A circular hole pose measurement method based on edge enhancement in binocular vision scenes

Abstract

The present invention discloses a circular hole pose measurement method based on edge enhancement in a binocular vision measurement scenario. First, in the preprocessing stage of the image super-resolution reconstruction process, a fuzzy filling operation is used to help the model learn the position and degree of super-resolution reconstruction; based on an enhanced adversarial generative network, edge enhancement of the circular hole super-resolution reconstructed image is achieved by adding cross-layer residual connections, introducing context loss in the perceptual loss term, and adding a pixel rearrangement mechanism; then, the sub-pixel level fine edge of the circular hole is extracted, and the parameters are fitted by an improved Hough transform method combined with geometric features; the binocular camera is calibrated by a multi-constraint fusion calibration method; and the circular hole pose parameter measurement is completed using stereo matching technology through the constructed binocular camera observation model. The present invention improves the accuracy and stability of circular hole pose measurement through edge enhancement, fine circular hole edge extraction and fitting methods.

Description

Technical Field

[0001] The present invention relates to a circular hole posture visual measurement method, and in particular to a circular hole posture measurement method based on edge enhancement in a binocular vision scene. Background Art

[0002] Circular holes are one of the most common features on parts, often found on sleeves, disc covers, and housings. In actual production, the assembly and docking of parts containing holes often utilize circular holes as features. Automated assembly and docking of parts with holes requires ensuring alignment between the axis of the shaft and hole. Therefore, achieving real-time, high-precision detection of the target circular hole's position is crucial.

[0003] Currently, existing methods for visually measuring the position and pose of circular holes are limited by the performance of visual acquisition system hardware, such as lenses, cameras, and image acquisition cards. This is also affected by factors such as changing lighting and complex backgrounds. This results in insufficient image resolution and unclear circular hole edges, leading to frequent large measurement errors. Therefore, fast and high-precision circular hole position and pose measurement remains a challenging task.

[0004] In 1992, Safaee-Rad et al. proposed a general equation for the path from a cone to its center in the IEEE Transactions on Robotics and Automation paper, "Three-dimensional location estimation of circular features for machine vision." This equation provides a closed-form analytical solution for measuring the pose of a circular hole and proposes two closed-form solutions, depending on whether the radius is known. In 2012, Xu et al. addressed the orientation duality problem in their paper "A pose measurement method of a non-cooperative GEO spacecraft based on stereo vision" published at the 12th International Conference on Control, Automation, Robotics and Vision. They proposed estimating the pose of the circular hole from the left and right images separately, and then removing the ambiguity of the monocular vision solution by constraining the pose difference between the left and right images to be within a threshold. In their 2016 paper, "Pose measurement of a non-cooperative spacecraft based on circular features," published in the IEEE International Conference on Real-time Computing and Robotics, Liu et al. proposed a closed-form solution for measuring the pose of circular holes of unknown radius. The solution used the quadratic form of the ellipse in the image and the pinhole projection model to obtain the quadratic curve of the 3D circle. The quadratic curve of the 3D circle, obtained by assuming the ellipses in the left and right images are equal, was then used to construct an equation to calculate the pose information of the circle. In their 2019 master's thesis, "Visual Detection of Circular Hole Pose Based on Image Super-resolution Reconstruction," Zhang proposed improving image resolution through image super-resolution reconstruction, thereby improving the accuracy of circular hole pose measurement. In their 2020 paper, "An accurate and robust method for the measurement of circular holes based on binocular vision," published in Measurement Science and Technology, Xia et al. proposed a method for obtaining high-precision circular hole edge points based on pixel intensity information and combining epipolar geometry constraints. These edge points were then reconstructed into spatial circular holes through three-dimensional reconstruction.

[0005] The above research has improved the accuracy of circular hole pose measurement, but it still faces problems such as blurred circular hole edges, inaccurate high-frequency information, and long time consumption. Summary of the Invention

[0006] In order to solve the problems existing in the background technology, the present invention proposes a circular hole pose measurement method based on edge enhancement in a binocular vision measurement scenario, which has the advantages of fast, high efficiency, and high measurement accuracy.

[0007] Firstly, in the process of image super-resolution reconstruction, the present invention realizes edge enhancement of the circular hole super-resolution reconstructed image through methods such as fuzzy filling, adding cross-layer residual connections to the network, introducing context loss in the perceptual loss term, and adding a pixel rearrangement mechanism. Secondly, the sub-pixel edge of the circular hole is extracted, and the parameters are fitted by the improved Hough transform method combined with geometric features. The binocular camera is calibrated through a multi-constraint fusion calibration method. Finally, the circular hole pose parameters are measured using stereo matching technology through the constructed binocular camera observation model.

[0008] The specific technical solutions adopted in the present invention include:

[0009] 1) Use a binocular camera to capture parts containing holes to obtain multiple high-resolution circular hole images. Preprocess each high-resolution circular hole image to obtain a corresponding blurred circular hole image. The dataset consists of each high-resolution circular hole image and its corresponding blurred circular hole image.

[0010] 2) Inputting the data set into the circular hole image super-resolution reconstruction deep network to train the network and obtain the trained circular hole image super-resolution reconstruction deep network;

[0011] 3) The binocular camera captures the high-resolution circular hole image to be reconstructed and inputs it into the trained circular hole image super-resolution reconstruction deep network. The network outputs the edge-enhanced high-resolution circular hole image.

[0012] 4) Using the circular hole edge sub-pixel extraction method, sub-pixel edge extraction is performed on the edge-enhanced high-resolution circular hole image to obtain a sub-pixel edge image;

[0013] 5) The parameters of the target circular hole are obtained by fitting the circular hole parameters of the sub-pixel edge image using the improved Hough transform method combined with geometric features;

[0014] 6) Based on the parameters of the target circular hole, the super-resolution reconstruction camera observation model is used to perform coordinate mapping to obtain the position information of the target circular hole in the camera coordinate system;

[0015] 7) According to the pose information of the target circular hole in the camera coordinate system, the stereo matching method based on the binocular camera is used to calculate the pose parameters of the target circular hole in the world coordinate system to complete the circular hole pose measurement.

[0016] In the step 1), each high-resolution circular hole image is subjected to degradation processing to obtain a corresponding low-resolution circular hole image, and each pair of circular hole images is formed by each high-resolution circular hole image and the corresponding low-resolution circular hole image. The circular hole image pairs are then fuzzy filled to obtain fuzzy circular hole images.

[0017] The circular hole image super-resolution reconstruction deep network consists of a cascade of a pixel rearrangement mechanism module and an improved enhanced generative adversarial network. The pixel rearrangement mechanism module is obtained by performing an inverse operation on the pixel upsampling module Pixel-shuffle, and the improved enhanced generative adversarial network is obtained by performing cross-layer residual connections on the three RRDB blocks in the enhanced generative adversarial network.

[0018] The L2 loss function in the perceptual loss function of the enhanced adversarial generative network is replaced by a contextual loss function.

[0019] The step 4) is specifically as follows:

[0020] 4.1) The edge-enhanced high-resolution circular hole image is input into the circular hole semantic segmentation model for prediction, and the circular hole edge mask is output;

[0021] 4.2) Use the edge extraction operator to extract the edge of the edge-enhanced high-resolution circular hole image to obtain the initial edge image of the part. Multiply the initial edge image of the part with the circular hole edge mask to obtain the rough edge image of the part.

[0022] 4.3) Perform fine edge extraction on the rough edge image of the part to obtain an accurate edge image;

[0023] 4.4) Based on the edge-enhanced high-resolution circular hole image and the accurate edge image, a sub-pixel edge image is calculated.

[0024] In step 4.1), the circular hole semantic segmentation model is an attention mechanism module added before the pyramid pooling ASPP module of the Deeplabv3+ semantic segmentation network, and the pyramid pooling ASPP module and the attention mechanism module are cascaded.

[0025] In the step 4.2), the edge extraction operator is specifically the Roberts operator.

[0026] The step 4.3) is specifically as follows:

[0027] The rough edge image of the part is sequentially subjected to hole edge fitting, connection and refinement to obtain an accurate edge image.

[0028] The step 4.4) is specifically as follows:

[0029] The sub-pixel points of the whole image of the predicted part with holes are extracted to obtain the sub-pixel points of the whole image. The sub-pixel points of the whole image are intersected with the accurate edge image to obtain the sub-pixel points of the circular hole edge, thereby obtaining the sub-pixel edge image.

[0030] The super-resolution reconstruction camera observation model is specifically a mapping relationship between the camera coordinate system of the binocular camera and the virtual coordinate system of the reconstructed image.

[0031] The beneficial effects of the present invention are:

[0032] 1) The training set is preprocessed by the fuzzy filling method, so that the super-resolution reconstruction deep network can learn the position and scale of super-resolution reconstruction, avoid over-sharpening of the reconstructed image, and make the network reconstruction process have a regularization effect.

[0033] 2) By adding cross-layer residual connections to the super-resolution reconstruction deep network, introducing context loss in the perceptual loss term, and introducing a pixel rearrangement mechanism module, edge enhancement of the reconstructed circular hole image is achieved, and a high-resolution image with rich high-frequency information and realistic edge details is reconstructed.

[0034] 3) The fine edge of the circular hole is extracted by the sub-pixel level extraction method, and the ellipse parameters are calculated by the improved Hough transform ellipse parameter fitting method combined with geometric features. The accuracy of the circular hole pose parameters measured by the stereo matching method is further improved by reconstructing the camera observation model based on super-resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the method of the present invention.

[0036] Figure 2 This is an example of a preprocessed image pair in the sample library of parts containing holes.

[0037] Figure 3 is an example of a reconstructed edge-enhanced circular hole image.

[0038] Figure 4 It is the sub-pixel extraction process of the circular hole edge.

[0039] Figure 5 It is the process of ellipse detection in circular hole image.

[0040] Figure 6 It is a flowchart of multi-constraint fusion dual-target positioning process.

[0041] Figure 7 This is an example of the measurement results of the circular hole pose of a typical part with a hole.

[0042] Figure 8 This is a structural diagram of the three RRDB blocks in the improved enhanced adversarial generation network. DETAILED DESCRIPTION

[0043] The present invention will be further described below with reference to the accompanying drawings and examples.

[0044] The embodiments of the present invention and their implementation process are as follows:

[0045] like Figure 1 As shown, the present invention includes the following steps:

[0046] 1) In the specific implementation, the binocular camera is calibrated by combining 2D and 3D fusion constraints to obtain a calibrated binocular camera. The calibrated binocular camera is used for subsequent acquisition. The calibration accuracy is improved by the camera calibration method of 2D-3D multi-constraint fusion. The calibration flow chart is as follows Figure 6 As shown in the figure, based on Zhang Zhengyou's camera calibration method, the camera calibration errors such as adjacent distance error, collinear error, and right angle error are calculated at the 2D image level; the reprojection error and epipolar error are calculated at the 3D level, that is, the constraints of the 2D image level and the 3D space level are integrated; and the Levenberg-Marquardt iterative algorithm is used to obtain the optimal solution for the camera parameters to complete the binocular camera calibration.

[0047] Use a binocular camera to shoot parts with holes to obtain multiple high-resolution circular hole images, such as Figure 2 As shown in (a), each high-resolution circular hole image is preprocessed to obtain the corresponding blurred circular hole image, and the data set is composed of each high-resolution circular hole image and the corresponding blurred circular hole image;

[0048] In step 1), each high-resolution circular hole image is degraded to obtain the corresponding low-resolution circular hole image, such as Figure 2 As shown in (b), specifically, the high-resolution circular hole image is subjected to noise addition, interpolation, and JPEG compression to generate the corresponding low-resolution circular hole image, wherein the three steps of noise addition, interpolation, and JPEG compression are not in any order. In a specific implementation, the high-resolution circular hole image is subjected to noise addition, interpolation, and JPEG compression in sequence to generate the corresponding low-resolution circular hole image. Each pair of circular hole images is composed of each high-resolution circular hole image and the corresponding low-resolution circular hole image, wherein the resolution between the high-resolution circular hole image and the corresponding low-resolution circular hole image is relative to each other. The circular hole image pair is then blurred and filled to obtain a blurred circular hole image, as shown in FIG. Figure 2 As shown in (c), specifically: the low-resolution circular hole image of the circular hole image pair is enlarged to the same size as the high-resolution image by upsampling, and then the circular hole area of the high-resolution image is cropped and filled into the corresponding area of the low-resolution circular hole image to obtain a blurred circular hole image.

[0049] 2) Inputting the data set into the circular hole image super-resolution reconstruction deep network to train the network and obtain the trained circular hole image super-resolution reconstruction deep network;

[0050] The deep network for super-resolution reconstruction of circular hole images consists of a cascade of a pixel rearrangement mechanism module and an improved enhanced generative adversarial network. The pixel rearrangement mechanism module is obtained by performing the inverse operation of the pixel upsampling module, the pixel shuffle. Taking the image 2x super-resolution reconstruction process as an example, the pixel upsampling module performs a 2x upsampling operation on the image to achieve the ×2 image super-resolution reconstruction goal. However, before the image super-resolution reconstruction, the present invention introduces the pixel rearrangement mechanism module, that is, it first performs a 2x downsampling operation on the reconstructed image to reduce the spatial size while rearranging the information into the channel dimension; then it performs a 4x super-resolution reconstruction, ultimately achieving the ×2 image super-resolution reconstruction goal.

[0051] The improved enhanced generative adversarial network is obtained by performing cross-layer residual connections on the three RRDB blocks in the enhanced generative adversarial network; that is, the original one-way linear information transmission mode of the three RRDB blocks in the enhanced generative adversarial network is transformed into a cross-layer residual connection, realizing information propagation between different residual blocks and building a cross-layer super-resolution reconstruction network based on the enhanced generative adversarial network architecture. Figure 8 As shown, specifically: the input of the first RRDB block is further input to the second RRDB block after the first convolution layer and the output of the first RRDB block is channel-added, the input of the second RRDB block is further input to the third RRDB block after the second convolution layer and the output of the second RRDB block is channel-added, the input of the third RRDB block is further input to the third convolution layer and the output of the third RRDB block is channel-added as the first output, the input of the first RRDB block is further input to the fourth convolution layer and the first output is channel-added as the input of the average pooling layer.

[0052] The L2 loss function in the perceptual loss function of the enhanced adversarial generation network is replaced by the context loss function.

[0053] The formula of the perceptual loss function is as follows:

[0054]

[0055] Among them, L G is the perceptual loss function value, L CX is the context loss function value, is the generation loss value, λ and η are the coefficient factors for adjusting the generation loss and content loss respectively, and L1 represents the edge-enhanced high-resolution circular hole image G(x i ) and the content loss of the 1-norm distance between the high-resolution hole image y.

[0056] The context loss function is as follows:

[0057]

[0058] Among them, L CX (·) represents the context loss at a certain point in the edge-enhanced high-resolution circular hole image obtained by network reconstruction, D(·) is the standard discriminant network, x and y are the low-resolution image blocks and high-resolution image blocks corresponding to the image pairs in the training set, respectively. is the kth low-resolution image block x k and the jth high-resolution image patch y j where exp(·) represents the exponential operation, max represents the maximum operation, and log(·) represents the logarithmic operation.

[0059] 3) The binocular camera collects the high-resolution circular hole image to be reconstructed and inputs it into the trained circular hole image super-resolution reconstruction deep network, and the network outputs the edge-enhanced high-resolution circular hole image; Figure 3 As shown, Figure 3 (a) is the high-resolution circular hole image to be reconstructed. Figure 3 (b) is a high-resolution circular hole image with edge enhancement.

[0060] 4) Using the circular hole edge sub-pixel extraction method, sub-pixel edge extraction is performed on the edge-enhanced high-resolution circular hole image to obtain a sub-pixel edge image;

[0061] Step 4) is specifically as follows:

[0062] 4.1) The edge-enhanced high-resolution circular hole image is input into the circular hole semantic segmentation model for prediction, and the circular hole edge mask is output; Figure 4 As shown, Figure 4 (a) is the output circular hole edge mask Mask, Figure 4 (b) is the result image after the circular hole edge mask is pasted to the image of the part with holes to be predicted.

[0063] The circular hole semantic segmentation model adds an attention mechanism module before the pyramid pooling ASPP module in the Deeplabv3+ semantic segmentation network. The pyramid pooling ASPP module and the attention mechanism module are cascaded. The modules placed before and cascaded with the pyramid pooling ASPP module in the Deeplabv3+ semantic segmentation network are also cascaded with the attention mechanism module. Images of parts containing holes are first filtered for feature information by the attention mechanism module, and then the results are subjected to targeted semantic segmentation.

[0064] 4.2) Use the edge extraction operator to extract the edge of the edge-enhanced high-resolution circular hole image to obtain the initial edge image of the part. Multiply the initial edge image of the part with the circular hole edge mask to obtain the rough edge image of the part.

[0065] In step 4.2), the edge extraction operator is specifically the Roberts operator.

[0066] 4.3) Perform fine edge extraction on the rough edge image of the part to obtain an accurate edge image;

[0067] Step 4.3) is specifically as follows:

[0068] The rough edge image of the part is sequentially subjected to hole edge fitting, connection and refinement to obtain an accurate edge image.

[0069] 4.4) Based on the edge-enhanced high-resolution circular hole image and the accurate edge image, a sub-pixel edge image is calculated.

[0070] Step 4.4) is specifically as follows:

[0071] The sub-pixel points of the whole image of the predicted part with holes are extracted to obtain the sub-pixel points of the whole image, such as Figure 5 As shown in (a), the sub-pixel points of the full image are intersected with the accurate edge image to obtain the sub-pixel points of the circular hole edge, thereby obtaining the sub-pixel edge image, as shown in Figure 5 As shown in (b).

[0072] 5) The parameters of the target circular hole are obtained by fitting the circular hole parameters of the sub-pixel edge image using the improved Hough transform method combined with geometric features. The parameters of the target circular hole include the coordinates of the center of the circle and the normal vector;

[0073] The improved Hough transform method combined with geometric features is specifically as follows:

[0074] First, edge detection is performed on the edges of the image, and the coordinates of the detected points are stored in an array; secondly, the maximum distance from the point in the array is calculated. The point with the smallest maximum distance among all points is the center of the ellipse, and the maximum distance is the length of the major axis of the ellipse; thirdly, the numerical value of the point obtained and the ellipse parameters obtained by the above process are substituted into the ellipse equation, and the remaining parameters are statistically analyzed to define the relevant thresholds; finally, the parameters whose output peak exceeds the threshold are the target ellipse parameters.

[0075] In theory, there is the following theorem: let the center of the ellipse on the plane be point c, and take any point on the plane different from point c, denoted as point p, then the maximum distance between point p and the point on the ellipse must be greater than the maximum distance between point c and the point on the ellipse.

[0076] According to the above principles, the calculation process of the ellipse parameters is as follows:

[0077] The major axis, minor axis, and rotation angle of the target ellipse are denoted as a, b, and θ respectively, and the coordinates of the center of the ellipse are marked as point (p, q). The calculation process of the ellipse equation is as follows:

[0078]

[0079] The specific detection process is as follows: edge detection is performed on the image edges, and the coordinates of the detected points are stored in an array. The maximum distance from the point in the array is calculated. The point with the smallest maximum distance among all points is the ellipse center (p, q), and this maximum distance is the length of the ellipse's major axis, a. The obtained point values and the ellipse parameters p, q, and a obtained in the above process are substituted into the above ellipse equation. The parameters b and θ are statistically analyzed, and relevant thresholds are set. The parameters whose peak value exceeds the threshold are then output as the target ellipse parameters.

[0080] 6) Based on the parameters of the target circular hole, the super-resolution reconstruction camera observation model is used to perform coordinate mapping to obtain the position information of the target circular hole in the camera coordinate system;

[0081] The super-resolution reconstruction camera observation model is specifically the mapping relationship between the camera coordinate system of the binocular camera and the virtual coordinate system of the reconstructed image. The virtual coordinate system of the reconstructed image is the image coordinate system after the image is reconstructed using the super-resolution reconstruction method.

[0082] 7) According to the pose information of the target circular hole in the camera coordinate system, the stereo matching method based on the binocular camera is used to calculate the pose parameters of the target circular hole in the world coordinate system to complete the circular hole pose measurement.

[0083] Specifically, the left and right cameras of the binocular camera system can obtain two normal vector measurements respectively, calculate the angle between the vectors, and take into account interference factors such as measurement error and accidental error, and take the average of the two vectors with the smallest measured angle as the final normal vector;

[0084]

[0085] Among them, the first and second normal vectors of the target circular hole measured by the left camera are and The first and second normal vectors of the target circular hole measured by the right camera are: and The final normal vector is represents the angle between the i-th normal vector measured by the left camera and the j-th normal vector measured by the right camera, It represents the angle between the first normal vector measured by the left camera and the first normal vector measured by the right camera. It represents the angle between the first normal vector measured by the left camera and the second normal vector measured by the right camera. It represents the angle between the second normal vector measured by the left camera and the first normal vector measured by the right camera. Represents the angle between the second normal vector measured by the left camera and the second normal vector measured by the right camera.

[0086] Similarly, the left and right cameras can measure two sets of circle center coordinate values respectively, obtain the final circle center coordinates according to the least squares method, and solve the circular hole radius value based on the obtained optimal circle center coordinate values to complete the calculation of the target circular hole pose parameters.

[0087] To test the algorithm's efficiency and verify the super-resolution reconstruction performance and edge enhancement effects of the present invention, this example first used a Basler acA2440-20gm high-resolution camera (2448px × 2048px) to capture 350 images of circular hole parts. These images were then expanded to 2000 images using image augmentation methods such as flipping and cropping to form a self-built sample library. Public super-resolution datasets, including DIV2K and Flickr2K, totaling 3650 image pairs, were introduced as supplementary datasets. These were randomly divided into training, test, and validation sets in a 96%:2%:2% ratio. The evaluation metrics used were the peak signal-to-noise ratio (PSNR) of the local image, the structural similarity (SSIM), and the image perceptual quality (PI). The comparison results are shown in Table 1. As can be seen from Table 1, compared to the current leading deep network models for super-resolution reconstruction, the proposed method achieves better reconstruction results that are more consistent with human visual perception.

[0088] Table 1 Comparison of super-resolution reconstruction results of circular hole parts

[0089]

[0090] In addition, through the method of the present invention, typical parts with holes are selected and their target circular hole positions are measured. The measurement results are shown in the figure below. Figure 7 As shown, Figure 7 Figures (a) and (b) show a binocular camera image of a hole-containing part with checkerboard corners, and the corresponding ellipse detection and fiducial point detection images, respectively. Table 2 shows a comparison with other methods. Compared to Zhang's measurement algorithm, which does not use super-resolution reconstruction, our method demonstrates superior measurement accuracy and stability.

[0091] Table 2 Experimental results when the distance between the parts and the binocular camera is 215 mm

[0092]

[0093] The experimental results above demonstrate that the present invention achieves edge enhancement during the super-resolution reconstruction of circular aperture images. Furthermore, the accuracy of the circular aperture pose information measured in binocular scenes is significantly improved. Furthermore, compared to existing algorithms, the present invention offers improved real-time performance, meeting the high-speed, high-efficiency, and high-precision measurement requirements of practical applications.

[0094] The above embodiments should not be regarded as limiting the present invention, but any improvements made based on the spirit of the present invention should be within the scope of protection of the present invention.

Claims

1. A circular hole pose measurement method based on edge enhancement in binocular vision scene, characterized in that: The following steps are involved: 1) Use a binocular camera to capture parts containing holes to obtain multiple high-resolution circular hole images. Preprocess each high-resolution circular hole image to obtain a corresponding blurred circular hole image. The dataset consists of each high-resolution circular hole image and its corresponding blurred circular hole image. 2) Inputting the data set into the circular hole image super-resolution reconstruction deep network to train the network and obtain a trained circular hole image super-resolution reconstruction deep network; 3) A binocular camera captures a high-resolution circular hole image to be reconstructed and inputs it into a trained deep network for circular hole image super-resolution reconstruction. The network outputs an edge-enhanced high-resolution circular hole image. The deep network for circular hole image super-resolution reconstruction consists of a cascade of a pixel rearrangement mechanism module and an improved enhanced generative adversarial network. The pixel rearrangement mechanism module is obtained by performing the inverse operation of the pixel upsampling module pixel-shuffle, and the improved enhanced generative adversarial network is obtained by performing cross-layer residual connections on the three RRDB blocks in the enhanced generative adversarial network. 4) Using the circular hole edge sub-pixel extraction method, sub-pixel edge extraction is performed on the edge-enhanced high-resolution circular hole image to obtain a sub-pixel edge image; The step 4) is specifically as follows: 4.1) The edge-enhanced high-resolution circular hole image is input into the circular hole semantic segmentation model for prediction, and the circular hole edge mask is output; 4.2) Use the edge extraction operator to extract the edge of the edge-enhanced high-resolution circular hole image to obtain the initial edge image of the part. Multiply the initial edge image of the part with the circular hole edge mask to obtain the rough edge image of the part. 4.3) Perform fine edge extraction on the rough edge image of the part to obtain an accurate edge image; 4.4) Calculate and obtain a sub-pixel edge image based on the edge-enhanced high-resolution circular hole image and the accurate edge image; 5) The parameters of the target circular hole are obtained by fitting the circular hole parameters of the sub-pixel edge image using the improved Hough transform method combined with geometric features; 6) Based on the parameters of the target circular hole, the super-resolution reconstruction camera observation model is used to perform coordinate mapping to obtain the position information of the target circular hole in the camera coordinate system; 7) According to the pose information of the target circular hole in the camera coordinate system, the stereo matching method based on the binocular camera is used to calculate the pose parameters of the target circular hole in the world coordinate system to complete the circular hole pose measurement.

2. The method for measuring circular hole pose based on edge enhancement in binocular vision scene according to claim 1, characterized in that: In the step 1), each high-resolution circular hole image is subjected to degradation processing to obtain a corresponding low-resolution circular hole image, and each pair of circular hole images is formed by each high-resolution circular hole image and the corresponding low-resolution circular hole image. The circular hole image pairs are then fuzzy filled to obtain fuzzy circular hole images.

3. The circular hole pose measurement method based on edge enhancement in a binocular vision scene according to claim 1 is characterized in that: The L2 loss function in the perceptual loss function of the enhanced adversarial generative network is replaced by a contextual loss function.

4. The method for measuring circular hole pose based on edge enhancement in binocular vision scene according to claim 1, characterized in that: In step 4.1), the circular hole semantic segmentation model is to add an attention mechanism module before the pyramid pooling ASPP module of the Deeplabv3+ semantic segmentation network, and the pyramid pooling ASPP module and the attention mechanism module are cascaded.

5. The circular hole pose measurement method based on edge enhancement in a binocular vision scene according to claim 1, characterized in that: In step 4.2), the edge extraction operator is specifically the Roberts operator.

6. The circular hole pose measurement method based on edge enhancement in a binocular vision scene according to claim 1, characterized in that: The step 4.3) is specifically as follows: The rough edge image of the part is sequentially subjected to hole edge fitting, connection and refinement to obtain an accurate edge image.

7. The method for measuring circular hole pose based on edge enhancement in binocular vision scene according to claim 1, characterized in that: The step 4.4) is specifically as follows: The sub-pixel points of the whole image of the predicted part with holes are extracted to obtain the sub-pixel points of the whole image. The sub-pixel points of the whole image are intersected with the accurate edge image to obtain the sub-pixel points of the circular hole edge, thereby obtaining the sub-pixel edge image.

8. The method for measuring circular hole pose based on edge enhancement in binocular vision scene according to claim 1, characterized in that: The super-resolution reconstruction camera observation model is specifically a mapping relationship between the camera coordinate system of the binocular camera and the virtual coordinate system of the reconstructed image.

Images