A method for constructing a multi-angle RGB image and a specification coordinate graph data set of a ship outfitting part

Abstract

The present application relates to the field of three-dimensional modeling and data set construction of ship outfitting parts, and specifically provides a ship outfitting part multi-angle RGB image and specification coordinate map data set construction method. The method comprises: performing direction label analysis and normalization processing on the outfitting part three-dimensional model based on a deep learning model, optimizing the view angle configuration file and generating multi-view configuration; adjusting the light source intensity and distribution according to the material characteristics of the outfitting part, and rendering the RGB image and the specification coordinate map; analyzing the geometric consistency and direction matching of the rendering result through multi-modal consistency checking, and generating an abnormal data list; expanding the diversity of the data set in combination with a data enhancement module; and finally evaluating the multi-view coverage and consistency of the data set, feeding back the direction label generation model and the view angle configuration file for optimization, and forming a closed-loop improvement mechanism. The present application can automatically generate high-quality multi-angle RGB image and specification coordinate map data sets, ensure the consistency and geometric alignment of the images, effectively support three-dimensional reconstruction tasks and deep learning model training, and improve the quality and adaptability of the data set.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of computer graphics, three-dimensional modeling and deep learning, and in particular to a method for constructing a multi-angle RGB image and a standard coordinate graph dataset of a ship outfitting part. BACKGROUND

[0002] In ship design and manufacturing, outfitting parts, as an important part of the ship, have complex geometric shapes and diverse material properties. Traditional three-dimensional modeling methods usually rely on manual modeling and rendering, especially in the details and geometric structures of outfitting parts, manual rendering is difficult to ensure the consistency and high-precision three-dimensional reconstruction effect of the image. In addition, in the generation process of multi-view images and standard coordinate graphs of ship outfitting parts, accurate rendering and checking need to be carried out for different angles, lighting and materials to ensure the geometric consistency between multi-angle images and the accurate matching of direction labels.

[0003] With the application of deep learning technology in three-dimensional reconstruction, automated and efficient data generation becomes particularly important. Especially for the generation of multi-angle images of ship outfitting parts, it is necessary to ensure the high quality and geometric consistency of the images to meet the needs of training data for deep learning models. However, existing technologies still face great challenges in multi-angle dataset generation, image consistency checking, geometric alignment and direction label matching. How to automatically generate accurate multi-angle RGB images and standard coordinate graph datasets and ensure the high quality and consistency of the datasets is still a key problem that needs to be solved in the current technical field.

[0004] The present application introduces an automated rendering and checking mechanism and combines deep learning technology to propose a new method that can efficiently generate high-quality multi-angle RGB images and standard coordinate graph datasets of outfitting parts, providing high-quality data support for three-dimensional reconstruction tasks and deep learning model training. SUMMARY

[0005] The present application provides a method for constructing a multi-angle RGB image and a standard coordinate graph dataset of a ship outfitting part to solve the problem of how to automatically generate multi-angle RGB images and standard coordinate graph datasets of a ship outfitting part to ensure the consistency, accurate geometric alignment and matching of direction labels of the images to support high-quality three-dimensional reconstruction tasks and deep learning model training.

[0006] To solve the above technical problems, the present application provides a method for constructing a multi-angle RGB image and a standard coordinate graph dataset of a ship outfitting part, comprising:

[0007] Based on a deep learning model, direction label analysis is performed on the three-dimensional model of the outfitting part, and the three-dimensional model is normalized to generate direction labels and default position parameters;

[0008] Based on the direction label and the default position parameter, the view angle configuration file is optimized, the view angle, focal length and position parameter of the camera are set, and a multi-view configuration of the outfitting part is generated;

[0009] Based on the view angle configuration file and the material characteristics of the outfitting part, the intensity and distribution of the light source are adjusted, the material of the high-reflective area is optimized, and a uniform light distribution scheme is generated;

[0010] Based on the view angle configuration file and the light distribution scheme, the RGB image and the specification coordinate image of the outfitting part are rendered in a loop, and the light consistency and geometric alignment are checked in real time during the rendering process to generate preliminary checking data;

[0011] The preliminary checking data and the rendered image are subjected to multi-modal consistency checking, the geometric consistency and direction matching of the rendering result are analyzed by using a deep learning model, and an abnormal data list is generated;

[0012] Based on the rendered image and the direction label, a standardized output file is generated, including RGB image, specification coordinate image and direction label; and through a data enhancement module, the diversity of the outfitting part dataset is expanded;

[0013] The multi-view coverage and specification coordinate consistency of the dataset are comprehensively evaluated, and the evaluation results are used to optimize the direction label generation model and the view angle configuration file, forming a closed-loop improvement mechanism for data generation. Figure One

[0014] Further, before the step of analyzing the direction label of the outfitting part three-dimensional model based on the deep learning model, it further includes:

[0015] The three-dimensional model of the outfitting part is imported from multiple standard three-dimensional model formats, and the geometric structure and material information of the model are extracted.

[0016] Further, the step of analyzing the direction label based on the deep learning model specifically includes:

[0017] The geometric features of the three-dimensional model are analyzed, the direction labels corresponding to the front, top and side of the model are generated, and the normalization scale factor is generated according to the size of the model.

[0018] Further, the step of optimizing the view angle configuration file specifically includes:

[0019] The six fixed view angles of the outfitting part are determined through the direction label, and the view angle, focal length and position of the camera are adjusted based on the size and geometric characteristics of the outfitting part.

[0020] Further, the step of adjusting the intensity and distribution of the light source specifically includes: ​

[0021] Based on the material characteristics of the outfitting, the position, intensity and angle of the light source are optimized to reduce the overexposure problem of the reflective area and adjust the light distribution.

[0022] Further, the step of rendering the RGB image and the canonical coordinate image of the outfitting in cycles includes:

[0023] According to the view angle configuration file and the light distribution scheme, the RGB image and the canonical coordinate image of each view angle are rendered using automated rendering control, and the light consistency and geometric alignment are verified.

[0024] Further, the step of multi-modal consistency verification includes:

[0025] The deep learning model is used to analyze the geometric consistency and the matching degree of the direction label between images, and to generate abnormal data about view angle deviation, geometric inconsistency and direction error.

[0026] Further, the step of the data augmentation module includes:

[0027] The rendered RGB image and the canonical coordinate image are subjected to light simulation, material transformation and geometric shape enhancement to generate new data set variants.

[0028] Further, the step of evaluating the multi-view coverage and canonical coordinate consistency of the comprehensive evaluation data set includes: Figure One

[0029] The image coverage area under each view angle is calculated and evaluated to verify the geometric alignment between views and the consistency of the canonical coordinate image, and an evaluation report is generated and fed back to the direction label generation model and the view angle configuration file optimization.

[0030] Further, a ship outfitting multi-angle RGB image and canonical coordinate image data set construction system includes:

[0031] The data acquisition module imports the three-dimensional model of the outfitting from multiple standard three-dimensional model formats, extracts the geometric structure and material information of the three-dimensional model, performs normalization processing, and generates direction labels and default position parameters;

[0032] The view angle configuration and parameter optimization module optimizes the view angle configuration file based on the direction labels and default position parameters, sets the view angle, focal length and position parameters of the camera, and generates the multi-view configuration file of the outfitting;

[0033] The light and material optimization module adjusts the intensity and distribution of the light source according to the material characteristics and geometric characteristics of the outfitting, optimizes the material of the high-reflective area, and generates a uniform light distribution scheme;

[0034] ​A rendering and verification module, based on the view configuration file and the illumination distribution scheme, cyclically renders the RGB image and the standard coordinate image of the outfitting part, verifies the illumination consistency and geometric alignment in real time, and generates preliminary verification data;

[0035] A multimodal consistency check module performs a multimodal consistency check on the preliminary verification data and the rendered image, uses a deep learning model to analyze the geometric consistency and directional matching of the rendering result, and generates a list of abnormal data;

[0036] A data set generation and enhancement module generates an RGB image, a standard coordinate image, and an orientation label based on the rendered image and the orientation label, and expands the diversity of the outfitting parts data set through a data enhancement module;

[0037] Closed-loop optimization and feedback module to comprehensively evaluate the multi-view coverage and standard coordinates of the dataset Figure One The evaluation results are used to optimize the direction label generation model and view configuration files, forming a closed-loop improvement mechanism for data generation, promoting continuous optimization of the system and improvement of dataset quality.

[0038] The key innovations of the present invention include:

[0039] (1) Automatic generation of multi-angle RGB images and standard coordinate images: Through automated viewing angle configuration and lighting optimization, the uniformity and consistency of multi-angle images under different conditions are ensured.

[0040] (2) Deep learning-assisted directional label analysis and geometric verification: The deep learning model is used to verify the geometric consistency and directional label matching of the rendering results, which improves the accuracy of the data.

[0041] (3) Closed-loop improvement mechanism: Through the evaluation and feedback optimization of the data set, a self-improving data generation closed loop is formed to ensure that the data quality is continuously improved to meet the high standards of 3D modeling and deep learning.

[0042] The main beneficial effects are as follows: the application provides an automatic ship outfitting multi-angle RGB image and specification coordinate graph dataset construction method, which can efficiently generate high-quality three-dimensional reconstruction datasets. Compared with traditional manual modeling and rendering methods, the application introduces a deep learning model to analyze direction labels, automatically optimize view angle configuration and camera parameters, and accurately control light and material, greatly improving the consistency of images, geometric alignment and accurate matching of direction labels, reducing errors and inconsistencies in human operation. In addition, through the automatic multi-view rendering and real-time verification mechanism, the application can monitor the rendering quality in real time to ensure that the images under different angles are accurately matched. The introduction of the data enhancement module further expands the diversity of the dataset and improves the training effect and adaptability of the deep learning model. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of a ship outfitting multi-angle RGB image and specification coordinate graph dataset construction method provided by an embodiment of the application is shown in the figure.

[0044] Figure 2 A structural block diagram of a ship outfitting multi-angle RGB image and specification coordinate graph dataset construction method provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the description herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the description and claims of this application as well as the above discussion of the related art are intended to be illustrative only and are not intended to be limiting upon the scope of the application; the terminology used in the description and the claims of this application and the above discussion of the related art all are intended to be given the broadest possible interpretation; the terms "comprising" and "having," and variations thereof, are intended to cover all possible inclusions and are not intended to be limiting; the terms "first," "second," and so on, are used to distinguish different objects and are not intended to describe a particular sequential order.

[0046] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. It is expressly understood that the embodiments described herein are merely examples from a much larger number of embodiments that can be claimed.

[0047] Embodiment One: Reference Figure 1 A flowchart of an end-to-end low-latency image transmission method provided by an embodiment of the application is shown in the figure, which can at least include steps S100-S700:

[0048] S100, based on the deep learning model, the direction label analysis is performed on the outfitting piece three-dimensional model, and the three-dimensional model is normalized to generate the direction label and the default position parameter.

[0049] S200, based on the direction label and the default position parameter, the view angle configuration file is optimized, the view angle, focal length and position parameter of the camera are set, and the multi-view configuration of the outfitting piece is generated.

[0050] S300, based on the view angle configuration file and the material characteristics of the outfitting piece, the intensity and distribution of the light source are adjusted, the material optimization of the high-reflective area is performed, and the uniform light distribution scheme is generated.

[0051] S400, based on the view angle configuration file and the light distribution scheme, the RGB image and the specification coordinate image of the outfitting piece are rendered in a loop, and the light consistency and geometric alignment are checked in real time during the rendering process to generate preliminary checking data.

[0052] S500, the multi-modal consistency of the preliminary checking data and the rendered image is checked, the deep learning model is used to analyze the geometric consistency and direction matching of the rendering result, and an abnormal data list is generated.

[0053] S600, based on the rendered image and the direction label, a standardized output file is generated, including RGB image, specification coordinate image and direction label; and the diversity of the outfitting piece data set is expanded through the data enhancement module.

[0054] S700, the multi-view coverage and specification coordinate Figure One consistency of the data set are comprehensively evaluated, and the evaluation results are used to optimize the direction label generation model and the view angle configuration file, forming a closed-loop improvement mechanism for data generation.

[0055] Step S100 includes at least steps S110-S130:

[0056] S110: Obtain the input three-dimensional model from the outfitting piece three-dimensional model import module, and perform coordinate normalization on the model to obtain a unitized outfitting piece model.

[0057] First, the three-dimensional model input data is obtained from the outfitting piece three-dimensional model import module, and the input data is an.obj or.gl b file supporting standard three-dimensional model format, which contains the geometric structure and material information of the outfitting piece.

[0058] Specifically, the input three-dimensional model is loaded into a default rendering environment, and the vertex coordinate data thereof is parsed to extract the boundary range of the model, including the minimum boundary point (min_x, min_y, min_z) and the maximum boundary point (max_x, max_y, max_z).

[0059] The size ratio of the outfitting part model is calculated based on the above boundary range, and the maximum boundary length is normalized to a unit length, while the center position of the model is adjusted to the coordinate origin (0, 0, 0) to form a unitized outfitting part model. During the unitization process, the generated scale factor and center offset value are recorded for subsequent steps.

[0060] S120: The direction of the normalized outfitting part model is analyzed using a deep learning model to generate a direction label for the outfitting part.

[0061] The vertex coordinates and center offset value of the unitized outfitting part model are obtained from the normalization output of S110 as input data for the deep learning model.

[0062] The normalized outfitting part model is analyzed for structural features using a pre-trained geometric feature direction classification model. The model determines the front direction, top direction, and side direction of the model by analyzing the geometric shape, vertex distribution, and edge characteristics of the outfitting part.

[0063] Further, the direction information is labeled as a direction label (front, top, side), and is output in the form of a label bound to the normalized model, providing a basis for direction correction for subsequent modules.

[0064] S130: Based on the direction label, the outfitting part model is adjusted to the default position and direction, and the processed three-dimensional model and direction label are output.

[0065] The direction label of the outfitting part is obtained from S120, and combined with the center offset value of the normalization output, the orientation and position of the outfitting part model are adjusted.

[0066] Specifically, according to the direction label, the front direction of the outfitting part model is corrected to the positive direction of the x-axis, the top direction is corrected to the positive direction of the z-axis, and the model is adjusted to the origin (0, 0, 0) position through the offset value.

[0067] The output of the processed outfitting part model and its direction label is output as input data for S200 module. The output includes the corrected three-dimensional geometric model, normalization scale factor, direction label, and position offset parameter.

[0068] Connection between modules: S110→S120: The unitized outfitting part model and normalization scale factor generated by S110 are directly input into S120 for geometric feature analysis and direction label generation.

[0069] S120→S130: The direction label generated by S120 is the key input for S130, and the direction label determines the rotation correction direction of the model in three-dimensional space.

[0070] S130→S200: The correction model and direction label output by S130 are used as the basis data for S200 to optimize the view angle configuration. The direction information is used to determine the specific settings of the six fixed view angles.

[0071] Through the above process, each step of the S100 module has a clear input and output relationship, forming a complete logical closed loop.

[0072] Step S200 includes at least steps S210-S230:

[0073] S210: Obtain the direction label and normalization model from the S100 module, and set the six fixed view angles (front, back, left, right, up, down) of the outfitting model.

[0074] Obtain the normalization data and direction label of the outfitting model from the S100 module. The normalization data includes the unitized model size and its center position, and the direction label includes the front, top, and side directions.

[0075] Specifically, based on the direction label, the positions of the six fixed view angles are determined, where the front view corresponds to the front of the model, the back view corresponds to the back of the model, the left and right views correspond to the side directions of the model, and the up and down views correspond to the top and bottom of the model, respectively.

[0076] Further, according to the size of the normalization model, the coordinate values and orientation of the camera in the three-dimensional space are initially set to ensure that the camera view angle can cover the entire model, providing basic position parameters for subsequent view angle optimization.

[0077] S220: Based on the size characteristics and direction label of the outfitting, optimize the focal length, position, and view angle parameters of the camera to ensure the uniformity of multi-view coverage.

[0078] Obtain the preliminary parameters of the six fixed view angles from the S210 module, including the initial position and direction of the camera, and the size range of the normalization model.

[0079] Specifically, according to the maximum boundary length of the outfitting model, the camera focal length is optimized to ensure that the view angle range contains the complete contour of the model, and the distance between the camera and the model is appropriately adjusted to avoid occlusion or view angle distortion problems.

[0080] Further, based on the direction label, the orientation angle of the camera is corrected so that the line of sight of each view angle is perpendicular to the model surface, ensuring that the projections of the model in the six view angles are uniformly distributed, providing parameter support for subsequent rendering consistency.

[0081] S230: Generate a multi-view configuration file containing view angle name, camera position, focal length, and angle parameters, which serves as input for subsequent modules.

[0082] The optimized camera parameters obtained from the S220 module include the names of the six perspectives, the three-dimensional spatial positions of the cameras, the focal lengths, the orientation angles, and the perspective ranges.

[0083] Specifically, the parameter information of each perspective is stored in the form of structured data as a multi-perspective configuration file, which contains the perspective name (such as front view, rear view, etc.), camera position coordinates (x, y, z), focal length (unit length), and rotation angle value.

[0084] The multi-perspective configuration file is output, providing input basis for the light optimization and rendering module, ensuring the consistency and parameter reproducibility of multi-perspective data.

[0085] The connection between modules is described as follows: S210→S220: The six fixed perspective preliminary parameters generated by S210, including camera position and direction, are directly input to S220 for further optimization of perspective range and focal length.

[0086] S220→S230: The camera focal length, position, and perspective parameters optimized by S220 are used as core data for the generation of configuration files by S230.

[0087] S230→S300: The multi-perspective configuration file generated by S230 is directly input to the S300 module for determining the light distribution and rendering perspective.

[0088] Through the above step description, the S200 module forms a close connection from the acquisition of direction labels to the generation of perspective configuration files, and provides clear data input support for subsequent modules.

[0089] Step S300 includes at least steps S310-S330:

[0090] S310: Obtain the perspective configuration file from the S200 module, analyze the material characteristics of the outfitting model, and determine the initial values of light source distribution and intensity.

[0091] From the perspective configuration file obtained from the S200 module, the camera position, orientation, and focal length parameters of the six fixed perspectives are parsed as input conditions for light source distribution analysis.

[0092] From the normalized model of the outfitting, further analyze its material characteristic data, including surface material type (such as metal, composite material) and its reflection characteristic parameters.

[0093] Based on the geometric relationship between the material characteristics and the perspective configuration file, the distribution position and intensity of the light source are preliminarily determined to ensure that the model is uniformly illuminated under the six fixed perspectives, providing basic parameters for subsequent optimization steps.

[0094] S320: Based on the geometric characteristics of the outfitting part model and the material reflection parameters, optimize the light source angle and intensity, reduce overexposure and shadow problems.

[0095] From the S310 module, obtain the initial distribution parameters of the light source and the geometric characteristics data of the outfitting part model, analyze the complex geometric structure (such as hollow design, protruding part) and the shielding relationship between the model and the light source.

[0096] Specifically, according to the reflection characteristics parameters of the model surface, adjust the angle and intensity distribution of the light source to ensure that the light can uniformly cover the model surface, avoid the light intensity of the overexposed area being too high and the light intensity of the shadow area being insufficient.

[0097] Further, combined with the material reflectivity, special light compensation is made for metal areas to reduce the influence of metal highlight reflection on the consistency of RGB images, and the projection effect of light in the hollow area is optimized.

[0098] S330: Generate a lighting distribution scheme file containing light source position, intensity and angle parameters, and output to the rendering module

[0099] From the S320 module, obtain the optimized light source position, intensity and angle parameters, and integrate them into a structured lighting distribution scheme file, which contains light source arrangement data corresponding to six fixed viewing angles.

[0100] Specifically, the lighting distribution scheme file contains the following contents:

[0101] The light source position coordinates (x, y, z) of each fixed viewing angle.

[0102] The intensity value of the light source, which adapts to the model material characteristics and geometric distribution.

[0103] The illumination angle and light diffusion range parameters of the light source.

[0104] Output the lighting distribution scheme file to the rendering module to ensure that subsequent rendering can generate high-quality RGB images and standard coordinate maps based on the optimized lighting environment.

[0105] Interconnection between modules: S310→S320: The initial distribution and intensity parameters of the light source determined by S310 are directly used as input conditions for S320 to optimize the light source angle and intensity.

[0106] S320→S330: The light source position, angle and intensity parameters optimized by S320 are used as core data for S330 to generate a lighting distribution scheme file.

[0107] S330→S400: The lighting distribution scheme file generated by S330 is directly input to the S400 module to guide the rendering of multi-view RGB images and standard coordinate maps.

[0108] Through the above step description, the S300 module realizes the complete process from the initial distribution analysis of the light source to the optimization of the light source parameters and the generation of the lighting scheme file, and forms a close connection with the preceding and subsequent modules.

[0109] The step S400 at least includes steps S410-S430:

[0110] S410: Obtain the view angle configuration file and the lighting distribution scheme from the S200 and S300 modules respectively, load the outfitting model and initialize the rendering environment.

[0111] Obtain the view angle configuration file from the S200 module, which includes the camera position, focal length and rotation angle information of the six fixed view angles.

[0112] Obtain the lighting distribution scheme from the S300 module, which includes the light source position, intensity and angle parameters as the basis for lighting arrangement in the rendering environment.

[0113] Specifically, load the normalized three-dimensional model of the outfitting into the rendering environment, set the camera parameters based on the view angle configuration file, and initialize the light source according to the lighting distribution scheme to ensure that the view angle and lighting environment of the model are consistent with the previous steps.

[0114] S420: Use automatic loop rendering control to render RGB images and standard coordinate images for each view angle, and real-time check the lighting consistency and geometric alignment.

[0115] Call the automatic rendering control script to loop through the six fixed view angles in the view angle configuration file, and load the corresponding camera parameters and lighting scheme one by one.

[0116] At each view angle, first render the RGB image of the outfitting, specifically map the surface color, light and texture information of the outfitting to the image output; then render the standard coordinate image through coordinate mapping technology to ensure that the three-dimensional position of the outfitting is accurately converted to unit cube coordinates.

[0117] Further, real-time check the consistency of the lighting distribution and the geometric alignment during each rendering process, specifically including:

[0118] Check the intensity distribution of the light source under the six view angles to determine whether there are overexposed or insufficient light areas.

[0119] Check the alignment degree of the geometric feature points of the outfitting in the rendered image to ensure that the image projections under different view angles are consistent.

[0120] S430: Generate the rendered RGB image and standard coordinate image, as well as preliminary check data (such as lighting distribution and alignment information)

[0121] Get the six perspective RGB images and canonical coordinate images from the S420 module, and store them in the specified path according to the perspective naming rules to ensure the correspondence between the images and the perspectives.

[0122] Specifically, generate preliminary verification data according to the real-time verification results, which include the illumination distribution information (such as intensity value distribution) of each perspective, the geometric alignment measurement value, and the integrity check record of the rendered image.

[0123] Output the RGB images, canonical coordinate images, and preliminary verification data to the next module for subsequent verification and data management.

[0124] The connection between modules is explained as follows: S410→S420: The model loaded by S410 and the rendering environment initialization parameters (perspective configuration file and illumination distribution scheme) are directly used as input conditions for the automatic loop rendering of S420.

[0125] S420→S430: The RGB images and canonical coordinate images rendered by S420 and the real-time verification results are directly used to generate preliminary verification data by S430.

[0126] S430→S500: The preliminary verification data and image files generated by S430 are used as input for further analysis by the subsequent multi-modal consistency verification module.

[0127] Through the above step description, S400 module completes rendering environment initialization, automatic loop rendering, and preliminary verification data generation based on perspective configuration file and illumination scheme, and realizes data flow and logical connection with subsequent modules.

[0128] Step S500 includes at least steps S510-S530:

[0129] S510: Obtain preliminary verification data and rendered image sequence from S400 module, perform multi-perspective geometric feature matching, and verify the consistency between images.

[0130] Obtain preliminary verification data from S400 module, which includes illumination distribution information of each perspective, geometric alignment measurement value, and RGB images and canonical coordinate images generated by rendering.

[0131] Load the RGB images and canonical coordinate images, and based on multi-perspective geometric feature matching technology, compare the spatial consistency of corresponding geometric feature points in six perspective images one by one. Specifically, extract the feature point distribution of the model under different perspectives through feature point matching algorithm.

[0132] Further, verify the geometric feature alignment between the six perspectives, generate a preliminary consistency verification report, including the feature offset value between each pair of perspectives and the image matching quality score.

[0133] S520: Analyze the matching degree of the direction label and the image sequence using the deep learning model, and determine whether there is an abnormal viewing angle or direction deviation.

[0134] Obtain the feature offset value and image matching score from the consistency check report and preliminary check data output from the S510 module, and combine the direction label generated by the S400 module.

[0135] Load the pre-trained deep learning model, input the image sequence of six viewing angles and their direction labels, and analyze the matching degree of each viewing angle image and direction label. Specifically, detect whether there is a direction error or viewing angle deviation problem.

[0136] Further, according to the analysis result of the deep learning model, mark the abnormal viewing angle and potential direction deviation, including the offset angle size, error type (such as viewing angle deviation, light abnormality, etc.) and the degree of inconsistency with the direction label.

[0137] S530: Output the abnormal data list, including abnormal viewing angle, direction deviation and geometric consistency problem description.

[0138] Integrate the verification results of S510 and S520 modules to generate an abnormal data list, which contains the following contents:

[0139] Abnormal viewing angle name (such as front view, left view, etc.).

[0140] Geometric consistency problem description, including the offset amount of feature points between viewing angles and matching score.

[0141] Direction deviation description, including the offset angle of viewing angle and direction label and deviation type.

[0142] Specifically, format the above abnormal data into a structured file, including the image file name of the abnormal viewing angle, matching score and offset details, and output to the subsequent module for further verification and repair.

[0143] Synchronize the abnormal data list to the data management module to record the verification log and provide traceability basis for data set quality management.

[0144] Connection between modules: S510→S520: The geometric feature matching check report and consistency score generated by S510 are directly used as the input condition for S520 deep learning model to analyze direction deviation.

[0145] S520→S530: The abnormal viewing angle and direction deviation analysis result marked by S520 is used for S530 to generate the abnormal data list.

[0146] S530→S600: The abnormal data list output by S530 is input as a check log for subsequent data management and optimization and processing of the enhancement module.

[0147] Through the above step description, the S500 module completely implements a closed-loop verification process from preliminary verification data loading, multi-view feature matching, deep learning direction analysis to abnormal data output, and maintains close connection with the front and rear modules.

[0148] Step S600 includes at least steps S610-S630:

[0149] S610: Integrate the rendered images generated by S400 and the direction labels generated by S100 to generate standardized output files (including RGB images, normalized coordinate images, and direction information).

[0150] From the S400 module, a sequence of rendered images of six views is obtained, including RGB images and normalized coordinate images; at the same time, the corresponding direction labels are obtained from the S100 module.

[0151] Specifically, the RGB images, normalized coordinate images, and direction labels are integrated into a standardized file, which is indexed by the unique identifier (U ID) of the fitting part, and records the image file name and direction information corresponding to each view.

[0152] Further, the rendered images are stored according to the naming rules, wherein the RGB images are named rgb_<view number>.png, the normalized coordinate images are named nocs_<view number>.png, and the direction labels and fitting part classification information are added in the associated file.

[0153] S620: Use the data enhancement module to diversify the RGB images and normalized coordinate images, including light simulation and material transformation, to expand the diversity of the data set.

[0154] Load the RGB images and normalized coordinate images from the standardized output files generated by the S610 module and input them into the data enhancement module for diversification.

[0155] Specifically, based on the light simulation technology, the intensity and distribution of the light source in the RGB image are adjusted to generate image variants under different lighting conditions; based on the material transformation technology, the surface material properties (such as metal gloss, surface roughness) of the fitting part are modified to generate images in a multi-material scene.

[0156] Further, the geometric shape enhancement algorithm is used to randomly scale and rotate the normalized coordinate images to generate variant normalized coordinate images, ensuring that the enhanced data set can cover more training scene requirements.

[0157] S630: Output the final dataset file, including the enhanced multi-view images, canonical coordinate maps and direction labels.

[0158] From the S620 module, obtain the enhanced RGB image and canonical coordinate map, and from the S610 module, obtain the direction label, and integrate them into the final dataset file.

[0159] Specifically, the enhanced image is stored in the specified path according to the standardized naming rules, and the associated file records the U ID of each fitting, the enhanced variant file name, the direction label and the enhancement type (such as illumination condition, material transformation).

[0160] Output the final dataset file, including the multi-view enhanced image, canonical coordinate map and direction label, and upload the dataset file to the data management system for subsequent deep learning model calling and verification.

[0161] The connection between modules is explained as follows: S610→S620: The standardized files (RGB image, canonical coordinate map and direction label) generated by S610 are used as input conditions for S620 data enhancement module.

[0162] S620→S630: The enhanced image and variant information from S620 are directly input into S630 module to generate the final dataset file.

[0163] S630→S700: The final dataset file and its description information output by S630 are used as the core input for S700 module quality evaluation.

[0164] Through the above step description, S600 module realizes the complete process from integration of rendered image and direction label, data enhancement processing to final dataset output, and forms a strict logical connection with the previous and subsequent modules.

[0165] Step S700 includes at least steps S710-S730:

[0166] S710: Obtain the generated dataset from S600 module, and analyze the consistency of multi-view coverage and canonical coordinate map.

[0167] From the S600 module, obtain the generated dataset, including the enhanced multi-view RGB image, canonical coordinate map and direction label, and load it into the evaluation module for coverage and consistency analysis.

[0168] Specifically, based on the view coverage analysis algorithm, the RGB image of six fixed views is evaluated to detect whether there are missing geometric features between views, and further analyze whether it meets the requirement of full view coverage.

[0169] Meanwhile, the geometric feature points of the outfitting parts are extracted from the standard coordinate graph, the consistency thereof in multi-view projection is verified, and potential inconsistent areas, such as alignment errors of the feature points caused by light or rendering deviation, are recorded.

[0170] S720: Based on the evaluation result, feedback optimization is performed on the direction label generation model and the view configuration file, and the model and the configuration parameters are updated.

[0171] The view coverage analysis result and the consistency deviation data of the standard coordinate graph are obtained from the evaluation report output by the S710 module.

[0172] According to the deviation data, the direction label generation model is further optimized, specifically including adjusting the geometric feature extraction strategy of the model to reduce the problem of insufficient multi-view coverage caused by inaccurate direction label generation.

[0173] Meanwhile, based on the evaluation result, the view configuration file is optimized, and the camera focal length, rotation angle and position parameters are updated to ensure that the view coverage in subsequent rendering is more uniform and to reduce the omission of geometric features caused by improper camera configuration.

[0174] S730: Complete the closed-loop optimization, and apply the optimized parameters and model to the next round of data set construction process.

[0175] The optimized parameters and the updated direction label generation model output by the S720 module are loaded into the data generation process for the construction of the next round of data set.

[0176] Specifically, the optimized view configuration file is input into the S200 module to reset the view, focal length and position parameters of the camera, and the optimized direction label generation model is applied to the S100 module to improve the accuracy of direction label generation.

[0177] Output the model and parameter version record after closed-loop optimization to ensure that the next round of data set generation process is based on the latest optimization result, and record the feedback log to track the optimization effect.

[0178] The connection between the modules is explained as follows: S710→S720: The view coverage evaluation result of S710 and the consistency deviation data of the standard coordinate Figure One are directly used as the input conditions for feedback optimization of the model and the configuration file of S720.

[0179] S720→S730: The optimized model and parameters of S720 are used as the input of S730 closed-loop optimization to ensure the update and application of the next round of data generation process.

[0180] S730→S100 / S200: The optimization results of S730 act on the S100 direction label generation module and the S200 view angle configuration module, forming a closed-loop improvement of data generation.

[0181] Through the above step description, the S700 module realizes the complete logical chain from data set evaluation, optimization feedback to closed-loop process update, effectively connecting the data flow and improvement mechanism with the front and rear modules.

[0182] The key innovations of the present application include:

[0183] (1) Automatically generate multi-angle RGB images and standard coordinate images: Through automatic view angle configuration and light optimization, the uniformity and consistency of multi-angle images under different conditions are ensured.

[0184] (2) Deep learning assisted direction label analysis and geometric verification: Use deep learning model to check the geometric consistency and direction label matching of the rendering result, improve the accuracy of the data.

[0185] (3) Closed-loop improvement mechanism: Through the evaluation and feedback optimization of the data set, a self-improving data generation closed loop is formed, ensuring continuous improvement of data quality and meeting the high standard requirements of three-dimensional modeling and deep learning.

[0186] The main beneficial effects are as follows: The present application provides an automatic ship outfitting multi-angle RGB image and standard coordinate image data set construction method, which can efficiently generate high-quality three-dimensional reconstruction data set. Compared with traditional manual modeling and rendering methods, the present application introduces deep learning model to analyze direction label, automatically optimizes view angle configuration and camera parameters, and accurately controls light and material, greatly improving the consistency, geometric alignment and accurate matching of direction label of the image, reducing the error and inconsistency in human operation. In addition, through the automatic multi-view rendering and real-time verification mechanism, the present application can monitor the rendering quality in real time to ensure accurate matching of images under different angles. The introduction of data enhancement module further expands the diversity of data set and improves the training effect and adaptability of deep learning model.

[0187] Embodiment two: Figure 2 A structural block diagram of a ship outfitting multi-angle RGB image and standard coordinate image data set construction method according to an embodiment of the present application is shown. As shown in Figure 2 The structure can include:

[0188] The data acquisition module 10 is responsible for extracting original data from the outfitting three-dimensional model and converting it into a format that can be used for subsequent processing. The specific functions include:

[0189] Import the 3D model of the fitting part from various standard 3D model formats (such as.obj and.glb), ensuring data diversity and universality.

[0190] Position normalization is performed on the 3D model, adjusting the fitting part model to a unified coordinate system and determining the model's direction label (such as front, top, side, etc.).

[0191] The processed model and material information are input into the subsequent module to ensure data consistency and accuracy.

[0192] The view configuration and parameter optimization module 20 is responsible for setting appropriate viewing angles and camera parameters based on the 3D model of the fitting part and its direction label obtained from the data acquisition module 10, to ensure uniform coverage of the model at different viewing angles. The specific functions of this module include:

[0193] Based on the direction label, six fixed viewing angles (front, back, left, right, top, bottom) are set, and corresponding camera positions, focal lengths, and rotation angles are configured for each viewing angle.

[0194] Optimize the distribution of viewing angles in combination with the geometric characteristics and model size of the fitting part, to ensure that all viewing angles can completely cover the geometric features of the fitting part.

[0195] Generate an optimized view configuration file for input to the subsequent rendering module.

[0196] The lighting and material optimization module 30 is responsible for adjusting the distribution and intensity of light sources based on the material properties and geometric features of the fitting part, optimizing rendering quality, and ensuring image consistency. The specific functions include:

[0197] Analyze the material properties of the fitting part (such as metal surface reflection, surface roughness, etc.), and optimize the lighting intensity and light source position based on the material properties.

[0198] Special processing is performed on high-reflectivity areas and complex geometric structures (such as hollow structures) to avoid overexposure or insufficient shadows.

[0199] Generate a lighting distribution scheme and output it to the subsequent rendering module to ensure uniformity and consistency of light sources during the rendering process.

[0200] The rendering and verification module 40 completes the rendering of the RGB image and the specification coordinate image of the fitting part based on the aforementioned view configuration file and lighting distribution scheme, and performs real-time verification on the rendering results to ensure image quality and geometric consistency. The specific functions of this module include:

[0201] Call the automated rendering control system to cyclically render the RGB image and the specification coordinate image based on the optimized view configuration file and the lighting distribution scheme.

[0202] During the rendering process, real-time verification of lighting consistency and geometric alignment is performed to ensure that images from different perspectives have consistent lighting distribution and geometric characteristics.

[0203] The output rendering results and preliminary verification data are provided for further verification by subsequent verification modules.

[0204] The multi-modal consistency verification module 50 is responsible for consistency verification between rendered images and standard coordinate maps. Deep learning algorithms are used to analyze the geometric alignment and matching degree of directional labels between images. The functions of this module include:

[0205] Feature points are extracted from the image sequence obtained from the rendering and verification module 40 for geometric consistency verification.

[0206] Deep learning models are used to analyze the matching of directional labels and image sequences to determine whether there are abnormal perspectives or directional deviations.

[0207] An abnormal data list is output, including abnormal perspectives, directional deviations, and geometric consistency issues, for subsequent data correction and optimization.

[0208] The data set generation and enhancement module 60 is responsible for integrating rendered images, standard coordinate maps, and directional labels, and expanding the diversity of the data set through data enhancement methods. The functions include:

[0209] Rendered images, corresponding directional labels, and standard coordinate maps are integrated into standardized data set files to ensure data format consistency.

[0210] Diverse training data is generated through data enhancement methods such as lighting simulation, material transformation, and geometric transformation to improve the diversity and adaptability of the data set.

[0211] The final data set file is output for deep learning model training and further use.

[0212] The closed-loop optimization and feedback module 70 continuously monitors and evaluates each module in the data set construction process to ensure continuous optimization of the system. The functions of this module include:

[0213] The final data set is obtained from the data set generation and enhancement module 60, and the multi-perspective coverage and consistency of the standard coordinate map are evaluated.

[0214] Based on the evaluation results, the directional label generation model and perspective configuration file are fed back for optimization to ensure that the next round of data set generation can provide higher quality images and annotations.

[0215] After completing the closed-loop optimization, the optimized parameters and models are applied to the next round of data set construction process to achieve continuous improvement of the system.

[0216] The ship outfitting piece multi-angle RGB image and specification coordinate map data set construction method system provided by the application can efficiently and accurately generate a high-quality multi-view data set that meets the training needs of deep learning.

[0217] The system has the following significant benefits:

[0218] 1. Automated data generation: Through automated steps such as view configuration, lighting optimization, and material rendering, manual intervention is reduced, and the efficiency and consistency of data generation are improved.

[0219] 2. High-precision verification: The multi-modal consistency verification module uses deep learning technology to accurately verify the rendering results, ensuring the geometric alignment and direction label consistency of the images.

[0220] 3. Data diversity enhancement: Through the data enhancement module, a diversified data set containing different lighting, materials, and view changes is generated, improving the robustness of the model.

[0221] 4. Closed-loop optimization mechanism: Based on the data set evaluation results, the direction label generation model and the view configuration file are optimized, realizing a closed-loop improvement mechanism for data generation, and continuously improving data quality.

[0222] The system not only supports the deep learning tasks of ship outfitting piece three-dimensional reconstruction, but also has strong scalability and can be applied to three-dimensional reconstruction and model training tasks in other fields.

[0223] Obviously, the above-described embodiments are only part of the embodiments of the present application, not all the embodiments, and the preferred embodiments of the present application are given in the drawings, but do not limit the patent scope of the present application. The present application can be implemented in many different forms, and on the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or equivalently replace some technical features. Any equivalent structure made by using the content of the specification and drawings, directly or indirectly applied to other related technical fields, is also within the scope of the patent protection of the present application.

Claims

1. A method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts, characterized in that: The steps include: Based on the deep learning model, the direction label analysis is performed on the three-dimensional model of the outfitting parts, and the three-dimensional model is normalized to generate the direction label and default position parameters; Based on the direction label and the default position parameters, the perspective configuration file is optimized, the perspective, focal length and position parameters of the camera are set, and a multi-perspective configuration of the outfitting part is generated; Based on the viewing angle configuration file and the material characteristics of the outfitting parts, the intensity and distribution of the light source are adjusted, the material of the highly reflective area is optimized, and a uniform illumination distribution scheme is generated; Based on the view configuration file and the illumination distribution scheme, cyclically rendering the RGB image and the standard coordinate image of the outfitting part, and verifying the illumination consistency and geometric alignment in real time during the rendering process to generate preliminary verification data; Performing a multimodal consistency check on the preliminary verification data and the rendered image, analyzing the geometric consistency and directional matching of the rendering results using a deep learning model, and generating a list of abnormal data; Based on the rendered image and the orientation label, a standardized output file is generated, including an RGB image, a standard coordinate image, and an orientation label; And expand the diversity of outfitting parts dataset through data enhancement module; The multi-view coverage and consistency of the canonical coordinate map of the dataset are comprehensively evaluated, and the evaluation results are used to optimize the direction label generation model and view configuration file, forming a closed-loop improvement mechanism for data generation.

2. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1 is characterized in that: Before the step of performing direction label analysis on the outfitting parts three-dimensional model based on the deep learning model, the method further includes: The three-dimensional model of the outfitting part is imported from a variety of standard three-dimensional model formats, and the geometric structure and material information of the model are extracted.

3. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The step of performing direction label analysis based on the deep learning model specifically includes: The geometric features of the three-dimensional model are analyzed to generate direction labels corresponding to the front, top and side of the model, and a normalized scale factor is generated according to the size of the model.

4. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The step of optimizing the viewing angle configuration file specifically includes: The six fixed viewing angles of the outfitting are determined by the direction tags, and the viewing angle, focal length and position of the camera are adjusted based on the size and geometric characteristics of the outfitting.

5. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The step of adjusting the intensity and distribution of the light source specifically includes: Based on the material properties of the outfitting parts, the position, intensity and angle of the light source are optimized to reduce overexposure in reflective areas and adjust the light distribution.

6. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The step of cyclically rendering the RGB image and the standard coordinate image of the outfitting part specifically includes: According to the view configuration file and the lighting distribution scheme, the RGB map and the standard coordinate map are rendered for each view using automated rendering control, and the lighting consistency and geometric alignment are verified.

7. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The multimodal consistency check step specifically includes: A deep learning model is used to analyze the geometric consistency between images and the matching degree of orientation labels, generating anomaly data on perspective deviation, geometric inconsistency, and orientation errors.

8. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The steps of the data enhancement module specifically include: Lighting simulation, material transformation and geometry enhancement are performed on the rendered RGB map and the canonical coordinate map to generate a new dataset variant.

9. The method for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts according to claim 1, characterized in that: The step of comprehensively evaluating the consistency between the multi-view coverage of the dataset and the standard coordinate map specifically includes: Calculate and evaluate the image coverage area at each viewpoint, verify the geometric alignment and consistency of the canonical coordinate map between views, generate an evaluation report and provide feedback to the orientation label generation model and viewpoint profile optimization.

10. A system for constructing a multi-angle RGB image and standard coordinate image dataset of ship outfitting parts, characterized in that: include: A data acquisition module imports the outfitting part 3D model from a variety of standard 3D model formats, extracts the geometric structure and material information of the 3D model, performs normalization processing, and generates direction labels and default position parameters; A view configuration and parameter optimization module, which optimizes the view configuration file based on the direction label and the default position parameters, sets the view angle, focal length and position parameters of the camera, and generates a multi-view configuration file for the outfitting part; The lighting and material optimization module adjusts the intensity and distribution of the light source according to the material properties and geometric features of the outfitting parts, optimizes the material of the highly reflective areas, and generates a lighting distribution plan for uniform light reception; A rendering and verification module, based on the view configuration file and the illumination distribution scheme, cyclically renders the RGB image and the standard coordinate image of the outfitting part, verifies the illumination consistency and geometric alignment in real time, and generates preliminary verification data; A multimodal consistency check module performs a multimodal consistency check on the preliminary verification data and the rendered image, uses a deep learning model to analyze the geometric consistency and directional matching of the rendering result, and generates a list of abnormal data; A data set generation and enhancement module generates an RGB image, a standard coordinate image, and an orientation label based on the rendered image and the orientation label, and expands the diversity of the outfitting parts data set through a data enhancement module; The closed-loop optimization and feedback module comprehensively evaluates the multi-view coverage and consistency of the canonical coordinate map of the dataset, and uses the evaluation results to optimize the direction label generation model and view configuration file, forming a closed-loop improvement mechanism for data generation, promoting continuous optimization of the system and improvement of dataset quality.

Images