CN-122015644-A - Workpiece three-dimensional reconstruction and dimension measurement method based on multi-view polarization binocular imaging

CN122015644ACN 122015644 ACN122015644 ACN 122015644ACN-122015644-A

Abstract

The invention discloses a workpiece three-dimensional reconstruction and dimension measurement method based on multi-view polarization binocular imaging. The method comprises the steps of constructing a multi-view polarization binocular imaging device, acquiring a multi-view image sequence which surrounds a workpiece by 360 degrees by utilizing a rotary platform, estimating the pose of a left camera at each view by adopting a motion restoration structure algorithm, and determining the pose of a right camera by combining the binocular fixed pose. And extracting polarization information of the image sequence, and introducing polarization constraint in binocular stereo matching to obtain the depth priori of the high-reflection workpiece. A polarization nerve implicit expression model is built based on depth priori, polarization information and camera pose, adaptive sampling is guided by the depth priori, and the three-dimensional shape of a workpiece is represented by a symbol distance field by combining polarizer rendering and normal Gaussian modeling supervision geometric learning, so that surface continuity is improved. And extracting the three-dimensional surface and recovering the real scale, and realizing the size measurement through geometric primitive fitting. The invention is suitable for three-dimensional reconstruction and accurate measurement of high-reflection and weak-texture workpieces.

Inventors

LIU YAN
CHEN TONG
YAN GAN
KONG WEIFENG
HUO GUANYING
LI QINGWU

Assignees

河海大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. The three-dimensional reconstruction and dimension measurement method for the workpiece based on multi-view polarized binocular imaging is characterized by comprising the following steps of: The method comprises the steps of constructing a multi-view polarized binocular imaging device, wherein the device comprises a polarized binocular camera (2), an annular light source (1) and a rotary platform, calibrating the binocular camera through a calibration plate (6) and obtaining a fixed phase-to-pose relation between an internal reference and the binocular camera; fixing the workpiece (5) to be detected on a rotary platform, controlling an imaging device to shoot around the workpiece (5) to be detected in a stepping mode according to a preset angle, continuously collecting multi-view polarized binocular image pairs, and accordingly obtaining an image sequence which surrounds the workpiece (5) to be detected and has high overlapping degree; Inputting the left camera image sequence into a motion restoration structure algorithm, restoring the spatial pose of the left camera at each view angle in a world coordinate system, and calculating the pose of the right camera by combining the binocular relative pose, so as to obtain the pose information of the binocular camera at all view angles; Extracting polarization information under the restriction of camera pose, introducing the polarization restriction in stereo matching, and obtaining multi-view depth information with confidence; Constructing a polarized nerve implicit expression model based on multi-view depth priori, polarization information and camera pose, adaptively sampling through depth priori guidance, simultaneously utilizing polarizer rendering to supervise geometry learning, inhibiting local normal fluctuation by combining normal Gaussian modeling constraint, and representing the three-dimensional surface of a workpiece (5) to be tested by a symbol distance field; And step six, extracting zero equivalent surface grids from the symbol distance field, recovering the real scale, identifying the plane, the cylinder and the free curved surface structure under the constraint of the normal line consistency and the neighborhood relation, realizing dimension measurement according to the geometric primitive fitting result, and automatically calculating key dimensions such as thickness, diameter, hole center distance, coaxiality and the like.
2. The method according to claim 1, wherein the polarized binocular camera (2) is fixed to the rotating platform by a rigid structure, the relative pose of the polarized binocular camera (2) remains unchanged during the acquisition process, the polarized binocular camera (2) moves synchronously with the annular light source (1) around the workpiece (5) to be measured fixed to the rotating platform, and a multi-view polarized binocular image sequence of about 60 views is acquired in steps at a preset angle of 6 ° to ensure sufficient overlap between adjacent views.
3. The method of claim 1, wherein in step four the total matching cost of the stereo matching is determined by the photometric matching cost Cost of agreement with polarization azimuth According to preset weight coefficient Weighted sum representation: ; Wherein, the And Respectively a left pixel point and a right image candidate matching point; cost of polarization azimuth uniformity The definition is as follows: ; Wherein, the And The surface normal azimuth angles corresponding to the left image pixel point and the right image pixel point are respectively obtained.
4. The method of claim 1, wherein in step four, the confidence level Confidence by matching cost Confidence of left-right parallax Local smoothing confidence The weighting structure is used for dynamically adjusting the expansion of the sampling interval in the follow-up self-adaptive ray sampling: ; Wherein, the Is a fixed experience weight; Confidence of the matching cost By optimally matching the saliency description in the candidate set, Is a pixel At optimum parallax The minimum matching total cost at which to place, For all candidate disparities within a search space The sum of the corresponding costs is given by: ; The left-right parallax confidence Is determined by the left and right parallax difference values obtained independently, For the optimal disparity found with reference to the left graph, For the optimal parallax back reversely matched with the right graph as a reference, the formula is as follows: ; Wherein the method comprises the steps of Parameters are adjusted for the dimensions.
5. The method of claim 1, wherein in step five, the neural implicit representation model incorporates a polarized physics model-based volume rendering calculation process in which an outgoing stokes vector From diffuse reflection components according to a polarized bidirectional reflection distribution function model And specular reflection component Linear superposition results in: ; Diffuse reflection polarization feature vector Constructed according to Fresnel's law of transmission, its azimuth of polarization From surface normal Determining the transmission coefficient And According to the refractive index and the observation direction of the material From the surface normal And calculating an included angle to obtain: ; Specular reflection polarization characteristic vector Constructed according to micro-surface reflection theory, the polarization azimuth angle From half-way vectors Determination of half-way vectors therein By directing incident light in a direction opposite to the direction of observation The sum vector of the reflection coefficients is normalized And From refractive index and half-distance vector of material And the direction of observation And calculating an included angle to obtain: ; The polarization distribution characteristics determined by the analysis of the physical law are required to be combined with the corresponding energy intensity items to represent the final emergent state. Diffuse reflection intensity From neural networks According to the space position Predicting specular reflection intensity From neural networks Combining spatial positions Direction of observation Surface normal Prediction is carried out, and finally, a predicted emergent Stokes vector is formed through linear superposition Based on the predicted value And observed value The difference between them can define the loss of polarization uniformity as 。
6. The method according to claim 1, wherein the fifth step further comprises a normal gaussian modeling method, comprising: at a spatial point In-neighborhood selection of (2) Sampling points Calculating corresponding unit normal vector according to gradient of symbol distance field at each sampling point Constructing a three-dimensional Gaussian distribution: ; Wherein the method comprises the steps of Representing the spatial point The mean vector of the normal line is located, Representing the corresponding normal covariance matrix, wherein the calculation modes are as follows: ; 。
7. the method of claim 6, wherein considering polarization information provides constraints only on the direction of a surface normal in an imaging plane, projecting a three-dimensional gaussian distribution of the normal as a two-dimensional gaussian, and constructing a normal gaussian consistency loss based on the two-dimensional gaussian and polarization information : ; Wherein the method comprises the steps of And (3) with For the two-dimensional gaussian mean and covariance obtained by projection, A priori for a normal projection direction constructed from polarization azimuth angles; it serves to constrain the degree of dispersion of the local normal distribution.
8. The method of claim 1, wherein in step five, the depth guided adaptive sampling refers to a sampling interval of rays Based on depth confidence And (3) dynamically adjusting: ; Wherein, the For depth priori The three-dimensional point obtained by back projection is at the parameter position corresponding to the camera ray; To take the following measures The half width of the radial parameter sampling interval is the center; and (3) taking the preset maximum value of the half width of the interval under the scene normalization condition.
9. The method according to claim 1, wherein in the sixth step, the dimension measurement comprises automatically identifying regular structural areas based on the relationship between the normal consistency of the surface sampling points and the spatial neighborhood on the three-dimensional model with the restored real dimensions, and fitting planar, cylindrical and freeform geometric primitives to realize dimension measurement of the workpiece (5) to be measured without manual point selection; performing plane fitting on the plane area to obtain a unit normal vector representing the spatial position and orientation of the plane and displacement parameters, and determining a plane equation; fitting the cylindrical region to obtain the axial direction, the on-axis point and the radius parameters of the cylinder by minimizing the distance error from the sampling point to the axis of the cylinder; And for the free-form surface area, a parameter surface modeling method is adopted, the free-form surface is expressed as a continuous mapping function of parameter coordinates, and the surface parameters of the free-form surface are automatically obtained by minimizing the distance error from the sampling point to the parameter surface.
10. The method of claim 9, wherein the calculating of the critical dimension comprises: Thickness calculation based on displacement parameters of a pair of parallel planes 、 Pressing down Calculating; diameter calculation based on cylinder radius Pressing down Calculating; calculation of hole center distance based on points on different cylinder axes And Pressing down Calculating; coaxiality calculation, namely the main shaft direction is Pressing down Calculating; Calculating the length of curved surface curve, namely selecting a parameter curve in the curved surface parameter domain The corresponding space curve length is ; Calculating curvature radius of curved surface based on main curvature of curved surface at parameter point The principal curvature at the point is The corresponding radius of curvature is 。

Description

Workpiece three-dimensional reconstruction and dimension measurement method based on multi-view polarization binocular imaging Technical Field The invention belongs to the technical field of computer vision and three-dimensional reconstruction, and particularly relates to a workpiece three-dimensional reconstruction and dimension measurement method based on multi-view polarization binocular imaging. Background The size and shape detection of industrial parts typically uses either contact or non-contact measurement methods. Although the contact measurement has higher accuracy, the contact measurement is easy to cause the surface abrasion of the workpiece, is limited by the geometric shape and the contact of the probe, and is difficult to be applied to complex or vulnerable workpieces. Non-contact measurement avoids physical contact risk, is widely used in industry, and main methods include machine vision measurement, laser triangulation, structured light measurement and laser scanning measurement. Among them, machine vision measurement has become a widely used non-contact measurement technique in industry because of low cost, high flexibility and easy integration of an automated production line. The method is usually used for carrying out two-dimensional edge detection based on image intensity or color information or obtaining three-dimensional information through binocular and multi-view stereo matching, but two-dimensional vision measurement only can obtain projection characteristics of a workpiece, and three-dimensional geometric structures such as complex curved surface morphology or step height and the like are difficult to describe completely. In order to realize accurate detection of the complete three-dimensional structure of the workpiece, a three-dimensional reconstruction technology based on multi-view images needs to be introduced, and a high-fidelity geometric model is recovered from a plurality of view angle observations. Neural implicit representation is used as an emerging three-dimensional reconstruction method, and can model space geometry through continuous function forms, so that the neural implicit representation has the potential of recovering complex geometry under the multi-view condition. However, the existing neural implicit reconstruction method is based on the assumption of image luminosity consistency, and has the defects of insufficient explicit modeling on the surface reflection characteristics of the workpiece, and is easy to be influenced by high light interference and effective characteristic missing on the high-reflection and weak-texture workpiece represented by metal, so that local geometric errors and reconstruction distortion are caused, and the accuracy of workpiece size measurement is further influenced. In order to overcome the difficult problem of measuring the high-reflection and weak-texture workpiece, polarization information can be introduced to assist in three-dimensional reconstruction. The information can reflect the physical characteristics of the interaction of light and the surface of the workpiece, and provide additional surface geometric constraint for three-dimensional reconstruction in a high-reflection and weak-texture area, so that the problems of inconsistent luminosity caused by specular reflection and unreliable matching caused by texture deletion are effectively suppressed, and the stability and measurement accuracy of a reconstruction result are improved. Disclosure of Invention The invention aims to provide a workpiece three-dimensional reconstruction and dimension measurement method based on multi-view polarization binocular imaging, which solves the problems that the prior art is easy to generate local geometric deviation and reconstruction distortion in non-contact measurement of a high-reflection and weak-texture workpiece. In order to achieve the above object, the present invention is realized by the following technical scheme. A workpiece three-dimensional reconstruction and dimension measurement method based on multi-view polarization binocular imaging comprises the following steps: The method comprises the steps of constructing a multi-view polarized binocular imaging device, wherein the device comprises a polarized binocular camera 2, an annular light source 1 and a rotary platform, calibrating the binocular camera through a calibration plate 6, and acquiring a fixed phase-to-pose relation between an internal reference and the binocular camera; fixing the workpiece 5 to be measured on a rotary platform, controlling an imaging device to shoot around the workpiece 5 to be measured in a stepping mode according to a preset angle, and continuously collecting multi-view polarized binocular image pairs so as to obtain an image sequence which surrounds the workpiece 5 to be measured for one circle and has high overlapping degree; Inputting the left camera image sequence into a motion restoration structure algorithm, restoring the spatial pose