources.

[0106] The reward is a set reward mechanism. For example, increasing the brightness of a 3D scene, reducing shadow areas, and enhancing the visibility of material details will result in a positive reward, while overexposure or excessive shadows will result in a negative reward.

[0107] Select a deep Q-learning model, use light source data to generate a frame of lighting image, and render the lighting image in a 3D scene environment;

[0108] Specifically, the current light source data is used to render the 3D scene and generate a frame of lighting image, which is done directly in the rendering engine;

[0109] If iterative training is required, then a rendering should be performed after each adjustment of the light source;

[0110] The quality of the rendered illuminated image is evaluated through image processing (such as contrast analysis, edge detection, brightness distribution, etc.), and a reward is given, which is fed back to the deep Q-learning model.

[0111] Specifically, image processing, such as connected component analysis, is used to identify and measure the size of shadow regions. The shadow area is quantified by pixel count or shadow percentage. Reducing the shadow area is to ensure that details and materials in the 3D scene are properly illuminated and visibility is improved.

[0112] Edge detection (such as Canny) is used to evaluate the sharpness of object details. The edge intensity or detail texture in the rendered lighting image is compared with the reference image or expected level of detail. High detail visibility means that the details of materials and geometry are fully displayed in the rendering, enhancing the realism and aesthetics of the lighting image.

[0113] By analyzing the histogram of the lighting image, the brightness distribution of pixels can be determined. Overexposed areas will have pixels concentrated at the high brightness end (close to 255), indicating that these areas are too bright. Conversely, underexposed areas will have pixels concentrated at the low brightness end (close to 0), indicating that these areas are too dark. In addition, HDR information or dynamic range indicators can be used to ensure that the image is not overexposed or underexposed, maintain the natural lighting effect of the 3D scene, and avoid brightness distortion.

[0114] To calculate the histogram of a lit image, an array is created, called the histogram array. The length of the array is equal to the number of brightness values ​​(e.g., 256 for an 8-bit grayscale image). Each position represents a brightness level from 0 to 255. The initial values ​​of the array are all zero.

[0115] Read the brightness value of each pixel in the illumination image. If it is a color image, first convert the color image to grayscale values.

[0116] Find the position corresponding to the brightness value in the histogram array and increment the counter at that position.

[0117] Observe the right end of the histogram (close to 255). If there are a large number of pixels, it indicates that some parts of the illuminated image may be overexposed.

[0118] Observe the left end of the histogram (close to 0). If there are many pixels, it means that the image may be underexposed.

[0119] Rewards are allocated based on the evaluation results; high-quality renderings receive high rewards, while low or negative rewards are given for poor renderings.

[0120] The reward is fed back to the deep Q-learning model, and the Q-value table of the deep Q-learning model is updated.

[0121] If the deep Q-learning model detects that the shadow area of ​​the illuminated image is too large, then add a light source to fill the shadow area;

[0122] If the deep Q-learning model detects that the lighting image is overexposed, it reduces the light source intensity to avoid overexposure and achieve a natural exposure effect.

[0123] Adjust the position and direction of the light source based on the reflectivity and geometry of the rendered lighting image;

[0124] Reflectivity is the reflectivity of a material under current lighting conditions, not an inherent property of the material itself;

[0125] Specifically, this involves moving or rotating the light source to better capture the reflectivity or reflective properties of the material;

[0126] Adjust the direction of the light source to distribute the light more evenly on the key geometry and highlight the details;

[0127] It should be noted that by optimizing geometry through deep learning algorithms, enhancing materials using generative adversarial networks (GANs), and optimizing light sources through reinforcement learning, this invention achieves a comprehensive improvement in 3D scene rendering. First, the accurate extraction and optimization of geometric features enhances the realism and detail of the model. Second, GANs enhance the realism and diversity of materials, enabling objects to exhibit natural effects under various lighting conditions. Finally, reinforcement learning dynamically adjusts the light source configuration, reducing shadows and overexposure issues and improving the natural harmony of lighting effects. Overall, these improvements not only significantly enhance the quality and visual effects of rendering but also increase rendering efficiency.

[0128] S2. Configure the sampling density strategy. Utilize Monte Carlo random sampling and a convolutional neural network model to dynamically adjust the sampling density strategy, generate and select effective ray paths and high-quality light maps.

[0129] Furthermore, based on the optimized rendering light source, a quick pre-render is performed to generate a low-resolution light map, which serves as the basis for evaluating the light intensity in different areas of the 3D scene.

[0130] Define a sampling density strategy and set a light intensity threshold;

[0131] For example, a light intensity of over 90% can be considered a highlight area, while less than 30% is a low light area;

[0132] In areas with high illumination intensity, when the illumination intensity exceeds the illumination intensity threshold, the sampling density is reduced because the contribution of highlight areas to the illumination image is usually more stable, and high-density sampling is not required to capture details. Adjustments can be made, such as reducing the number of sampling paths or the sampling resolution.

[0133] For example, it is stipulated that in areas where the light intensity exceeds 80%, only 5 samples are taken per square pixel instead of the standard 10.

[0134] In areas with low light intensity, if the light intensity does not exceed the light intensity threshold, the sampling density is increased to ensure that details in low-light or shadow areas can be accurately captured by increasing the number of light paths or improving the resolution of path sampling.

[0135] For example, in areas with light intensity below 30%, each square pixel may be sampled 20 times to ensure that details in shadows are captured;

[0136] Based on real-time light source data during rendering, the sampling density strategy is adjusted. For example, if an unexpected highlight appears in a certain area, the sampling density of that area can be dynamically reduced. Conversely, if changes in lighting conditions in a certain area cause loss of detail, the sampling density is increased. A controller is used to automatically adjust the sampling density based on the current rendering results (such as through histogram analysis or brightness evaluation).

[0137] Light paths are randomly generated from light source data using Monte Carlo random sampling, simulating the light path that starts from the light source, passes through the geometric surfaces of the 3D scene, and reaches the light path rendered by the camera.

[0138] Specifically, Monte Carlo random sampling is used to randomly generate a large number of light paths from the light source data. These paths simulate how light rays start from the light source, pass through the 3D scene, and undergo reflection and refraction until they reach the position rendered by the camera.

[0139] The simulation of light paths is analyzed using a convolutional neural network model, and the effective light paths for rendering are selected.

[0140] Specifically, it calculates the total path length of light rays from the light source to the camera rendering; the expression is:

[0141]

[0142] Among them, L total The total path length from the light source to the camera is given by d. y Let y be the length of the y-th segment in the ray path, and n be the total number of segments in the ray path. y represents starting from 1, representing the first segment or the first point of the ray path, and increasing to n.

[0143] When simulating the path of light rays, the number of reflections is counted to obtain the number of reflections.

[0144] Specifically, whenever a ray interacts with and is reflected by the surface of an object in the 3D scene, the value of the reflection counter is increased, and the total number of reflections in the ray path is recorded, reflecting the complexity of the ray path and the number of times the light propagates.

[0145] Record the material reflectivity for each reflection along the light path;

[0146] Specifically, at each reflection or refraction, the material properties of that point or segment are queried, and the reflectivity of that material is recorded;

[0147] The path length, number of reflections, and material reflectivity are combined to form the path feature;

[0148] Define the path features as the training set of the convolutional neural network model and input them into the convolutional neural network model;

[0149] Specifically, the path features (length, number of reflections, reflectivity) are converted into structured data and standardized. The standardized path features are then used as the training set and input into the convolutional neural network model.

[0150] Local features of the light path are extracted using convolutional layers and used as feature maps.

[0151] Specifically, convolutional layers are used to scan the input path features to capture local spatial features, such as changes in path direction and reflectivity distribution. Multiple feature maps are generated through multiple convolutional layers, and each feature map represents a different level of abstraction of a feature.

[0152] Reduce the dimensionality of the feature map by using pooling layers;

[0153] Specifically, max pooling or average pooling is applied to reduce the size of the feature map while retaining the most important features, such as the most significant number of reflections or the material reflectivity region.

[0154] The feature map is converted into a ray path importance score through a fully connected layer, and effective ray paths are selected.

[0155] Specifically, the feature map processed by the convolutional and pooling layers is flattened into a one-dimensional feature vector, integrating information such as total path length, number of reflections, and material reflectivity. The fully connected layer calculates weights to convert this feature vector into a score (between 0 and 1), representing the contribution of the ray path to the initial rendered image. Based on the score, the ray path that contributes the most to the quality of the initial rendered image is selected for rendering.

[0156] Calculate the lighting contribution of light intensity in a 3D scene to obtain a high-quality lightmap;

[0157] Specifically, each light path starts from the light source, passes through the 3D scene, undergoes reflection and refraction, and finally reaches the camera for rendering;

[0158] The light intensity is obtained through the attenuation function, expressed as:

[0159]

[0160] Where I is the illuminance, representing the intensity of light from a light source at a specific distance at that point; I0 ​​is the initial illuminance of the light source, representing the illuminance of the light source without distance attenuation; and d... 2 The square of the distance is used because the light intensity spreads in three-dimensional space, following the relationship between the increase in area and the square of the distance (for example, the area covered by the light emitted by the light source increases quadratically with the increase in distance).

[0161] The intensity of light reaching the surface of the geometry is calculated based on the reflectivity of the material, expressed as:

[0162] I surf =I×R;

[0163] Among them, I surf Surf is the light intensity when it reaches the surface of the geometry. It represents the light intensity of the light on a specific surface after attenuation. Surf is the surface of the geometry, and R is the reflectivity of the geometry surface, which represents the proportion or ability of the geometry surface to reflect light. Its value is usually between 0 and 1. R=1 indicates complete reflection (such as a mirror), R=0 indicates no reflection of light (such as a completely absorbed black surface), and values ​​between 0 and 1 indicate partial reflection and partial absorption.

[0164] For each light path, the light intensity of each light path is calculated using the attenuation function to obtain the reflection and refraction properties of the material;

[0165] By simulating multiple reflections using Monte Carlo simulation and accumulating the contribution of light, the final illumination intensity is obtained, expressed as:

[0166] I path =I×∏R i ;

[0167] Among them, I path R represents the final light intensity after all reflections and refractions along the light path, where path is the entire light path, ∏ is the product operator, meaning multiplying all these reflectivities or transmissivities together. i The reflectivity of each reflection, where i is the index of the i-th reflection in the ray path;

[0168] For each pixel or geometric surface point of each high-quality lightmap, one or a series of high-quality lightmaps are generated by weighted accumulation of the contributions of all lighting paths for subsequent rendering.

[0169] It should be noted that by configuring a sampling density strategy and dynamically adjusting it using Monte Carlo random sampling and a convolutional neural network model, efficient and high-quality ray path selection and high-quality light map generation are achieved. First, a light intensity threshold is set based on the pre-rendering results, and the sampling density of highlight and low-light areas is intelligently adjusted to ensure detail capture while reducing computational resource consumption. Next, the Monte Carlo method is used to simulate ray paths, and effective ray paths are selected through convolutional neural network analysis, which improves rendering efficiency and rendered image quality. Finally, high-quality light maps are generated based on accurately calculated lighting contributions, enhancing the realism and visual effects of the scene, making 3D animation rendering more efficient and realistic.

[0170] S3. Perform preliminary image rendering using optimized geometry, enhanced geometry materials, and high-quality lightmaps to generate preliminary rendered image results.

[0171] Furthermore, preliminary image rendering is performed using optimized geometry and enhanced geometry materials;

[0172] It should be noted that the preliminary rendered image here is a comprehensive effect including geometry, materials and lighting. It is not the same as a lighting image, which is an evaluation of the lighting distribution and does not have the details of the materials.

[0173] Draw an initial image of each geometry using the graphics API;

[0174] Specifically, the optimized geometry vertex data is fed into the vertex buffer of the graphics API;

[0175] Use a graphics API (such as OpenGL or DirectX) to initiate drawing commands, and render the vertex data into an initial image for each geometry using a list of triangles or other primitive types;

[0176] By applying model matrix, view matrix, and projection matrix transformations, the geometry is transformed from local coordinates to world coordinates, ensuring that the geometry is drawn in the correct position and viewpoint;

[0177] A model matrix is ​​used to transform geometry from its defined local space to world space, including transformation operations such as translation, rotation, and scaling. For example, if you have a chair model, you can use a model matrix to place it in a certain position in the room and rotate it to face the table.

[0178] The view matrix transforms all objects in world space into the camera view space. It is created using the camera's position, orientation, and up vector to perform camera position transformations and rotations.

[0179] The projection matrix projects geometry from view space to clip space. There are two types: perspective projection and orthographic projection. Perspective projection is used to simulate the visual effect of the human eye, making distant objects appear smaller, while orthographic projection keeps the size of objects unchanged, regardless of distance.

[0180] Use vertex and fragment shaders to handle the basic appearance of the geometry, but do not consider complex lighting;

[0181] By combining the color, reflectivity, and transparency of the high-quality lightmap with the material data, the final color value of each pixel in the initial rendered image is calculated and applied to the initial image of each geometry (i.e., in the fragment shader of the rendering pipeline, the intensity I of the lightmap is combined with the material data of the geometry; these final color values ​​are applied to the pixels of the initial image, replacing or overwriting the original initial colors, forming an image with lighting and material effects). The expression is:

[0182] C f =(C b ×I×R)×A+C g ×(1-A);

[0183] Among them, C f For the final color value, C b The base color is a color sampled from the material map, where A is the transparency and C is the base color. g Background color refers to the background color at the current pixel or fragment position, and the color that needs to be mixed when considering transparency;

[0184] The initial images of all geometries are merged into an image frame, which serves as the initial rendered image result;

[0185] Specifically, the initial images of all rendered geometry are blended into a single image frame through a frame buffer, and basic anti-aliasing techniques (such as MSAA - multisampling anti-aliasing) are applied to smooth the image edges.

[0186] Multiple rendering targets are used to draw in layers, and then the final image is synthesized as the initial rendered image result;

[0187] It should be noted that by using optimized geometry and enhanced materials for initial rendering, combined with high-quality lightmaps, the realism and visual effects of the 3D scene can be significantly improved. The geometry is drawn efficiently using the graphics API, and the model, view, and projection matrices are applied to ensure correct spatial transformation and perspective representation. Color, reflectivity, transparency, and light intensity are comprehensively considered in the shader to accurately calculate the final color value of each pixel, achieving delicate material representation and lighting effects. Image frames are merged through frame buffers and anti-aliasing technology is used to smooth the edges, generating clear and realistic initial rendering image results.

[0188] S4. Perform semantic segmentation of the image using U-Net, and perform post-processing based on the semantic segmentation results and SRGAN to generate a high-quality rendered image.

[0189] Furthermore, U-Net is used to segment the initially rendered image, identify the semantic meaning of the human region, building region, sky region and background region in the initially rendered image, and generate segmentation results;

[0190] Specifically, a pre-trained U-Net model is used;

[0191] The format of the initially rendered image is converted into a size and format that the U-Net model can accept (such as normalization). Typically, the input to the U-Net model is a 3-channel (RGB) image.

[0192] The input preliminary rendered image is passed through a series of convolutional layers (encoders) to gradually reduce the resolution of the preliminary rendered image and extract features;

[0193] At the deepest level of the encoder, feature information is compressed;

[0194] The features are gradually restored to the resolution of the original initially rendered image through the upsampling layer (decoder).

[0195] The unique feature of the U-Net model is its "skip connection," which directly passes high-resolution feature maps from the encoding stage to the decoding stage, ensuring that details are preserved.

[0196] Finally, the class label segmentation map of each pixel is output, which is converted into a probability map through a softmax layer, and then a threshold or argmax is applied to determine the class of each pixel (such as human region, building region, sky region, and background region, etc.).

[0197] Based on the segmentation results, the human figure area is sharpened and the background area is blurred to enhance the sense of depth.

[0198] Specifically, a sharpening filter is applied to the segmented human figure area to enhance edge details, making the human figure appear clearer and more prominent in the image;

[0199] Apply a blur filter (such as Gaussian blur) to the background area to simulate a natural depth-of-field effect, making the background area blurry relative to the foreground and increasing the contrast between the foreground and the background.

[0200] Based on the image semantics of the people area, building area, sky area, and background area, adjust the hue, saturation, and brightness of the initially rendered image;

[0201] Specifically, different tone mappings are applied according to the semantics of each region (character region, building region, sky region, and background region). For example, the tone of the sky is adjusted to simulate the lighting effects at different times of the day.

[0202] Enhance or reduce the color saturation of specific areas, such as increasing the color saturation of buildings to make them more vivid, or reducing the saturation of the background to highlight the foreground;

[0203] Adjust the brightness to increase or decrease the visibility of specific areas, such as making architectural details stand out more in terms of brightness, or making the background darker to enhance the sense of depth.

[0204] The depth of field of the camera is simulated using the segmentation results;

[0205] Specifically, by combining the previous blurring and sharpening processes, the distribution of focus is precisely controlled using the segmentation results to ensure that foreground objects (such as people) are clear, while the background (such as buildings or part of the sky) gradually becomes blurred, simulating the depth-of-field effect of a camera in reality.

[0206] Use the generator in SRGAN to convert the initially rendered image (i.e., the low-resolution image) into a high-resolution image;

[0207] The discriminator in SRGAN is used to distinguish high-resolution images from real high-resolution images, enhancing the texture details and sharpness of the initially rendered images;

[0208] Output high-quality rendered images using the generator in SRGAN;

[0209] Specifically, SRGAN consists of a generator and a discriminator. The generator generates a high-resolution image from a preliminary rendered image (i.e., a low-resolution image), while the discriminator evaluates whether the generated preliminary rendered image looks realistic.

[0210] The image to be enhanced (which can be a pre-rendered image or a segmented pre-rendered image portion) is input into SRGAN to transform the pre-rendered image (i.e., the low-resolution image) into a high-resolution image.

[0211] SRGAN's generator learns to extract image features from the initial rendered input image through a deep convolutional network model, and uses these features to generate a higher resolution version.

[0212] Specifically, convolutional layers are used to extract image features from the initial rendered input image;

[0213] Increase the resolution of the initial rendered image by using an upsampling layer, while preserving or enhancing details;

[0214] Use residual learning to gradually increase detail and clarity;

[0215] The discriminator distinguishes between generated high-resolution images and real high-resolution images, helping the generator to generate more realistic details. Through this adversarial learning, the generator continuously improves its generation quality.

[0216] The initial rendered image or the rendered image with insufficient detail is input into the trained SRGAN generator to directly generate a high-resolution and high-quality rendered image with rich detail.

[0217] The generator outputs a high-quality rendered image, which has significantly improved texture detail and sharpness;

[0218] It should be noted that U-Net is used for image semantic segmentation to accurately identify and distinguish different regions (such as people, buildings, and the sky). The sharpness, blur, hue, and saturation of each region are adjusted accordingly to enhance the sense of depth and visual hierarchy. SRGAN technology is used to convert the preliminary rendered image or the rendered image with insufficient details into a high-resolution image, which significantly improves texture details and clarity, making the preliminary rendered image more realistic and delicate. This process not only enhances the overall visual effect of the preliminary rendered image, but also ensures the superiority of the preliminary rendered image in terms of detail through meticulous local processing and high-quality magnification, ultimately generating a high-quality rendered image that is both artistic and highly realistic.

[0219] This embodiment also provides a computer device applicable to the case of an AI-based 3D rendering analysis method for animated images, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI-based 3D rendering analysis method for animated images as proposed in the above embodiment.

[0220] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0221] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the artificial intelligence-based 3D rendering analysis method for animated images as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0222] In summary, this invention optimizes 3D scene rendering through deep learning algorithms, generative adversarial networks (GANs), and reinforcement learning, achieving efficient and intelligent 3D animation image rendering. First, deep learning algorithms are used to optimize geometry, extracting and aggregating shape features and other information, enhancing the model's detail and providing high-quality data support for subsequent steps, thus improving the realism of the rendered image. Second, GANs are used to enhance the geometric material, creating multiple variations that allow the material to exhibit natural effects under different lighting conditions, greatly enriching the object's representation and improving the quality of the rendered image. Finally, reinforcement learning is used to dynamically adjust the light source configuration, automatically optimizing the light source position and intensity based on the rendering results, reducing manual intervention, improving shadow and overexposure issues, and enhancing the natural harmony of the lighting effects.

[0223] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for analyzing 3D rendering of animated images based on artificial intelligence, characterized in that: include, The geometry in the 3D scene is optimized by using deep learning algorithms, the material of the geometry is enhanced by using generative adversarial network models, and the rendering light source of the 3D scene is optimized by using reinforcement learning. Configure a sampling density strategy, utilize Monte Carlo random sampling and a convolutional neural network model to dynamically adjust the sampling density strategy, generate and select effective light paths and high-quality light maps; Preliminary image rendering is performed using optimized geometry, enhanced geometry materials, and high-quality lightmaps to generate preliminary rendered image results; Image semantic segmentation is performed using U-Net, and post-processing is performed based on the semantic segmentation results and SRGAN to generate high-quality rendered images. The configured sampling density strategy utilizes Monte Carlo random sampling and a convolutional neural network model to dynamically adjust the sampling density strategy, generating and selecting effective ray paths and high-quality light maps. The specific steps are as follows: Based on the optimized rendering light source, pre-rendering is performed to generate a low-resolution light map, which serves as the basis for evaluating the light intensity in different areas of the 3D scene. Define a sampling density strategy and set a light intensity threshold; In areas with high light intensity, when the light intensity exceeds the light intensity threshold, the sampling density is reduced. In areas with low light intensity, if the light intensity does not exceed the light intensity threshold, the sampling density is increased. The sampling density strategy is adjusted based on real-time light source data during rendering; Light paths are randomly generated from light source data using Monte Carlo random sampling, simulating the light path that starts from the light source, passes through the geometric surfaces of the 3D scene, and reaches the light path rendered by the camera. The simulation of light paths is analyzed using a convolutional neural network model, and the effective light paths for rendering are selected. Calculate the lighting contribution of light intensity in a 3D scene to obtain a high-quality lightmap.

2. The AI-based 3D rendering analysis method for animated images as described in claim 1, characterized in that: The specific steps for optimizing the geometry in a 3D scene using deep learning algorithms are as follows: Using PointNet++, representative points are selected as center points in each layer of the 3D scene by sampling from distant points; Use a spherical neighborhood query to find the neighboring points around the center point and divide them into local regions. domain; Multilayer perceptron is used to extract the geometric shape features, geometric normal direction features, and geometric curvature features of neighboring points in a local region; Max pooling aggregates geometric shape features, geometric normal direction features, and geometric curvature features into global features.

3. The method for analyzing 3D rendering of animated images based on artificial intelligence as described in claim 2, characterized in that: The specific steps for enhancing the material of the geometry using a generative adversarial network model are as follows: Read the color, reflectivity, and transparency of the geometry as material data; Material data is used as a training set and input into the generative adversarial network model; Create multiple variations and use data augmentation to rotate, scale, and adjust the brightness of material data; The generator of the generative adversarial network model generates new material data based on the trained material data; The properties and textures of the new material data are reapplied to the geometry, replacing the original material.

4. The method for analyzing 3D rendering of animated images based on artificial intelligence as described in claim 3, characterized in that: The specific steps for optimizing the rendering light source of the 3D scene through reinforcement learning are as follows: Read light source data from the 3D scene using the rendering engine API; Define the state, actions, and rewards of the 3D scene environment; Select a deep Q-learning model, use light source data to generate a frame of lighting image, and render the lighting image in a 3D scene environment; The quality of the rendered lighting image is evaluated through image processing, and a reward is given, which is then fed back to the deep Q-learning model. If the deep Q-learning model detects that the shadow area of ​​the illuminated image is too large, then add a light source to fill the shadow area; If the deep Q-learning model detects that the lighting image is overexposed, it reduces the intensity of the light source; Adjust the position and direction of the light source based on the reflectivity and geometry of the rendered lighting image.

5. The AI-based 3D rendering analysis method for animated images as described in claim 4, characterized in that: The specific steps for analyzing and simulating light paths using a convolutional neural network model to select effective light paths for image rendering are as follows: Calculate the total path length of light rays from the light source to the camera rendering; When simulating the path of light rays, the number of reflections is counted to obtain the number of reflections. Record the material reflectivity for each reflection along the light path; The path length, number of reflections, and material reflectivity are combined to form the path feature; Define the path features as the training set of the convolutional neural network model and input them into the convolutional neural network model; Local features of the light path are extracted using convolutional layers and used as feature maps. Reduce the dimensionality of the feature map by using pooling layers; The feature map is converted into a ray path importance score through a fully connected layer, and the effective ray path is selected.

6. The AI-based 3D rendering analysis method for animated images as described in claim 5, characterized in that: The process involves using optimized geometry, enhanced geometry materials, and high-quality lightmaps for initial image rendering to generate a preliminary rendered image result. The specific steps are as follows: Initial image rendering is performed using optimized geometry and enhanced geometry materials; Draw an initial image of each geometry using the graphics API; The high-quality lightmap is combined with the color, reflectivity, and transparency of the material data to calculate the final color value of each pixel in the initial rendered image and apply it to the initial image of each geometry. The initial images of all geometries are merged into an image frame, which serves as the initial rendered image result.

7. The AI-based 3D rendering analysis method for animated images as described in claim 6, characterized in that: The process involves image semantic segmentation using U-Net, followed by post-processing based on the semantic segmentation results and SRGAN to generate a high-quality rendered image. The specific steps are as follows: U-Net is used to segment the initially rendered image, identify the semantic meaning of the human, building, sky and background regions in the initially rendered image, and generate segmentation results. Based on the segmentation results, the human figure area is sharpened, and the background area is blurred. Based on the image semantics of the people area, building area, sky area, and background area, adjust the hue, saturation, and brightness of the initially rendered image; The depth of field of the camera is simulated using the segmentation results; Use the generator in SRGAN to convert the initially rendered image into a high-resolution image; The discriminator in SRGAN is used to distinguish high-resolution images from real high-resolution images, enhancing the texture details and sharpness of the initially rendered images; High-quality rendered images are output through the generator in SRGAN.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the artificial intelligence-based three-dimensional rendering analysis method for animated images as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the AI-based three-dimensional rendering analysis method for animated images as described in any one of claims 1 to 7.

Images