Cross-platform rendering consistency verification method and system based on neural radiation field

Abstract

The invention discloses a cross-platform rendering consistency verification method and system based on a neural radiation field, and the method comprises the steps: carrying out the multi-view scene data collection of a target scene, training a NeRF model through hierarchical sampling and light termination probability prediction, and generating a standardized scene file; loading the standardized scene file on the target platform, and rendering a target screenshot; carrying out resolution matching on the target screenshot and the reference screenshot of the standard platform through an adaptive sampler, and eliminating geometric and color differences through a feature alignment algorithm; according to the invention, the NeRF is utilized to generate a standardized scene-containing file, a cross-platform unified physical real reference is provided, and traditional manual comparison without a quantitative standard is replaced; matching resolution is adaptively sampled, platform rendering differences are eliminated in combination with a feature alignment algorithm, automatic quantitative comparison is realized, and manual intervention is reduced; meanwhile, the sensitivity of a traditional method to resolution / color space is avoided through the standardized process, and the problem of low detection efficiency is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of graphics rendering, in particular to a cross-platform rendering consistency verification method and system based on neural radiance field. BACKGROUND

[0002] In the prior art, traditional rendering tests have many problems: strong platform dependency, dependence on manual comparison, difficulty in quantifying differences; lack of a cross-platform standard rendering scene library based on physical reality, i.e., lack of a benchmark; existing image comparison methods (such as SSIM, PSNR) are sensitive to resolution scaling and color space conversion, and have low detection efficiency.

[0003] Therefore, under the cross-platform rendering scene, how to achieve consistency verification and ensure that multiple terminal users obtain the same visual experience is a technical problem to be solved at present. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a cross-platform rendering consistency verification method and system based on neural radiance field, which solves the problems of benchmarking, platform dependency, and low detection efficiency.

[0005] To solve the above technical problems, the technical solution adopted by the present application is: A cross-platform rendering consistency verification method based on neural radiance field, comprising the steps of: S1, multi-view scene data acquisition is performed on a target scene, a NeRF model is trained through hierarchical sampling and ray termination probability prediction, and a standardized scene file is generated; S2, the standardized scene file is loaded on a target platform, and a target screenshot is rendered; S3, the target screenshot is matched with a benchmark screenshot of a standard platform in resolution through an adaptive sampler, and geometric and color differences are eliminated through a feature alignment algorithm.

[0006] To solve the above technical problems, another technical solution adopted by the present application is: A cross-platform rendering consistency verification system based on neural radiance field, comprising: A neural radiance field benchmark construction module: multi-view scene data acquisition is performed on a target scene, a NeRF model is trained through hierarchical sampling and ray termination probability prediction, and a standardized scene file is generated; A cross-platform screenshot comparison module: the standardized scene file is loaded on a target platform, and a target screenshot is rendered; the target screenshot is matched with a benchmark screenshot of a standard platform in resolution through an adaptive sampler, and geometric and color differences are eliminated through a feature alignment algorithm.

[0007] The beneficial effects of the present application are that the neural radiation field-based cross-platform rendering consistency verification method and system of the present application generates a standardized scene file containing NeRF, provides a cross-platform unified physical real benchmark, and replaces the traditional non-quantitative standard manual comparison; adaptive sampling matches the resolution, combines the feature alignment algorithm (geometry + color) to eliminate platform rendering differences (such as resolution, color gamut), realizes automatic quantitative comparison, and reduces manual intervention; at the same time, the standardized process avoids the sensitivity of the traditional method to the resolution / color space, and solves the problem of low detection efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A brief process example diagram of a neural radiation field-based cross-platform rendering consistency verification method of an embodiment of the present application; Figure 2 A process example diagram of a neural radiation field-based cross-platform rendering consistency verification method of an embodiment of the present application; Figure 3 A structure example diagram of a neural radiation field-based cross-platform rendering consistency verification system of an embodiment of the present application. DETAILED DESCRIPTION

[0009] To explain the technical content, purposes and effects of the present application in detail, the following will be described in conjunction with the embodiments and the accompanying drawings.

[0010] Please refer to Figure 1 and Figure 2 A neural radiation field-based cross-platform rendering consistency verification method, comprising the steps of: S1, multi-view scene data acquisition is performed on a target scene, a NeRF model is trained through hierarchical sampling and ray termination probability prediction, and a standardized scene file is generated; S2, the standardized scene file is loaded on a target platform, and a target screenshot is rendered; S3, the target screenshot is matched with a reference screenshot of a standard platform through an adaptive sampler, and geometric and color differences are eliminated through a feature alignment algorithm.

[0011] From the above description, the beneficial effects of the present application are that the neural radiation field-based cross-platform rendering consistency verification method of the present application generates a standardized scene file containing NeRF, provides a cross-platform unified physical real benchmark, and replaces the traditional non-quantitative standard manual comparison; adaptive sampling matches the resolution, combines the feature alignment algorithm (geometry + color) to eliminate platform rendering differences (such as resolution, color gamut), realizes automatic quantitative comparison, and reduces manual intervention; at the same time, the standardized process avoids the sensitivity of the traditional method to the resolution / color space, and solves the problem of low detection efficiency.

[0012] Further, the step S1 comprises the steps of: S11, multi-view scene data acquisition is performed on the target scene, and the NeRF model is input; S12, according to the multi-view scene data, the light depth point is sampled through a hierarchical sampling strategy; S13, the termination probability of each sampling point is predicted through the MLP network, and the light ray tracing is terminated in advance if the cumulative termination probability is greater than a preset value; S14, a joint loss function based on color loss and depth loss is used to measure the difference between the predicted value and the true value, the MLP parameters are updated based on the loss gradient, and iterative optimization is performed until convergence; S15, a standardized scene file containing geometry, material and illumination information is output.

[0013] As can be seen from the above description, coarse sampling quickly locates high-density areas, fine sampling focuses on details, and the light termination strategy is combined to reduce redundant calculation and shorten the training time; the joint loss function optimizes color and depth at the same time, reduces geometric error, and ensures the physical reality of the standardized scene.

[0014] Further, the joint loss function is specifically: ; Wherein, r represents a single light ray; represents the predicted color value of the light ray r by the NeRF model; C ( r ) represents the true color value corresponding to the light ray r; represents the predicted depth value of the light ray r by the NeRF model; D ( r ) represents the true depth value corresponding to the light ray r; represents the weight coefficient of the depth loss.

[0015] As can be seen from the above description, the weight of color and depth error is adjusted by λ to avoid model bias (such as optimizing color only and ignoring geometry), and the comprehensive accuracy of scene reconstruction is ensured to provide a basis for cross-platform geometric consistency detection (such as depth difference).

[0016] Further, the elimination of geometric and color differences through the feature alignment algorithm comprises the steps of: S31, using SIFT to extract key points, and eliminating affine transformation differences through RANSAC; S32, using a color gamut mapping function to unify the color space.

[0017] Further, the step S31 is specifically: SIFT algorithm is used to extract key points from the target screenshot and the reference screenshot, and an affine transformation matrix is fitted through RANSAC algorithmM : ; satisfy: ; in,( x , y (The coordinates are used as the reference for the screenshot.) The coordinates of the target screenshot.

[0018] Furthermore, step S32 specifically includes: Construct a color gamut mapping function to convert the color space of the target screenshot to the sRGB space of the reference screenshot: ; in, W It is a 3×3 color gamut conversion matrix; b The bias term is obtained by fitting the color patch calibration data of the two images using the least squares method; C dev Represents the original color values ​​in the color space of the target screenshot; C std This represents the standard color values ​​in the sRGB space of the baseline screenshot.

[0019] As described above, SIFT extracts key points, RANSAC fits the affine transformation matrix to eliminate translation / rotation / scaling differences, and ensures pixel-level geometric consistency; the target screenshot color space is converted to the reference sRGB through the color gamut mapping function (W matrix + b bias) to avoid misjudgment caused by device color gamut differences and improve detection robustness.

[0020] Furthermore, it also includes the following steps: S4. Combine structural similarity with perceptual hashing to calculate local dissimilarity, and superimpose geometric error and apparent error to generate a multidimensional heat map.

[0021] Furthermore, combining structural similarity with perceptual hashing to calculate local dissimilarity includes: The baseline screenshot and the target screenshot are divided into local blocks, and a structural similarity value (SSIM) is calculated for each local block. ; in, Indicates the ( x , y The average pixel value of each local block; Indicates the ( x , y The pixel variance of each local block; This represents the covariance of the local block at the same location in the baseline screenshot and the target screenshot; C 1 andC 2 represents a stability constant; scaling and grayscale conversion are performed on the reference screenshot and the target screenshot, discrete cosine transform is calculated, and low-frequency coefficients are taken, and a hash value corresponding to the number of bits is generated; The Hamming distance is calculated for the hash values of the reference screenshot and the target screenshot, and normalized to obtain a difference value: ; wherein, d pHash represents the difference value; d h represents the Hamming distance of the hash values of the reference screenshot and the target screenshot; m represents the number of bits of the hash value; For each local block, the comprehensive difference is calculated by weighting: ; And threshold judgment is performed, if: ; The corresponding local block is marked as an abnormal block; wherein, d ab represents a preset abnormal threshold; Merge adjacent abnormal blocks to form a connected region, determine the coordinates of the abnormal region, and determine the error level according to the comprehensive difference.

[0022] Further, superimposing geometric error and apparent error to generate a multi-dimensional heat map includes the steps of: Obtaining a pixel-level depth map from the NeRF model D NeDF , comparing the depth map rendered by the target platform D render , determining the geometric error E geo : ; wherein, scale represents a depth dimension conversion coefficient; Calculating color error based on CIE Lab color space E color : ; wherein, L 1, a 1 and b 1 represent the brightness and chroma components of the CIE Lab color space of the reference screenshot; L 2, a 2 and b2 represents the brightness and chromaticity components of the CIE Lab color space of the target screenshot; The geometric error E geo is mapped to the red channel, the color error E color is mapped to the green channel, superimposed to generate a multi-dimensional heat map: ; ; wherein, represents a transparency coefficient; R represents a red channel value; G represents a green channel value; OriginalImage represents a reference image.

[0023] As can be seen from the above description, the SSIM captures structural similarity, the pHash perceives semantic features, and the weighted calculation comprehensively calculates the difference, covering pixel-level and semantic-level differences, avoiding single indicator missing detection; the geometric error (red) and the color error (green) are superimposed to generate a heat map, supporting hierarchical display and interactive query, intuitively displaying error distribution, and improving abnormal area positioning efficiency.

[0024] A cross-platform rendering consistency verification system based on a neural radiation field comprises: A neural radiation field reference construction module: multi-view scene data acquisition is performed on a target scene, a NeRF model is trained through hierarchical sampling and ray termination probability prediction, and a standardized scene file is generated; A cross-platform screenshot comparison module: the standardized scene file is loaded on a target platform, and a target screenshot is rendered; the target screenshot is compared with a reference screenshot of a standard platform, the resolution is matched through an adaptive sampler, and geometric and color differences are eliminated through a feature alignment algorithm.

[0025] As can be seen from the above description, the cross-platform rendering consistency verification system based on the neural radiation field has the advantages that: the system generates a standardized scene file containing NeRF, provides a cross-platform unified physical real reference, and replaces traditional non-quantitative standard manual comparison; the resolution is matched through adaptive sampling, and platform rendering differences (such as resolution and color gamut) are eliminated through a feature alignment algorithm (geometry + color), so that automatic quantitative comparison is realized and manual intervention is reduced; at the same time, the standardized process avoids the sensitivity of traditional methods to resolution / color space, and solves the problem of low detection efficiency.

[0026] The cross-platform rendering consistency verification method and system based on the neural radiation field are suitable for cross-platform graphics rendering testing.

[0027] Please refer to Figure 1 andFigure 2 Embodiment one of the present application is: A neural radiance field-based cross-platform rendering consistency verification method, comprising the steps of: S1, multi-view scene data acquisition is performed on the target scene, a NeRF model is trained through hierarchical sampling and ray termination probability prediction, and a standardized scene file is generated; Step S1 includes the steps of: S11, multi-view scene data acquisition is performed on the target scene, and the NeRF model is input; In this embodiment, RGB-D images and pose data of the scene are collected by a camera array or a virtual camera to realize multi-view scene data acquisition.

[0028] S12, according to the multi-view scene data, the light depth points are sampled through a hierarchical sampling strategy; In this embodiment, a hierarchical sampling strategy and ray termination probability prediction are used to generate a high-precision radiance field model. The hierarchical sampling strategy is: Coarse sampling stage: 64 points are uniformly sampled along the light rays to preliminarily estimate the density distribution of the scene.

[0029] Fine sampling stage: according to the coarse sampling result, 128 points are densely sampled in the area with high density (such as the surface of the object) to optimize the geometric details.

[0030] S13, the termination probability of each sampling point is predicted through an MLP network, and the ray tracing is terminated in advance if the cumulative termination probability is greater than a preset value; In this embodiment, the ray termination probability prediction is: The termination probability T(t) of each sampling point is predicted through an MLP network, which represents the possibility of the light ray terminating at depth t.

[0031] Dynamic adjustment of the sampling interval: if the cumulative termination probability satisfies: ; That is, if the cumulative termination probability is greater than 0.99, the ray tracing is terminated in advance to reduce redundant calculation.

[0032] S14, a joint loss function based on color loss and depth loss is used to measure the difference between the predicted value and the true value, the MLP parameters are updated based on the loss gradient, and iterative optimization is performed until convergence; The joint loss function is specifically: ; Wherein, r represents a single light ray; represents the predicted color value of the light ray r by the NeRF model; C ( r ) represents the true color value corresponding to the light ray r. This represents the predicted depth value of ray r by the NeRF model; D ( r ) represents the true depth value corresponding to ray r; The weighting coefficients represent the depth loss. In this embodiment, the weighting coefficient =0.1, this method reduces training time by 30% and geometric error to 0.08mm (compared to 0.25mm in the traditional method).

[0033] S15. Output a standardized scene file containing geometry, material, and lighting information.

[0034] Standard scene generation: Outputs a standardized scene file containing geometry, material, and lighting information.

[0035] S2. Load the standardized scene file on the target platform and render the target screenshot; In this embodiment, the baseline screenshot is an image rendered on a reference platform (such as a high-performance PC) of a standardized scene generated by NeRF. The target screenshot is an image of the same scene rendered on the platform under test (such as a mobile device or AR glasses).

[0036] S3. Take the target screenshot and compare it with the benchmark screenshot of the standard platform, match the resolution with the adaptive sampler, and eliminate geometric and color differences through the feature alignment algorithm; In this embodiment, the rendering sampling rate is dynamically adjusted according to the target platform resolution.

[0037] Eliminating geometric and color differences using feature alignment algorithms includes the following steps: S31. Use SIFT to extract key points and use RANSAC to eliminate affine transformation differences. Step S31 is as follows: The SIFT algorithm is used to extract key points (such as edges and corners) from the target screenshot and the reference screenshot, and the affine transformation matrix is ​​fitted using the RANSAC algorithm. M : ; satisfy: ; in,( x , y (The coordinates are used as the reference for the screenshot.) The coordinates of the target screenshot.

[0038] In-point filtering: Only retain matching points with a reprojection error of less than 1.5 pixels (in-point ratio ≥ 80%).

[0039] S32. Use a color gamut mapping function to unify the color space; Step S32 is specifically: A color gamut mapping function is constructed to convert the target screenshot color space to the sRGB space of the reference screenshot: ; Wherein, W is a 3x3 color gamut conversion matrix; b is a bias term, obtained by fitting the color block calibration data of the two pictures by least squares method; C dev represents the original color value of the target screenshot color space; C std represents the standard color value of the sRGB space of the reference screenshot.

[0040] S4, combined with structural similarity and perceptual hash to calculate local difference, and superimposed geometric error and apparent error to generate a multi-dimensional heat map; Combined with structural similarity and perceptual hash to calculate local difference includes: The reference screenshot and the target screenshot are divided into local blocks, and the structural similarity value SSIM is calculated for each local block: ; Wherein, represents the pixel mean value of the (i, j)th local block; x , y represents the pixel mean value of the (i, j)th local block; represents the pixel variance of the (i, j)th local block; x , y represents the pixel variance of the (i, j)th local block; represents the covariance of the local blocks at the same position of the reference screenshot and the target screenshot; C 1 and C 2 represent stable constants; In this embodiment, the reference picture and the target picture are divided into local blocks of 32x32 pixels (stride=16, overlap 50%). The SSIM value is calculated for each block. Among them, C 1=(0.01L) 2 , C 2=(0.03L) 2 , L=255, is the pixel value range.

[0041] The reference screenshot and the target screenshot are scaled and converted into grayscale pictures, the discrete cosine transform is calculated, the low frequency coefficients are taken, and the hash value corresponding to the bit number is generated; In this embodiment, the image is scaled to 32x32 pixels and converted into a grayscale picture. The discrete cosine transform (DCT) is calculated, and the upper left 8x8 low frequency coefficients are taken. A 64-bit hash value is generated: if the coefficient is greater than the mean value, it is 1, otherwise it is 0.

[0042] The Hamming distance is calculated for the hash values of the reference screenshot and the target screenshot, and is normalized to obtain a difference value: ; wherein, d pHash represents the difference value; d h represents the Hamming distance of the hash values of the reference screenshot and the target screenshot; m represents the number of bits of the hash value; For each local block, the comprehensive difference is calculated by weighting: ; And threshold judgment is performed, if: ; The corresponding local block is marked as an abnormal block; wherein, d ab represents a preset abnormal threshold; Merge adjacent abnormal blocks to form a connected region, determine the coordinates of the abnormal region, and determine the error level according to the comprehensive difference; In this embodiment, adjacent abnormal blocks are merged to form a connected region, and the coordinates of the minimum circumscribed rectangle and the maximum d local As the error level (0.15-0.3 is a warning, and >0.3 is serious).

[0043] Superimposing geometric error and apparent error to generate a multi-dimensional heat map includes the steps of: Obtaining a pixel-level depth map from the NeRF model D NeDF Comparing the depth map rendered by the target platform D render Determine the geometric error E geo : ; wherein, scale represents a depth dimension conversion coefficient; Calculate the color error based on the CIE Lab color space E color : ; wherein, L 1、 a 1 and b 1 represent the luminance and chrominance components of the CIE Lab color space of the reference screenshot; L 2、 a 2 and b2 indicates the luminance and chromaticity components of the target screenshot in the CIE Lab color space; In this embodiment, a threshold ΔE>5 is set as a significant difference (perceptible to the human eye).

[0044] The geometric error E geo Mapped to the red channel, the color error E color Mapped to green channels, overlaid, to generate a multidimensional heatmap: ; ; in, Indicates the transparency coefficient; R Indicates the red channel value; G Indicates the green channel value; OriginalImage This represents the reference image.

[0045] In this embodiment, an interactive heatmap was generated, and the generation method is as follows: 1. Data Preprocessing Store the difference detection results in GeoJSON format, including: { "type": "FeatureCollection", "features": [ { "geometry": {"type": "Polygon", "coordinates": [[x1,y1],...[xn,yn]]}, "properties": { "error_level": 0.45, "error_type": "geometry", / / or "color" "details": {"SSIM": 0.82, "pHash": 0.28, "depth_diff": 1.2mm} } } ] } 2. Visual Engine Construction WebGL rendering: Use Canvas to draw the original image as the underlying layer.

[0046] Generating heatmaps using WebGL shaders: Vertex Shader: Convert GeoJSON coordinates to screen position.

[0047] Fragment Shader: Mix colors according to error level and type (geometric error -> red gradient, color error -> blue gradient).

[0048] Transparency Control: Non-linear mapping of error level.

[0049] 3. Interactive Function Implementation Event Listening: Mouse Hover: Detect heat map area by Raycasting, display Tooltip (contains error value and type).

[0050] Click Event: Trigger Ajax request, load local zoom-in image (200x200 pixels, stored as Base64).

[0051] Layered Display Control: Error type overlay is achieved through CSS blend mode (mix-blend-mode: multiply).

[0052] Checkbox controls layer visibility (e.g., show only geometric errors).

[0053] 4. Performance Optimization Quadtree index is used to accelerate spatial queries, supporting 4K image interaction response within ≤10ms.

[0054] Heat map data compression: RLE (Run-Length Encoding) encoding for adjacent similar error areas, reducing volume by 70%.

[0055] Example Interaction Flow: User clicks on a red highlighted area in the heat map.

[0056] System returns a local zoom-in comparison chart for the area (left: reference image, right: target platform image).

[0057] Drag the slider to view the difference details, and the right panel displays quantitative indicators (SSIM=0.76, ΔE=6.2).

[0058] Embodiment Two of the Invention: A neural radiance field-based cross-platform rendering consistency verification method, which differs from Embodiment One in that this embodiment describes a specific application scenario for a neural radiance field-based cross-platform rendering consistency verification method: Mobile game adaptation testing: normal mapping Figure 1 Consistency detection: Implementation steps: Data preparation phase: Import character model in Unity engine, use 12-megapixel industrial camera array (4096x2160 resolution) to capture 360° RGB-D images and pose data (every 5° one frame).

[0059] Generate benchmark normal map (precision 0.01mm) through NeRF training, output as.nerf format standardized file.

[0060] Cross-platform rendering: Android side: call Adreno 650 GPU's Vulkan API, set anisotropic filtering level to MipmapLevel 4.

[0061] iOS side: call Apple A15 GPU's Metal API, enable the same Mipmap Level parameter.

[0062] Render 4K resolution character screenshots (PNG format, 32-bit color depth) on both platforms.

[0063] Difference detection: Feature alignment: use SIFT algorithm to extract normal map feature points (≥500 per image), calculate affine transformation matrix through RANSAC (inlier rate threshold 80%).

[0064] Error calculation: calculate the angle between the aligned normal vectors pixel by pixel (formula: θ=arccos(mathbf{n}_1·mathbf{n}_2)), set threshold θ≤2° as qualified.

[0065] Heat map generation: red indicates θ>2° area (transparency 50%), green indicates θ≤0.5° area (transparency 20%).

[0066] Embodiment three of Benmin House is: A cross-platform rendering consistency verification method based on neural radiation field, which is different from embodiment one in that this embodiment describes a specific application scenario of a cross-platform rendering consistency verification method based on neural radiation field: AR glasses rendering verification: transparent material refractive index difference detection: Implementation steps: Benchmark scene construction: Create a virtual object with glass material in Blender (refractive index preset as 1.33), generate a.vol file containing refractive light path data through NeRF training.

[0067] Load the scene on Hololens 2 and Magic Leap 2 devices respectively, simulate an environment with a natural light incident angle of 45°.

[0068] Light tracing rendering: Hololens 2: Use the WMR rendering pipeline, call the refractive index parameter IOR=1.33.

[0069] Magic Leap 2: Enable Lumin SDK, set IOR=1.35 (simulate device differences).

[0070] Intercept the edge area of the transparent material (200x200 pixels) for 10x supersampling anti-aliasing (SSAAx10).

[0071] Difference quantization: Key point detection: Establish a 10x10 pixel ROI at the center of the refracted light spot, calculate the centroid coordinate offset (formula: Δd=sqrt{(x_2-x_1)^2+(y_2-y_1)^2}).

[0072] Criteria: Δd≤2 pixels is qualified, >2 pixels triggers a warning (error level 3).

[0073] Data output: Generate a JSON report recording each light spot ID, offset, and corresponding material area UV coordinates.

[0074] Embodiment four of Benminfang is: A cross-platform rendering consistency verification method based on neural radiation field, which is different from embodiment one. In this embodiment, a specific application scenario of a cross-platform rendering consistency verification method based on neural radiation field is described: Edge aliasing detection for cloud game streaming QA: Implementation steps: Resolution adaptation: Render the original 4K scene on the cloud server (NVIDIA A100), and generate a 1080P downsampled image using an adaptive sampler (Lanczos3 algorithm).

[0075] The client receives 1080P streaming data and locally upsamples it to 4K (bicubic interpolation, parameter b=0.75).

[0076] Jagged detection algorithm: Edge extraction: Apply Canny edge detection to the original 4K and upsampled 4K images respectively (low threshold 50, high threshold 150).

[0077] Difference comparison: Calculate the Hausdorff distance of two edge maps (formula: H(A, B) = max{h(A, B), h(B, A)}), set threshold H≤3 pixels.

[0078] Jagged score: Divide the level according to the jagged density (the number of jagged points per 100x100 pixel area): 0-5 (excellent), 6-10 (warning), >10 (failure).

[0079] Visual output: Heat map overlay: Overlay a semi-transparent blue mask on the edge area of the game screen, and the higher the jagged density, the deeper the blue.

[0080] Interactive function: Click the heat map area to view the local enlarged image and the original-contrast image slider control.

[0081] Please refer to Figure 3 , the fifth embodiment of the present application is: A cross-platform rendering consistency verification system based on neural radiation field, comprising: NeRF reference construction module: multi-view scene data acquisition is performed on the target scene, a NeRF model is trained through hierarchical sampling and ray termination probability prediction, and a standardized scene file is generated; Cross-platform screenshot comparison module: load the standardized scene file on the target platform and render the target screenshot; compare the target screenshot with the reference screenshot of the standard platform, match the resolution through an adaptive sampler, and eliminate geometric and color differences through a feature alignment algorithm.

[0082] Difference heat map generation module: calculate the local difference degree by combining structural similarity and perceptual hash, and generate a multi-dimensional heat map by superimposing geometric error and apparent error.

[0083] The steps in the cross-platform rendering consistency verification method based on neural radiation field according to any one of the above embodiments one to four are realized by the NeRF reference construction module, the cross-platform screenshot comparison module and the difference heat map generation module.

[0084] In summary, the cross-platform rendering consistency verification method and system based on neural radiation field provided by the present application use NeRF to generate a standardized scene file, provide a unified physical real reference across platforms, and replace the traditional non-quantitative standard manual comparison; adaptive sampling matches the resolution, and combines a feature alignment algorithm (geometry + color) to eliminate platform rendering differences (such as resolution, color gamut), realize automatic quantitative comparison, and reduce manual intervention; at the same time, the standardized process avoids the sensitivity of traditional methods to resolution / color space, and solves the problem of low detection efficiency.

[0085] The above merely illustrates the embodiments of the present application, and is not intended to limit the patent scope of the present application. Any equivalent transformation or direct or indirect application in the related technical field based on the content of the present application specification and drawings is also included in the patent protection scope of the present application.

Claims

1. A cross-platform rendering consistency verification method based on neural radiation fields, characterized in that, Including the following steps: S1. Collect multi-view scene data of the target scene, train the NeRF model through hierarchical sampling and ray termination probability prediction, and generate a standardized scene file. S2. Load the standardized scene file on the target platform and render the target screenshot; S3. Take the target screenshot and compare it with the benchmark screenshot of the standard platform, match the resolution with an adaptive sampler, and eliminate geometric and color differences through a feature alignment algorithm.

2. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 1, characterized in that, Step S1 includes the following steps: S11. Collect multi-view scene data of the target scene and input it into the NeRF model; S12. Based on the multi-view scene data, sample the light depth points using a layered sampling strategy; S13. Predict the termination probability of each sampling point through an MLP network, and terminate ray tracing in advance if the cumulative termination probability is greater than a preset value. S14. Measure the difference between the predicted value and the true value using the joint loss function based on color loss and depth loss, update the MLP parameters based on the loss gradient, and iterate until convergence. S15. Output a standardized scene file containing geometry, material, and lighting information.

3. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 2, characterized in that, The joint loss function is specifically as follows: ； Where r represents a single ray; This represents the color value predicted by the NeRF model for ray r; C ( r ) represents the true color value corresponding to ray r; This represents the predicted depth value of ray r by the NeRF model; D ( r ) represents the true depth value corresponding to ray r; The weighting coefficients represent the depth loss.

4. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 1, characterized in that, Eliminating geometric and color differences using feature alignment algorithms includes the following steps: S31. Use SIFT to extract key points and use RANSAC to eliminate affine transformation differences. S32. Use a color gamut mapping function to unify the color space.

5. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 4, characterized in that, Step S31 is as follows: Key points were extracted from the target screenshot and the reference screenshot using the SIFT algorithm, and the affine transformation matrix was fitted using the RANSAC algorithm. M : ； satisfy: ； in,( x , y (The coordinates are used as the reference for the screenshot.) The coordinates of the target screenshot.

6. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 4, characterized in that, Step S32 is as follows: Construct a color gamut mapping function to convert the color space of the target screenshot to the sRGB space of the reference screenshot: ； in, W It is a 3×3 color gamut conversion matrix; b The bias term is obtained by fitting the color patch calibration data of the two images using the least squares method; C dev Represents the original color values ​​in the color space of the target screenshot; C std This represents the standard color values ​​in the sRGB space of the baseline screenshot.

7. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 1, characterized in that, It also includes the following steps: S4. Combine structural similarity with perceptual hashing to calculate local dissimilarity, and superimpose geometric error and apparent error to generate a multidimensional heat map.

8. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 7, characterized in that, Combining structural similarity and perceptual hashing to calculate local dissimilarity includes: The baseline screenshot and the target screenshot are divided into local blocks, and a structural similarity value (SSIM) is calculated for each local block. ； in, Indicates the ( x , y The average pixel value of each local block; Indicates the ( x , y The pixel variance of each local block; This represents the covariance of the local block at the same location in the baseline screenshot and the target screenshot; C 1 and C 2 represents the stability constant; The baseline screenshot and the target screenshot are scaled and converted to grayscale, the discrete cosine transform is calculated, the low-frequency coefficients are taken, and the corresponding bit hash value is generated. For the hash values ​​of the baseline screenshot and the target screenshot, calculate the Hamming distance, and normalize it to obtain the difference value: ； in, d pHash Indicates the degree of difference; d h The Hamming distance between the hash values ​​of the baseline screenshot and the target screenshot; m Indicates the number of bits in the hash value; For each local block, calculate the weighted overall difference: ； And perform threshold determination, if: ； The corresponding local block is then marked as an abnormal block; in, d ab This indicates the preset abnormal threshold. Adjacent abnormal blocks are merged to form a connected region, the coordinates of the abnormal region are determined, and the error level is determined based on the comprehensive difference degree.

9. The cross-platform rendering consistency verification method based on neural radiation fields according to claim 7, characterized in that, The steps involved in generating a multidimensional heatmap by superimposing geometric and apparent errors are as follows: Obtaining pixel-level depth maps from NeRF models D NeDF Compare the depth map rendered by the target platform. D render Determine geometric error E geo : ； in, scale Indicates the depth dimension conversion coefficient; Color error calculated based on CIE Lab color space E color : ； in, L 1. a 1 and b 1 indicates the luminance and chromaticity components of the CIE Lab color space in the reference screenshot; L 2. a 2 and b 2 indicates the luminance and chromaticity components of the target screenshot in the CIE Lab color space; The geometric error E geo Mapped to the red channel, the color error E color Mapped to green channels, overlaid, to generate a multidimensional heatmap: ； ； in, Indicates the transparency coefficient; R Indicates the red channel value; G Indicates the green channel value; OriginalImage This represents the reference image.

10. A cross-platform rendering consistency verification system based on neural radiation fields, characterized in that, include: Neural Radiation Field Benchmark Construction Module: Collects multi-view scene data of the target scene, trains the NeRF model through hierarchical sampling and ray termination probability prediction, and generates standardized scene files; Cross-platform screenshot comparison module: Loads the standardized scene file on the target platform and renders the target screenshot; matches the resolution of the target screenshot with the benchmark screenshot of the standard platform through an adaptive sampler, and eliminates geometric and color differences through a feature alignment algorithm.

Images