Online measurement method for rolling deformation geometry state of hot ring based on deep learning

Abstract

The application provides a hot ring rolling deformation geometry state online measurement method based on deep learning, collects a hot ring image, trains a template image by using collected image samples, matches a test image and an input image by using a template matching algorithm, enhances a real-time image, captures a higher-precision ring edge, detects actual ring edge points, and realizes the function of online measurement of the hot ring rolling deformation geometry state under the interference of non-uniform temperature color difference, flying sparks and rolling block shielding and the like by fitting. The application solves the problem that the ring geometry state feature cannot be quickly and accurately extracted due to the shielding of the rolling block to the target ring and the flying of the oxide skin, gas mist and the like during the hot ring rolling process, ensures the matching accuracy, speeds up the target matching operation efficiency, and improves the overall robustness of the algorithm.

Description

Technical Field

[0001] This invention belongs to the field of machine vision technology, specifically relating to an online measurement method for the geometric state of hot-state rolling deformation of ring parts based on deep learning. Background Technology

[0002] As core components in the automotive, shipbuilding, metallurgy, chemical, and aerospace industries, the quality and performance of ring components are of paramount importance. Ring rolling is an advanced forming and manufacturing technology for high-performance seamless ring components. Large ring components are typically produced using hot rolling. However, because ring components are at high temperatures during the initial stage of hot rolling, existing contact measurement technologies are difficult to apply to such high-temperature environments.

[0003] Because large rings are at a high temperature of 1250°C in the initial stage of hot rolling, the visible light radiated by the hot ring body interferes with optical measurements, increasing measurement errors. At different rolling stages, the rolls also provide some shielding to the hot ring, and there are also issues such as the outer skin of the ring peeling off and splashing, as well as misting. Therefore, non-contact measurement methods, such as visual measurement, struggle to guarantee accuracy and efficiency during real-time monitoring.

[0004] Machine vision offers significant advantages in measuring the geometric state of hot ring components due to its non-contact nature, high efficiency, and comprehensive information acquisition. However, the visible light radiated by the hot ring component interferes with optical measurements, increasing measurement errors. Furthermore, the drastic and time-varying deformation state of the ring component during rolling places high demands on the measurement system's response. Existing vision measurement methods generally suffer from low efficiency, poor robustness, and low accuracy, making it impossible to achieve precise visual measurements of the object under test.

[0005] Therefore, there is an urgent need for a visual measurement method and equipment to solve the above problems. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide an online measurement method for the deformation geometry of hot-state rolling rings based on deep learning, which is used to measure the deformation geometry parameters of rings during hot rolling under disturbance.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: an online measurement method for the geometric state of hot-state rolling deformation of ring parts based on deep learning, comprising the following steps:

[0008] S1: Obtain rolling images of the hot ring in the hot rolling experiment using an image acquisition device, and perform target detection on the rolling images of the hot ring using a deep learning algorithm to determine the corresponding template image and feature image;

[0009] S2: Use an image edge matching algorithm to match the template image and the rolling image;

[0010] S3: Use a weighted function to fit the boundary profile of the hot ring and calculate the outer diameter of the hot ring.

[0011] According to the above scheme, the specific steps in step S1 are as follows:

[0012] S11: Acquire real-time rolling images of the hot ring during the hot rolling experiment using an image acquisition device;

[0013] S12: The rolling image is used as an image sample and randomly divided into the training set, test set and validation set according to a preset ratio. The weight parameters of the convolutional neural network are trained and updated iteratively using the training set through a deep learning algorithm. The rolling image acquired by the image acquisition device is used to perform coarse target recognition to obtain the template image corresponding to the rolling image. The target area is selected on the template image to capture the target hot ring component.

[0014] S13: Slice the original-sized image sample to obtain the first feature map; convolve the first feature map with a convolution kernel to obtain the feature image.

[0015] Furthermore, step S1 also includes the following steps:

[0016] The actual dimensions of the hot ring component were calibrated using the Zhang Zhengyou calibration method and a calibration plate to correct lens distortion.

[0017] Furthermore, in step S2, the specific steps are as follows:

[0018] S21: Input a template image, perform affine transformations on the template image including translation, rotation, and scaling at different angles or scales, and calculate the feature points and feature vectors of the template image;

[0019] S22: Input the rolling image, calculate the pixel position, and construct the gradient response map;

[0020] S23: Use image edge matching to calculate the similarity between the rolling image and the template image to complete template matching.

[0021] Furthermore, in step S3, the specific steps are as follows:

[0022] S31: Enhance the image obtained in step S2, perform linear transformation on the pixels, highlight bright areas, suppress dark areas, improve edge contrast, and capture high-precision hot ring edges.

[0023] S32: Use a caliper tool to generate caliper rectangles at equal intervals on the circular contour fitted by the algorithm, perform edge point detection, and filter out the actual edge points with larger gradient magnitudes within the caliper.

[0024] S33: Introduce a weighting function to iteratively reweight and fit the actual edge points, use weighted least squares to fit and suppress outliers, and calculate the geometric parameters of the hot ring.

[0025] Furthermore, in step S22, the specific steps for constructing the gradient response map are as follows:

[0026] Discretize the gradient into several directions with equal angular spacing, and use binary bits to represent the quantized gradient directions to simplify the amount of information in the gradient representation.

[0027] The computational process of comparing the template image with the search image at all locations by quantizing the gradient to build a pre-response table is transformed into a lookup table process.

[0028] The template matching similarity measure is obtained by finding the corresponding data in the linear storage table based on the feature points of the template image and adding the corresponding data of the feature points.

[0029] Furthermore, in step S31, the specific steps for enhancing the image are as follows:

[0030] Let GMin be the gray value, g be the current gray value, Mult be the multiplier, and Add be the additional offset value. Image enhancement is performed according to equations (1) and (2):

[0031] Add = -Mult*GMin (1),

[0032] g' = g*Mult + Add (2).

[0033] An online measurement system for the geometric state of hot-state rolling deformation of a ring component based on deep learning includes an image acquisition device, an industrial control computer, and a light source. The image acquisition device uses an industrial camera to acquire images of the ring component. The industrial control computer is used to process the acquired images and calculate the geometric state parameters of the hot-state ring component. The light source is used to provide illumination or supplementary lighting.

[0034] A computer storage medium storing a computer program executable by a computer processor, the computer program executing a deep learning-based online measurement method for the geometric state of hot-state rolling deformation of a ring.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. The present invention provides an online measurement method for the geometric state of hot ring rolling deformation based on deep learning. This method acquires images of the hot ring, uses the acquired image samples for training to obtain template images, uses a template matching algorithm to match the test image with the input image, enhances the real-time image, captures the ring edge with high precision, detects the actual edge points of the ring, and achieves the function of online measurement of the geometric state of hot ring rolling deformation under interferences such as non-uniform temperature color difference, flying chips, and roll obstruction through fitting.

[0037] 2. This invention and method combine coarse target recognition and fine matching. It uses deep learning algorithms to match the image edge templates to obtain a high-precision fitted contour of the target, and uses a weighted function to correct the fitting effect, thereby obtaining the real-time geometric state of the hot ring rolling process. By using deep learning algorithms and training on the dataset, the method can intelligently capture the target ring and measure its geometric state using rolling images acquired by the camera, thus solving the problem of not being able to accurately obtain the target image.

[0038] 3. This invention solves the problem that the geometric features of the ring cannot be extracted quickly and accurately during the hot ring rolling process due to the obstruction of the target ring by the rolls and the presence of splashed oxide scale, mist, etc. While ensuring matching accuracy, it speeds up the target matching calculation efficiency and improves the overall robustness of the algorithm.

[0039] 4. This invention solves the problems of time-consuming and safety hazards associated with manual contact measurement, and addresses the issue of low target extraction accuracy caused by the detachment and splashing of the outer skin of the ring during rolling. It improves the speed of the image matching process based on edge matching optimization and improves the fitting accuracy based on iterative reweighted fitting, thus ensuring the online measurement effect and accuracy requirements during the hot ring rolling process.

[0040] 5. This method uses a deep learning algorithm to perform target detection on the hot ring image, employs an image edge matching algorithm to perform template matching between the ring template image and the rolled ring image, and uses a weight function to fit the ring boundary contour and calculate the outer diameter of the ring. The calculation process of comparing the template model with the search images at all positions is transformed into a lookup table process, which improves the lookup speed of the matching process and its robustness to small positions and rotations. Attached Figure Description

[0041] Figure 1 This is a flowchart of an embodiment of the present invention.

[0042] Figure 2 This is a system layout diagram according to an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram of the measurement results according to an embodiment of the present invention. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0045] See Figure 1 The embodiments of the present invention include an image acquisition device and an industrial control computer; the image acquisition device uses an industrial camera to acquire images of the ring component; the industrial control computer is used to process the acquired images and calculate the geometric state parameters of the hot ring component; the light source is used to provide illumination or supplementary lighting.

[0046] The present invention provides an online measurement method for the geometric state of hot-state rolling deformation of a ring component based on deep learning, comprising the following steps:

[0047] (1) Obtain a real-time image of the hot ring under test and determine the template image corresponding to the image, wherein the region of interest has been selected on the template image.

[0048] (2) The actual dimensions of the hot-state ring are calibrated using a calibration plate to correct lens distortion and ensure accurate ring dimensions. The calibration method adopted is the Zhang Zhengyou calibration method.

[0049] (3) Hot-state ring images are collected from the hot rolling test of the ring. The hot-state ring images are divided into training, testing and validation sets according to a certain ratio for training to obtain template images. The template images are trained by a convolutional neural network and have easily identifiable feature points and feature vectors.

[0050] (4) Further, the template information and the test image are input into the system, and the similarity of feature points is calculated to complete the template matching. The template information is trained by scaling at different angles and scales before input.

[0051] (5) Further, by enhancing the image and performing linear transformation on the pixels, the edge contrast is improved, and the edges of the ring are captured with higher precision. The actual edge points with large gradient magnitudes within the calipers are detected by using a caliper tool, and a weighting function is introduced to weight the data and perform circle fitting.

[0052] (6) The hot ring rolling image captured in real time by the industrial camera is transmitted to the industrial control computer. The industrial control computer processes the image synchronously and calculates the size of the hot ring by fitting the above-mentioned circle.

[0053] An industrial computer controls an industrial camera to acquire images of a hot ring component. The acquired image samples are used to train template images. A template matching algorithm is then used to match the test image with the input image. This enhances the real-time images obtained by the industrial camera, capturing high-precision ring component edges and detecting the actual edge points. Through fitting, the dimensions of the hot ring component are measured online. The specific operation is as follows:

[0054] Images of hot-state ring components obtained from hot rolling experiments are used as samples and randomly divided into training, testing, and validation sets according to proportions. Training is then performed on these images. The training process includes slicing the input images to obtain feature maps, and then convolving these feature maps with convolutional kernels to obtain more feature maps. Predicted bounding boxes are output based on initial anchor boxes, and the network parameters are updated iteratively after calculating the discrepancies.

[0055] The test image captured by the industrial camera and the aforementioned template feature information are imported into an image edge matching algorithm to calculate the similarity of feature points. Before template matching, the template information is trained by scaling it at different angles and scales to obtain the feature points and feature vectors of the template image. Pixel positions are calculated on the input test image to construct a gradient response map.

[0056] The aforementioned gradient response map is constructed by discretizing the gradient into several directions with equal angular spacing, and representing the quantized gradient directions using binary bits, thus simplifying the information content of the gradient representation. The computational process of comparing the template model with the search image at all locations is transformed into a lookup table process. Based on the template feature points, the corresponding data in the linear storage table is looked up, and the corresponding data for each feature point are summed to obtain the template matching similarity metric.

[0057] Image enhancement is performed according to equations (1) and (2):

[0058] Add=-Mult*GMin (1)

[0059] g'=g*Mult+Add (2)

[0060] Linear transformation is applied to pixels to improve edge contrast and capture high-precision ring edges.

[0061] The profile of the hot ring component was fitted to a circle using a fitting algorithm, and actual edge points with large gradient magnitudes within the caliper were detected. Weighted least squares fitting was then used to suppress outliers, resulting in better measurement results for the hot ring component's dimensions.

[0062] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A deep learning-based online measurement method for the geometric state of hot-state rolling deformation of ring parts, characterized in that: Includes the following steps: S1: Obtain rolling images of the hot ring in the hot rolling experiment using an image acquisition device, and perform target detection on the rolling images of the hot ring using a deep learning algorithm to determine the corresponding template image and feature image; The specific steps are as follows: S11: Acquire real-time rolling images of the hot ring during the hot rolling experiment using an image acquisition device; S12: The rolling image is used as an image sample and randomly divided into the training set, test set and validation set according to a preset ratio. The weight parameters of the convolutional neural network are trained and updated iteratively using the training set through a deep learning algorithm. The rolling image acquired by the image acquisition device is used to perform coarse target recognition to obtain the template image corresponding to the rolling image. The target area is selected on the template image to capture the target hot ring component. S13: Slice the original-sized image sample to obtain the first feature map; convolve the first feature map with a convolution kernel to obtain the feature image; S2: Use an image edge matching algorithm to match the template image and the rolling image; S3: Fit the boundary profile of the hot ring component using a weighted function and calculate the outer diameter of the hot ring component; the specific steps are as follows: S31: Enhance the image obtained in step S2, perform linear transformation on the pixels, highlight bright areas, suppress dark areas, improve edge contrast, and capture high-precision hot ring edges. S32: Use a caliper tool to generate caliper rectangles at equal intervals on the circular contour fitted by the algorithm, perform edge point detection, and filter out the actual edge points with larger gradient magnitudes within the caliper. S33: Introduce a weighting function to iteratively reweight and fit the actual edge points, use weighted least squares to fit and suppress outliers, and calculate the geometric parameters of the hot ring.

2. The online measurement method for the geometric state of hot-state rolling deformation of a ring based on deep learning according to claim 1, characterized in that: Step S1 further includes the following steps: The actual dimensions of the hot ring component were calibrated using the Zhang Zhengyou calibration method and a calibration plate to correct lens distortion.

3. The online measurement method for the geometric state of hot-state rolling deformation of a ring based on deep learning according to claim 1, characterized in that: The specific steps in step S2 are as follows: S21: Input a template image, perform affine transformations on the template image including translation, rotation, and scaling at different angles or scales, and calculate the feature points and feature vectors of the template image; S22: Input the rolling image, calculate the pixel position, and construct the gradient response map; S23: Use image edge matching to calculate the similarity between the rolling image and the template image to complete template matching.

4. The online measurement method for the geometric state of hot-state rolling deformation of a ring based on deep learning according to claim 3, characterized in that: In step S22, the specific steps for constructing the gradient response map are as follows: Discretize the gradient into several directions with equal angular spacing, and use binary bits to represent the quantized gradient directions to simplify the amount of information in the gradient representation. The computational process of comparing the template image with the search image at all locations by quantizing the gradient to build a pre-response table is transformed into a lookup table process. The template matching similarity measure is obtained by finding the corresponding data in the linear storage table based on the feature points of the template image and adding the corresponding data of the feature points.

5. The online measurement method for the geometric state of hot-state rolling deformation of a ring based on deep learning according to claim 1, characterized in that: In step S31, the specific steps for enhancing the image are as follows: set up GMin Grayscale value g This is the current grayscale value. Mult The coefficient being multiplied. Add To add offset values, image enhancement is performed according to equations (1) and (2): (1)， (2)。 6. A measurement system for the online measurement method of the geometric state of hot-state rolling deformation of a ring as described in any one of claims 1 to 5, characterized in that: Includes image acquisition device, industrial control computer and light source; The image acquisition device uses an industrial camera to acquire images of the components. The industrial computer is used to process acquired images and calculate the geometric state parameters of hot-state components; Light sources are used to provide illumination or supplemental lighting.

7. A computer storage medium, characterized in that: It contains a computer program that can be executed by a computer processor, which performs the online measurement method for the geometric state of hot ring rolling deformation based on deep learning as described in any one of claims 1 to 5.

Images