Search

CN-121883501-B - Image data processing method for mosaic positioning of decoration panel

CN121883501BCN 121883501 BCN121883501 BCN 121883501BCN-121883501-B

Abstract

The invention belongs to the technical field of image data processing and discloses an image data processing method for mosaic positioning of a decoration panel, which comprises the following steps of S1, synchronously acquiring two-dimensional image data and three-dimensional point cloud data of a decoration panel and an mosaic, S2, carrying out gray correction, 3X 3 kernel Gaussian filtering and self-adaptive threshold segmentation preprocessing on the two-dimensional image data, denoising and downsampling the three-dimensional point cloud data by adopting a statistical filtering algorithm, carrying out coordinate registration on the three-dimensional point cloud data and the preprocessed two-dimensional image data to establish a mapping relation between pixel coordinates and physical coordinates, S3, inputting the preprocessed two-dimensional image data and the three-dimensional point cloud data into a CNN convolutional neural network model trained by decoration panel mosaic scene samples, S4, establishing an XYA reference coordinate base based on decoration panel design drawing, respectively calculating independent detection coordinates for four types of characteristic information, outputting final detection coordinates after fusion by a weighted average algorithm, and S5, calculating uncompensated offset values of the final detection coordinates and the reference coordinates.

Inventors

  • HUANG QINGYAO
  • MA YI

Assignees

  • 威准（厦门）自动化科技有限公司

Dates

Publication Date
20260512
Application Date
20260323

Claims (9)

  1. 1. The image data processing method for mosaic positioning of the decoration panel is characterized by comprising the following steps of: s1, synchronously acquiring two-dimensional image data and three-dimensional point cloud data of a decoration panel and an insert, wherein the two-dimensional image data are collected at different visual angles, and the three-dimensional point cloud data are profile shape data of the decoration panel and the insert; s2, performing gray correction, 3×3 kernel Gaussian filtering and adaptive threshold segmentation preprocessing on the two-dimensional image data, denoising and downsampling the three-dimensional point cloud data by adopting a statistical filtering algorithm, and then performing coordinate registration with the preprocessed two-dimensional image data to establish a mapping relation between pixel coordinates and physical coordinates; S3, inputting the preprocessed two-dimensional image data and the three-dimensional point cloud data into a CNN convolutional neural network model trained by a decoration panel mosaic scene sample, and merging three-dimensional depth and two-dimensional texture features to synchronously identify four feature information of a decoration datum line, an mosaic groove contour, an mosaic piece appearance and a texture alignment mark; S4, establishing an XYA reference coordinate library based on the design drawing of the decoration panel, respectively calculating independent detection coordinates for the four types of characteristic information, and outputting final detection coordinates after fusion by a weighted average algorithm; s5, calculating uncompensated offset values of the final detection coordinates and the reference coordinates, substituting the real-time working temperature of the equipment into a temperature compensation formula for correction, and obtaining offset values after compensation; The calculation formula of the uncompensated offset value in step S5 is as follows: , , The temperature compensation formula is: , , , wherein, 、 、 Respectively an X-axis coordinate, a Y-axis coordinate and a rotation angle around a Z axis which are finally detected after fusion; 、 、 Respectively presetting an X-axis reference coordinate, a Y-axis reference coordinate and a rotation reference angle around a Z axis for a product design drawing; 、 、 respectively an uncompensated X-axis offset value, a uncompensated Y-axis offset value and a rotating offset angle around a Z axis; 、 、 The offset value is an X-axis offset value, a Y-axis offset value and a Z-axis rotation offset angle after temperature compensation; The real-time working temperature of the machine is set, The thermal expansion coefficient of the decoration panel material is 25, which is the standard reference temperature, namely the product design reference temperature; simultaneously introducing a humidity compensation factor to carry out secondary correction on the offset value after temperature compensation: , , , , wherein, As the humidity compensation factor, a temperature compensation factor, As the coefficient of influence of the humidity, For the actual ambient humidity, 50 is the standard reference humidity, 、 、 Respectively carrying out secondary correction on the X-axis offset value, the Y-axis offset value and the rotation offset angle around the Z axis; and S6, checking the positioning accuracy of the offset value after compensation by adopting a root mean square error algorithm, and if the positioning accuracy exceeds a preset accuracy threshold value, returning to the step S1 to acquire the data again until the positioning accuracy meets the preset requirement.
  2. 2. The method for processing image data for mosaic positioning of a decoration panel according to claim 1, wherein when two-dimensional image data are acquired in step S1, four groups of light compensating light sources with different wavelengths, namely 450nm blue light, 530nm green light, 650nm red light and 780nm near infrared light, are dynamically switched based on real-time detection results of the material, texture density and reflection intensity of the surface of the decoration panel, and the light intensity of the light sources is adaptively adjusted within a preset range according to a fixed gradient.
  3. 3. The method for processing image data for mosaic positioning of decorative panels according to claim 1, wherein the statistical filtering denoising algorithm formula adopted by the three-dimensional point cloud data in the step S2 is as follows: , wherein, For the three-dimensional point cloud data point set remained after filtering, p is the three-dimensional point cloud data point to be screened, For the set separator, the subsequent condition is indicated to be satisfied, For the number of points in the p-neighborhood, For p and the neighborhood point Is used for the distance of euclidean distance, As the distance average value of the neighborhood points, Is the standard deviation of the distance; and removing discrete noise points which deviate from the mean value by 3 times of standard deviation by the statistical filtering denoising algorithm.
  4. 4. The method for processing image data for mosaic positioning of decoration panels according to claim 1, wherein in step S2, an iterative closest point algorithm is adopted for coordinate registration, a dynamic weight factor is introduced in the registration process to optimize the iterative convergence speed, and the dynamic weight factor calculation formula is as follows: , wherein, Is the first The weight factor of the number of iterations, The initial weight factor is 0.8-1.0, The attenuation coefficient is 0.05 to 0.1, The iteration times; The coordinate registration is to unify the coordinate system of the three-dimensional point cloud data and the pixel coordinate system of the two-dimensional image data to a physical coordinate system, the error of the coordinate registration is less than or equal to 0.002mm, and the registration time is less than or equal to 0.5s.
  5. 5. The method for processing image data of mosaic positioning of a decoration panel according to claim 1, wherein in the step S3, a shallow layer convolution layer of the CNN convolution neural network model extracts basic features of edges and textures through a 3×3 convolution kernel, a middle layer convolution layer strengthens local feature response through cavity convolution, a deep layer fuses three-dimensional depth information and two-dimensional texture features through a cross-dimensional attention mechanism to realize complementary enhancement of two types of data, a feature confidence assessment mechanism is built in the CNN convolution neural network model, and when single type feature identification confidence does not reach a preset confidence threshold, a standby feature extraction channel is automatically called to re-extract and calculate the type of features.
  6. 6. The method for processing image data for mosaic positioning of decorative panel according to claim 5, wherein in step S3, the characteristic fusion loss function formula of the CNN convolutional neural network model is: , wherein, The total loss function is fused for the features, Cross entropy loss is identified for 2D image features, A mean square error loss is identified for the 3D point cloud features, For contrast loss of 2D and 3D feature fusion, 、 、 Is a loss weight coefficient, and 。
  7. 7. The method for processing image data for mosaic positioning of decorative panels according to claim 1, wherein in step S4, the XYA reference coordinate library comprises X-axis reference coordinates, Y-axis reference coordinates, rotation reference angles around Z-axis, and allowable deviation ranges of four types of features, and the fusion formula of the weighted average algorithm is as follows: , , , wherein, 、 、 Respectively an X-axis coordinate, a Y-axis coordinate and a rotation angle around a Z axis which are finally detected after fusion; 、 、 The X-axis coordinate, the Y-axis coordinate and the rotation angle around the Z-axis are independently detected for the characteristic of the decoration datum line; 、 、 the X-axis coordinate, the Y-axis coordinate and the rotation angle around the Z axis are independently detected for the profile features of the inlaid grooved pulley; 、 、 the method comprises the steps of independently detecting X-axis coordinates, Y-axis coordinates and rotation angles around a Z axis of the appearance characteristics of the insert; 、 、 The X-axis coordinate, the Y-axis coordinate and the rotation angle around the Z axis are independently detected for the texture alignment mark characteristics; Meanwhile, calculating the deviation contribution rate of each feature detection coordinate and the reference coordinate: , , , wherein, 、 、 Respectively represent the first The bias contribution rate of the class features in the directions of X axis, Y axis and rotation angle around Z axis, 、 、 Respectively the first The class features are preset in an XYA reference coordinate library as X-axis reference coordinates, Y-axis reference coordinates and rotated around a Z-axis reference angle, 、 、 Respectively the first The X-axis coordinate, Y-axis coordinate and the rotation angle around the Z-axis are independently detected by the class features, 、 、 Respectively the first The X-axis coordinate, Y-axis coordinate and the rotation angle around the Z-axis are independently detected by the class features, 、 、 Respectively the first The class features are preset in an XYA reference coordinate library as X-axis reference coordinates, Y-axis reference coordinates and rotated around a Z-axis reference angle, Serial numbers corresponding to the four types of characteristics, 1 corresponding to a decoration datum line, 2 corresponding to an embedded groove outline, 3 corresponding to an embedded part outline, 4 corresponding to a texture alignment mark, For the traversal sequence number of the four types of features, Representing a summation of the absolute values of the corresponding deviations of the four classes of features.
  8. 8. The method for processing image data for mosaic positioning of decorative panel according to claim 7, wherein the formula of the root mean square error algorithm in step S6 is: 0.01 is an angle error conversion coefficient, converts the rotation angle offset into an equivalent length offset, and ensures the unification of error calculation dimensions; And (3) presetting a positioning accuracy threshold value to be 0.005mm, returning to the step (S1) to trigger re-acquisition positioning when the RMSE is more than 0.005mm, and optimizing the parameters of the light supplementing light source and the acquisition density of the point cloud.
  9. 9. The method for processing image data of mosaic positioning of a decoration panel according to claim 1, further comprising the steps of S7, after the positioning accuracy is checked to be qualified, storing a two-dimensional positioning original image, a three-dimensional contour data image, a reference coordinate, a final detection coordinate, a compensated offset value, processing parameters of each step, environment temperature and humidity data and feature recognition confidence level and a two-dimensional code of the decoration panel product in a distributed database in a correlated manner, simultaneously transmitting the compensated offset value to an executing mechanism in real time to realize linkage control of positioning and a subsequent process, setting a parameter abnormality feedback mechanism in the linkage process, triggering an alarm and suspending the subsequent process to execute when the linkage parameter deviation reaches an abnormality threshold value, and resuming operation after confirmation or correction.

Description

Image data processing method for mosaic positioning of decoration panel Technical Field The invention relates to the technical field of image processing, in particular to an image data processing method for mosaic positioning of a decoration panel. Background The surface embedding technology of the miniature decoration panel is a core link of processing high-end miniature home ornaments and precise ornaments, and the embedding precision directly determines the artistic value and the use reliability of the product. The key technical bottleneck of the process is that the positioning accuracy is controlled, the sizes of the decoration panel and the inserts are tiny (the width and the depth of the insert grooves are mostly micro-scale), four types of features of the decoration datum line, the outline of the insert grooves, the appearance of the inserts and the texture alignment mark are required to be synchronously identified, and the traditional image data processing method is difficult to realize the accurate co-positioning of multiple features. In the prior art, single 2D image recognition or simple 3D point cloud analysis is adopted for image data processing, and the problems that a single 2D image is easily subjected to surface texture, chromatic aberration and reflective interference of a decoration panel to cause micro-scale feature recognition deviation, the recognition capability of the single 3D point cloud to two-dimensional features such as plane texture alignment marks and the like is insufficient, point cloud data easily contain discrete noise points, and meanwhile, a standardized multi-feature coordinate fusion mechanism is lacking, the influence of factors such as environment temperature and humidity on positioning is not considered, the final positioning offset value exceeds a preset precision threshold (+ -0.005 mm), and further the problems of glue overflow embedding, insufficient adhesive strength and the like are caused, and the high precision requirement of micro decoration panel embedding cannot be met. Aiming at the core technical problem, an image data processing method which fuses two-dimensional image and three-dimensional point cloud data and has multi-feature collaborative recognition and accurate coordinate calculation capability is needed, and the precision and stability of mosaic positioning of the decoration panel are fundamentally improved. Disclosure of Invention The invention aims to provide an image data processing method which fuses two-dimensional images and three-dimensional point cloud data and has multi-feature collaborative recognition and accurate coordinate calculation capability, and aims to solve the problem that the precision and stability of mosaic positioning of a decoration panel are poor. In order to achieve the above purpose, the present invention adopts the following technical scheme: an image data processing method for mosaic positioning of decoration panels comprises the following steps: s1, synchronously acquiring two-dimensional image data and three-dimensional point cloud data of a decoration panel and an insert, wherein the two-dimensional image data are collected at different visual angles, and the three-dimensional point cloud data are profile shape data of the decoration panel and the insert; s2, performing gray correction, 3×3 kernel Gaussian filtering and adaptive threshold segmentation preprocessing on the two-dimensional image data, denoising and downsampling the three-dimensional point cloud data by adopting a statistical filtering algorithm, and then performing coordinate registration with the preprocessed two-dimensional image data to establish a mapping relation between pixel coordinates and physical coordinates; S3, inputting the preprocessed two-dimensional image data and the three-dimensional point cloud data into a CNN convolutional neural network model trained by a decoration panel mosaic scene sample, and merging three-dimensional depth and two-dimensional texture features to synchronously identify four feature information of a decoration datum line, an mosaic groove contour, an mosaic piece appearance and a texture alignment mark; S4, establishing an XYA reference coordinate library based on the design drawing of the decoration panel, respectively calculating independent detection coordinates for the four types of characteristic information, and outputting final detection coordinates after fusion by a weighted average algorithm; s5, calculating uncompensated offset values of the final detection coordinates and the reference coordinates, substituting the real-time working temperature of the equipment into a temperature compensation formula for correction, and obtaining offset values after compensation; and S6, checking the positioning accuracy of the offset value after compensation by adopting a root mean square error algorithm, and if the positioning accuracy exceeds a preset accuracy threshold value, returning to the step S1 to acqui