An ultra-resolution image construction method for hole shaft high-precision assembly pose measurement

Abstract

The application discloses a kind of for hole shaft high-precision assembly pose measurement super-resolution image construction method, the super-resolution image construction method for hole shaft high-precision assembly pose measurement includes steps as follows: S1: first generate the hole shaft assembly super-resolution image training dataset for training neural network;S2: construct the super-resolution neural network structure for hole shaft high-precision assembly pose measurement, a kind of for hole shaft high-precision assembly pose measurement super-resolution image construction method of the application, super-resolution data acquisition method for hole shaft structure image is designed, the problem that training effect is not good caused by downsampling dataset is solved;Proposed EESR with EEDBB as core structure, new super-resolution network structure has advantage in improving pose measurement precision.

Description

Technical Field

[0001] This invention relates to a method for constructing super-resolution images, and particularly to a method for constructing super-resolution images for high-precision assembly pose measurement of holes and shafts, belonging to the field of machine vision measurement technology. Background Technology

[0002] Image resolution is crucial for the accuracy of visual measurements. However, on the one hand, increasing the resolution of acquisition equipment is prohibitively expensive; on the other hand, image resolution inevitably decreases when photographing distant objects, which is particularly common in the pose measurement of large hole-shaft structures. For visual measurements, image quality is critical to accuracy. The higher the image resolution, the richer and clearer the texture details presented. Image resolution largely determines the accuracy of visual measurements. The most direct way to improve image resolution is to improve the hardware used for acquisition or imaging. However, on the one hand, the high cost limits the widespread adoption of such methods; on the other hand, in practice, even with high-quality, high-resolution acquisition equipment, low resolution is still encountered when photographing distant targets. In the pose measurement of large hole-shaft structures, due to the large size of the object being measured and the assembly equipment, visual measurements are often performed at large working distances (approaching or exceeding 1m), while successful assembly requires high measurement accuracy (better than 0.03mm).

[0003] In recent years, although deep learning super-resolution technology has been widely developed, especially in the fields of intelligent monitoring, medical imaging and remote sensing, it is still rarely used in the field of pose measurement. The reasons are as follows: First, there is a lack of corresponding training datasets. The existing dataset images are too different from the images of the objects to be measured, resulting in poor performance in actual use. Second, the existing network architectures either have too many parameters, making it difficult to meet the requirements of industrial efficiency, or they are ineffective and cannot meet the stringent measurement accuracy requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a super-resolution image construction method for high-precision assembly pose measurement of hole shafts, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a super-resolution image construction method for high-precision assembly pose measurement of hole shafts, comprising the following steps:

[0006] S1: First, generate a training dataset of super-resolution images of hole-axis assemblies for training the neural network;

[0007] S2: Construct a super-resolution neural network structure for high-precision assembly pose measurement of hole shafts.

[0008] As a preferred embodiment of the present invention, step S1 includes the following steps:

[0009] A1: Data acquisition. First, use the same short focal length to capture all the required low-resolution aperture axis feature images, and then switch to a telephoto lens to capture the corresponding high-resolution aperture axis feature images.

[0010] A2: Image alignment. A checkerboard image is displayed on a monitor. Before the separate acquisition processes at low and high resolutions, the checkerboard images displayed on the monitor are captured using short-focal-length and long-focal-length cameras, respectively. The corner features of the checkerboard images are identified to calculate the alignment relationship between the high-resolution and low-resolution images.

[0011] As a preferred embodiment of the present invention, in step A1, a black image is displayed on the monitor at the beginning of each shooting process, and then subtracted from each captured image; the subject of the shooting is an ultra-high-quality monitor, the tool used is a digital camera, and all operations are remotely operated in a dark room to prevent any interference from ambient light and any interference from the monitor and camera.

[0012] As a preferred embodiment of the present invention, the neural network structure in step S2 consists of: 1 convolutional layer, 24 residual blocks, 8 EEDBB layers, 1 convolutional layer, 1 upsampling layer, and 1 convolutional layer.

[0013] As a preferred embodiment of the present invention, the EEDBB layer output is performed according to the following formula:

[0014] R c =R n +R in +R s +R l

[0015] In the formula, Rn is the image output after the input image undergoes normal convolution, used to ensure the basic performance of the neural network structure; Rin is the image output after the input image undergoes shortcut connections, used to realize the residual; Rs is an operator including horizontal and vertical Sobel filters. Where Sx and Sy are the horizontal and vertical Sobel filters, and Wsx and Wsy are the coefficients of the convolution with the Sobel filter. is the convolution operator; Rl is the operator containing the Laplacian filter. Where Lx is the Laplacian filter, and Wlx are the coefficients of the convolution with the Laplacian filter. It is a convolution operator.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention proposes a super-resolution image construction method for high-precision assembly pose measurement of hole shafts, including a new deep learning super-resolution network structure and a matching hole shaft image super-resolution dataset acquisition method. The deep learning super-resolution network enhances the perception of edge information in the image through the design of the core structure.

[0017] (1) A super-resolution data acquisition method for hole shaft structure images was designed, which solved the problem of poor training effect caused by downsampled datasets;

[0018] (2) An EESR with EEDBB as the core structure was proposed. The new super-resolution network structure has advantages in improving the accuracy of pose measurement. Attached Figure Description

[0019] Figure 1 This is the neural network structure of the present invention;

[0020] Figure 2 This is the low-resolution-high-resolution image alignment process in this invention;

[0021] Figure 3 This is the EEDBB network structure in this invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention have been clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1-3 This invention provides a super-resolution image construction method for high-precision assembly pose measurement of hole shafts. The super-resolution image construction method for high-precision assembly pose measurement of hole shafts includes the following steps:

[0024] S1: First, generate a training dataset of super-resolution images of hole-axis assemblies for training the neural network;

[0025] S2: Construct a super-resolution neural network structure for high-precision assembly pose measurement of hole shafts.

[0026] Step S1 includes the following steps:

[0027] A1: Data acquisition. First, use the same short focal length to capture all the required low-resolution aperture axis feature images, and then switch to a telephoto lens to capture the corresponding high-resolution aperture axis feature images.

[0028] A2: Image alignment. A checkerboard image is displayed on a monitor. Before the separate acquisition processes at low and high resolutions, the checkerboard images displayed on the monitor are captured using short-focal-length and long-focal-length cameras, respectively. The corner features of the checkerboard images are identified to calculate the alignment relationship between the high-resolution and low-resolution images.

[0029] Furthermore, in step A1, a black image is displayed on the monitor at the beginning of each shooting process, and then subtracted from each captured image; the subject of the shooting is an ultra-high-quality monitor, the tool used is a digital camera, and all operations are remotely operated in a dark room to prevent any interference from ambient light and any interference from the monitor and camera.

[0030] Furthermore, the neural network structure in step S2 consists of: 1 convolutional layer, 24 residual blocks, 8 EEDBB layers, 1 convolutional layer, 1 upsampling layer, and 1 convolutional layer.

[0031] Preferably, the output of the EEDBB layer is performed according to the following formula:

[0032] R c =R n +R in +R s +R l

[0033] In the formula, Rn is the image output after the input image undergoes normal convolution, used to ensure the basic performance of the neural network structure; Rin is the image output after the input image undergoes shortcut connections, used to realize the residual; Rs is an operator including horizontal and vertical Sobel filters. Where Sx and Sy are the horizontal and vertical Sobel filters, and Wsx and Wsy are the coefficients of the convolution with the Sobel filter. is the convolution operator; Rl is the operator containing the Laplacian filter. Where Lx is the Laplacian filter, and Wlx are the coefficients of the convolution with the Laplacian filter. It is a convolution operator.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A super-resolution image construction method for hole shaft high-precision assembly pose measurement, characterized in that, The method comprises the following steps: Step S1: first generate a hole shaft assembly super-resolution image training data set for training a neural network; Step S2: construct a super-resolution neural network structure for hole shaft high-precision assembly pose measurement; The neural network structure level of step S2 is: 1 convolution layer, 24 residual blocks, 8 EEDBB layers, 1 convolution layer, 1 up-sampling layer and 1 convolution layer; The EEDBB layer output is according to the following formula: R c = R n + R in + R s + R l wherein R n is the output image of the input image after general convolution, used to ensure the basic performance of the neural network structure; R in is the output image of the input image after shortcut connection, used to realize the residual; R s is an operator including horizontal and vertical Sobel filters, R s = (S x W sx ) + (S y W sy ), wherein S x and S y are horizontal and vertical Sobel filters, W sx and W sy are coefficients for convolution with the Sobel filter, is a convolution operator; R l is an operator containing a Laplacian filter, R l =L x W lx , wherein L x is a Laplacian filter, and W lx is a coefficient for convolution with the Laplacian filter.

2. The super-resolution image construction method for hole shaft high-precision assembly pose measurement according to claim 1, characterized in that: Step S1 comprises the following steps: Step A1: data acquisition, first capture all the required low-resolution hole shaft feature images using the same short focal length, then replace the long focal lens to capture the corresponding high-resolution hole shaft feature images; Step A2: image alignment, display the checkerboard image on the monitor, and capture the checkerboard image displayed on the monitor using the short focal length and long focal length cameras before the respective acquisition processes of the low-resolution and high-resolution images respectively, identify the corner features of the checkerboard image for calculating the alignment relationship between the high-resolution and low-resolution images.

3. The super-resolution image construction method for hole shaft high-precision assembly pose measurement according to claim 2, characterized in that: In step A1, a black image is displayed on the display at the beginning of each shooting process, and then the black image is subtracted from each shot image; the shooting object is a super high quality display, the tool used is a digital camera, and all operations are remotely controlled in a dark room to prevent any environmental light interference and any interference action of the monitor and the camera.

Images