A deep convolutional regression laser coining algorithm

Abstract

The present application belongs to the technical field of computer vision and laser steering gear calibration application, and particularly relates to a deep convolution regression laser calibration algorithm, which realizes calibration of the position of laser emitted by a low-precision steering gear laser emitter on a plane through computer vision means, and realizes laser irradiation on a target position on a fixed plane. In the process, the most core calibration process of the steering gear turning angle and the laser irradiation position is changed from a traditional precise measurement mode to an end-to-end computer vision calibration mode. The scheme greatly reduces the requirement of the steering gear device precision and the calibration time, and improves the device precision. The noise enhancement function proposed in the present application can only control the overall intensity and range of noise, but also can adjust the detailed characteristics of noise, so as to meet more complex data enhancement requirements. The use of a multi-scale deep convolution regression model in the present application makes the two-way cross learning improve the positioning ability of the laser landing point and reduce external environmental interference.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and laser servo calibration application technology, and particularly relates to a deep convolutional regression laser calibration algorithm. Background Technology

[0002] Currently, the calibration and control process of rotary servos largely relies on precise measurement of their physical position, followed by calibrating the servo. This method requires sophisticated measuring tools and fixed equipment, and is highly susceptible to external environmental influences, making it inflexible. Relative calibration of the servo from the current perspective is time-consuming and increases the demands on the measuring and fixed equipment. Furthermore, the accuracy of the servo can lead to random instability. Summary of the Invention

[0003] This invention addresses the technical problems existing in the calibration control process of rotary servo motors by proposing a deep convolutional regression laser calibration algorithm that is rationally designed, simple in method, theoretically sound, and can save calibration time, reduce the requirements for measuring and fixing equipment, and suppress random instability.

[0004] To achieve the above objectives, the technical solution adopted by this invention is: a deep convolutional regression laser calibration algorithm, comprising the following steps:

[0005] S1. Data Acquisition: First, fix the positions of the screen, servo-laser emitter, and camera. The computer initiates commands to control the servo's rotation and laser illumination. The servo rotates 1 degree vertically (h-axis) and horizontally (v-axis) each time. Therefore, if the rotation angle from the leftmost to the rightmost edge of the screen is 60 degrees when h is fixed, and similarly, the rotation angle from top to bottom is 30 degrees when v is fixed, the servo will sample a total of 30 × 60 = 1800 times. Each time the servo switches angles, the camera captures an image, naming the image with the hv value at the time of acquisition, and recording the angle position of the laser servo device that generated the image. The data set of angle positions is obtained by summarizing them. ), to obtain the original images, and construct 1800 basic data images;

[0006] S2. Complete the missing laser impact points to reconstruct the original acquired image. ,in and These represent the original image and the noisy image, respectively. The noisy image is filled with missing laser points using an approximation method.

[0007] S3. Noise enhancement: The laser impact points in the original acquired images are labeled, and a target detection base model, named model_base, is trained based on the YOLOv8 model. This model can effectively identify the pixel positions of the laser impact points in the image. ), and the pixel location data set is obtained by summarizing them. ), retrieve data pairs ( ),( Based on the data set, a multinomial regression model is trained and named Reg_location, and the following applies:

[0008]

[0009] Among them, when inputting a pixel position on the screen At that time, the polynomial model Reg_location can calculate the angular position of the laser servo device that generated the image. A certain number of (x,y) coordinate pairs are randomly generated within the range of (x,y) and input into the Reg_location model to obtain the base (h,v) values. New (x,y) coordinate values ​​are then generated based on the noise enhancement function. The calculation of the noise enhancement function for x is shown below:

[0010]

[0011] in, The main control is the basic intensity of noise. , The combined effect acts on the frequency and phase of noise, influencing the morphology of the noise pattern. Used to fine-tune noise intensity, providing more precise control. To adjust the range and complexity of the noise, a term combining sine and cosine was added, making the noise patterns more diverse.

[0012] S4. Multi-scale deep convolutional regression: A weight mask W is constructed using a standard normal distribution, with the same size as the original image. The weight mask W is used to adjust the distribution of noise. The calculation method is as follows:

[0013]

[0014] in, This is the image at step t. It is noise sampled from a standard normal distribution. To control the overall noise intensity, a multi-scale deep convolutional regression model is constructed. The model input is an image and the angular position of the laser servo device that generated the image. The training model is required to determine the angular position of the laser servo device at the moment of laser impact in a given image. A dual-input network architecture is constructed to process two input streams simultaneously, corresponding to the original image and the noisy image respectively. Finally, these two input streams are fused at a certain point to predict the angular position of the laser servo device. The multi-scale deep convolutional regression model introduces a feature interaction layer between two parallel feature extraction sub-networks to dynamically adjust the feature extraction process for the noisy image. and noisy images Features are extracted through their respective CNN subnetworks, denoted as . and noisy images Then use As a query Q, As keys K and values ​​V, weighted noisy image features are calculated using an attention mechanism. Then use the obtained weighted features The feature representation of the noisy image is updated to guide the learning of the noisy image feature extraction subnetwork. The calculation process is as follows:

[0015]

[0016] in, It is a function that decreases with t, used to control the degree of influence of the original image features on the update of noisy image features. Its calculation process is as follows:

[0017]

[0018] Where a(t) is the amplitude adjustment function, which controls the main variation of λ(t), and is a function that gradually decreases as t decreases, as follows:

[0019]

[0020] in, It is the initial amplitude. The rate at which the amplitude decreases is controlled. b(t) is a frequency adjustment function that determines the frequency of change of λ(t). It is set as a function that changes in stages to simulate the changing noise control requirements at different stages. The formula for b(t) is:

[0021]

[0022] in, The phase shift determines the initial change pattern of λ(t), which can be set according to the requirements during model initialization. Finally, the updated noise image features are... and original image features The fusion is performed, and the predicted angle position of the laser servo device is output through the fully connected layer.

[0023] Preferably, the multi-scale depth convolutional regression model in step S4 is divided into two parts: a dual-path localization module and a depth image regression module. The dual-path localization module takes the original image and the noise-enhanced image as inputs and enables dual-path cross-learning to improve the localization capability of the laser landing point. The depth image regression module obtains the angle position of the laser servo device required to generate the laser landing point.

[0024] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0025] This invention proposes a deep convolutional regression laser calibration algorithm. Using computer vision, it calibrates the position of a laser emitted by a low-precision servo laser emitter on a plane, achieving laser illumination of a target position on a fixed plane. The core calibration process, the alignment of the servo steering angle with the laser illumination position, is transformed from traditional precision measurement to end-to-end computer vision calibration. This significantly reduces the precision requirements of the servo equipment and the calibration time, while simultaneously improving equipment accuracy. The proposed noise enhancement function not only controls the overall intensity and range of noise but also adjusts its detailed characteristics, thus meeting more complex data augmentation needs. This invention uses a multi-scale deep convolutional regression model, enabling dual-path cross-learning to improve the localization capability of the laser impact point and reduce external environmental interference. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A top-view diagram showing the location of an embodiment of the present invention;

[0028] Figure 2 This is a frontal view diagram of the position provided in an embodiment of the present invention;

[0029] Figure 3 This is a laser impact point distribution diagram provided for an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the structure of a deep convolutional regression laser calibration algorithm provided in an embodiment of the present invention. Detailed Implementation

[0031] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0032] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0033] In this embodiment, considering that the calibration control process of a rotary servo largely relies on precise measurement of its physical position for calibration, this method requires sophisticated measuring tools and fixed equipment. Furthermore, it is highly susceptible to external environmental influences, inflexible, and wastes time performing relative calibration of the servo from the current viewpoint, increasing the demands on the measuring and fixed equipment. Additionally, the servo's accuracy can lead to random instability. Therefore, this invention proposes a depthwise convolutional regression laser calibration algorithm.

[0034] like Figure 1 , Figure 2 As shown, this solution first describes the application scenario of the servo motor and laser emitter. The laser emitter is first fixed to a servo motor with rotational function. The aim is to control the emission direction of the laser emitter by changing the rotation angle (h,v) of the servo motor, thereby controlling the laser's landing point on a fixed plane. The requirement for the entire device is to determine any position of the target displayed on the current screen using computer vision methods, and simultaneously illuminate the center point of that position with a laser. However, due to limitations in the stability, accuracy, and placement of the servo motor, traditional methods can only ensure that the laser beam emitted by the servo motor at an angle of (0,0) is perpendicular to the screen through precise measurement. This requirement is extremely difficult to measure, and the accuracy of the device is often affected by external environmental changes (accidental collisions, vibrations from personnel movement, etc.), making it difficult to obtain this ideal position. Therefore, this solution proposes to calibrate the laser-servo motor combination device by acquiring the laser's landing point on the screen using a vision device. This improves calibration accuracy and speed while eliminating dependence on the external environment, meaning the device can be placed anywhere within the screen area. The detailed method is as follows.

[0035] First, data acquisition: The positions of the screen, servo-laser emitter, and camera are fixed. The computer initiates commands to control the servo's rotation and laser illumination. The servo rotates 1 degree vertically (h-axis) and horizontally (v-axis) each time. Therefore, if the rotation angle from the leftmost to the rightmost edge of the screen is 60 degrees when h is fixed, and similarly, the rotation angle from top to bottom is 30 degrees when v is fixed, the servo will sample a total of 30 × 60 = 1800 times. Each time the servo switches angles, the camera captures an image, naming the image with the hv value at the time of acquisition, and recording the angle position of the laser servo device that generated the image. The data set of angle positions is obtained by summarizing them. The original images were obtained, and 1800 basic data images were constructed.

[0036] After the data acquisition is complete, the laser irradiation point corresponding to each angle is visualized onto a base plate image, such as... Figure 3 As shown, some laser impact points are missing. Furthermore, when the v value remains constant, the laser impact points generated by different h values ​​should be perpendicular to the bottom edge of the image; however, the impact points fluctuate vertically and are not strictly perpendicular. Based on this, this solution proposes introducing noise to fill in the missing laser impact points. Specifically, to fill in the missing laser impact points, the original acquired image is reconstructed as follows: ,in and The images represent the original image and the noisy image, respectively. The noisy image uses an approximation-based completion scheme to fill in the missing laser impact points. In this scheme, noise samples are not defined as negative samples, but rather as new samples generated by a specific strategy to fill in the missing laser impact points. For the missing impact point generation process, this scheme uses an approximation-based completion method, that is, selecting the average coordinates of the four nearest points (up, down, left, and right) around the missing location as the completion coordinates. Unlike conventional missing value completion, the aforementioned impact point fluctuation problem is a positional shift caused by the external environment and can be defined as noise; therefore, the completed data also contains noise.

[0037] To effectively determine the true coordinates of the laser impact point, it is unavoidable to remove or minimize the offset caused by noise. However, this type of noise is random and related to different devices and external environments. Therefore, this solution aims to train a model that can simulate and generate noise while suppressing it.

[0038] Specifically, noise enhancement is performed, the laser impact points in the original acquired images are labeled, and a target detection base model, named model_base, is trained based on the YOLOv8 model. This model can effectively identify the pixel positions of the laser impact points in the image. ), and the pixel location data set is obtained by summarizing them. ), retrieve data pairs ( ),( Based on the data set, a multinomial regression model is trained and named Reg_location, and the following applies:

[0039]

[0040] Among them, when inputting a pixel position on the screen At that time, the polynomial model Reg_location can calculate the angular position of the laser servo device that generated the image. A certain number of (x,y) coordinate pairs are randomly generated within the range of (x,y) and input into the Reg_location model to obtain the base (h,v) values. New (x,y) coordinate values ​​are then generated based on the noise enhancement function. The calculation of the noise enhancement function for x is shown below:

[0041]

[0042] in, The main control is the basic intensity of noise. , The combined effect acts on the frequency and phase of noise, influencing the morphology of the noise pattern. Used to fine-tune noise intensity, providing more precise control. To address this, a term combining sine and cosine was added to adjust the range and complexity of the noise, making the noise patterns more diverse. Through this design, we can not only control the overall intensity and range of the noise, but also adjust the detailed characteristics of the noise, thereby meeting more complex data augmentation needs.

[0043] Through the above operations, more sample data can be obtained. Then, using the (x,y) coordinate values ​​generated by the noise enhancement function, new sample images are created. That is, given an original acquired image, noise is gradually introduced into the image, generating a series of images until the image is entirely composed of noise. In the noise introduction process, unlike in the image generation domain, our goal is to adjust the position and shape of the laser impact point in the image (reflections at different angles will cause changes in the shape of the laser impact point), while ignoring the surrounding image parts. This is achieved by concentrating the noise mainly around the specified coordinates while reducing the influence of noise in other areas. The specific operations are as follows:

[0044] Multi-scale deep convolutional regression uses a standard normal distribution to construct a weight mask W, with the same size as the original image. The weight mask W is used to adjust the distribution of noise, and the calculation method is as follows:

[0045]

[0046] in, This is the image at step t. It is noise sampled from a standard normal distribution. The overall intensity of the noise is controlled in this way, so that the noise is mainly concentrated near the (x,y) coordinates, while in other areas it is modulated by W, and the noise has a smaller impact. A multi-scale deep convolutional regression model is constructed. The model input is an image and the angular position (h,v) of the laser servo device that generated the image. The training model is required to be able to determine the angular position of the laser servo device when the laser impact point is generated in the given image. A dual-input network architecture is constructed to process two input streams simultaneously, corresponding to the original image and the noisy image respectively. Finally, the two input streams are fused at a certain point to predict the angular position of the laser servo device. The multi-scale deep convolutional regression model is divided into two parts: a dual-path localization module and a depth image regression module. The dual-path localization module takes the original image and the noise-enhanced image as inputs respectively, and the dual-path cross-learning improves the localization ability of the laser impact point and reduces external environmental interference. The depth image regression module obtains the angular position of the laser servo device required to generate the laser impact point. The multi-scale deep convolutional regression model introduces a feature interaction layer between two parallel feature extraction sub-networks. This layer aims to enable features extracted from the original image to guide the learning process of the noisy image sub-network. Using an attention mechanism as the core of the feature interaction, it dynamically adjusts the noise image feature extraction process by calculating the influence weights of the original image features on the noise image features. and noisy images Features are extracted through their respective CNN subnetworks, denoted as . and noisy images Then use As a query Q, As keys K and values ​​V, weighted noisy image features are calculated using an attention mechanism. Then use the obtained weighted features The feature representation of the noisy image is updated to guide the learning of the noisy image feature extraction subnetwork. The calculation process is as follows:

[0047]

[0048] in, It is a function that decreases with t, used to control the degree of influence of the original image features on the update of noisy image features. Its calculation process is as follows:

[0049]

[0050] Where a(t) is the amplitude adjustment function, which controls the main variation of λ(t), and is a function that gradually decreases as t decreases, as follows:

[0051]

[0052] in, It is the initial amplitude. The rate at which the amplitude decreases is controlled. b(t) is a frequency adjustment function that determines the frequency of change of λ(t). It is set as a function that changes in stages to simulate the changing noise control requirements at different stages. The formula for b(t) is:

[0053]

[0054] in, The phase shift determines the initial change pattern of λ(t), which can be set according to the requirements during model initialization. Finally, the updated noise image features are... and original image features The system is fused together, and the predicted angular position of the laser servo device is output through the fully connected layer. After calibration using this solution, the entire process takes 15 minutes, with an average error of approximately 0.1 degrees, achieving the minimum accuracy currently available for servos. Traditional calibration methods take approximately 2 hours, with an accuracy of only 0.63 degrees. Therefore, this solution effectively reduces calibration time while significantly improving device accuracy, and allows for arbitrary placement of the device within a specified range, reducing installation costs.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep convolutional regression laser calibration algorithm, characterized in that, Includes the following steps: S1. Data Acquisition: First, fix the positions of the screen, servo-laser emitter, and camera. The computer initiates commands to control the servo's rotation and laser illumination. The servo rotates 1 degree vertically (h-axis) and horizontally (v-axis) each time. Therefore, if the rotation angle from the leftmost to the rightmost edge of the screen is 60 degrees when h is fixed, and similarly, the rotation angle from top to bottom is 30 degrees when v is fixed, the servo will sample a total of 30 × 60 = 1800 times. Each time the servo switches angles, the camera captures an image, naming the image with the hv value at the time of acquisition, and recording the angle position of the laser servo device that generated the image. The data set of angle positions is obtained by summarizing them. This yielded 1800 original images. S2. Complete the missing laser impact points to reconstruct the original acquired image. ,in and These represent the original image and the noisy image, respectively. The noisy image is filled with missing laser points using an approximation method. S3. Noise enhancement: The laser impact points in the original image are labeled, and a target detection base model, named model_base, is trained based on the YOLOv8 model. This model can effectively identify the pixel position of the laser impact points in the image. This results in a collection of all original image pixel location data. ), retrieve data pairs ( ),( Based on the data set, a multinomial regression model is trained and named Reg_location, and the following applies: ， Among them, when inputting a pixel position on the screen At that time, the polynomial model Reg_location can calculate the angular position of the laser servo device that generated the image. A certain number of (x,y) coordinate pairs are randomly generated within the range of (x,y) and input into the Reg_location model to obtain the base (h,v) values. New (x,y) coordinate values ​​are then generated based on the noise enhancement function. The calculation of the noise enhancement function for x is shown below: ， in, The main control is the basic intensity of noise. , The combined effect acts on the frequency and phase of noise, influencing the morphology of the noise pattern. Used to fine-tune noise intensity, providing more precise control. To adjust the range and complexity of the noise, a term combining sine and cosine was added, making the noise patterns more diverse. S4. Multi-scale deep convolutional regression: A weight mask W is constructed using a standard normal distribution, with the same size as the original image. The weight mask W is used to adjust the distribution of noise. The calculation method is as follows: ， in, This is the image at step t. It is noise sampled from a standard normal distribution. To control the overall noise intensity, a multi-scale deep convolutional regression model is constructed. The model input is an image and the angular position of the laser servo device that generated the image. The training model is required to determine the angular position of the laser servo device at the moment of laser impact in a given image. A dual-input network architecture is constructed to process two input streams simultaneously, corresponding to the original image and the noisy image respectively. Finally, these two input streams are fused at a certain point to predict the angular position of the laser servo device. The multi-scale deep convolutional regression model introduces a feature interaction layer between two parallel feature extraction sub-networks to dynamically adjust the feature extraction process for the noisy image. and noisy images Features are extracted through their respective CNN subnetworks, denoted as . and noisy images Then use As a query Q, As keys K and values ​​V, weighted noisy image features are calculated using an attention mechanism. Then, the obtained weighted noise image features are used. The feature representation of the noisy image is updated to guide the learning of the noisy image feature extraction subnetwork. The calculation process is as follows: ， in, It is a function that decreases with t, used to control the degree of influence of the original image features on the update of noisy image features. Its calculation process is as follows: ， Where a(t) is the amplitude adjustment function, which controls the main variation of λ(t), and is a function that gradually decreases as t decreases, as follows: ， in, It is the initial amplitude. The rate at which the amplitude decreases is controlled. b(t) is a frequency adjustment function that determines the frequency of change of λ(t). It is set as a function that changes in stages to simulate the changing noise control requirements at different stages. The formula for b(t) is: ， in, The phase shift determines the initial change pattern of λ(t), which is set according to the requirements during model initialization. Finally, the updated noise image features are... and original image features The fusion is performed, and the predicted angle position of the laser servo device is output through the fully connected layer.

2. The deep convolutional regression laser calibration algorithm according to claim 1, characterized in that, The multi-scale depth convolutional regression model in step S4 is divided into two parts: a dual-path localization module and a depth image regression module. The dual-path localization module takes the original image and the noise-enhanced image as inputs and uses dual-path cross-learning to improve the localization ability of the laser landing point. The depth image regression module obtains the angle position of the laser servo device required to generate the laser landing point.

Images