Auto-collimation angle measurement method based on multi-scale residual convolutional neural network
Patent Information
- Application Number
- CN202510586799.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-08
- Publication Date
- 2025-08-29
AI Technical Summary
[0004]针对自准直成像光斑定位算法无法准确获取光斑特征进行光斑中心定位而造成的自准直角度测量精度低的问题
[0033] 1. The present invention utilizes multi-scale learning to extract features of large, medium, and small scales of the light spot, and utilizes the residual structure to significantly increase the trainable parameters, thereby improving the nonlinear learning ability of positioning from features to the center of the light spot. This can better compensate for the angle measurement errors caused by the aberrations of the autocollimation optical system and the manufacturing and assembly errors of the optical-mechanical system, thereby improving the accuracy of the autocollimation angle measurement.
Smart Images
Figure CN120558129A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a self-collimation angle measurement method, in particular to a self-collimation angle measurement method based on a multi-scale residual convolutional neural network, and belongs to the technical field of self-collimation angle measurement. Background Art
[0002] Convolutional neural networks (CNNs) are a special type of artificial neural network that combines artificial neural networks with convolution operations. They feature sparse connections and weight sharing, significantly reducing the number of network parameters. CNNs have demonstrated excellent performance in recognizing various objects and handling distortions and deformations, and are therefore widely used in computer vision tasks. Residual networks (RNNs) are a type of deep convolutional neural network that address the vanishing and exploding gradient problems in deep neural networks by introducing residual connections. These connections allow the network to more easily propagate gradients and information during training, thereby improving training efficiency and performance. Residual networks are also widely used in object detection tasks to improve detection accuracy and speed. With the rise of deep learning, multi-scale object detection methods based on CNNs have gradually become mainstream. These methods automatically learn image features to extract more expressive features, resulting in better results for subsequent classification, detection, and other tasks.
[0003] Autocollimation angle measurement technology is a high-precision angle measurement method widely used in fields such as machinery manufacturing and aerospace. Traditional autocollimation angle measurement methods rely primarily on optical components and mechanical structures, such as autocollimators and optical indexing heads. However, these methods suffer from limited measurement accuracy and complex operation. Summary of the Invention
[0004] To address the problem of low autocollimation angle measurement accuracy caused by the inability of the autocollimation imaging spot positioning algorithm to accurately acquire spot features for spot center location, a high-precision autocollimation angle measurement method based on a multi-scale residual convolutional neural network is proposed. This method uses three convolution kernels of different sizes to extract multi-scale features of the spot, significantly improving the neural network's ability to extract the various scale features of the autocollimator sensor imaging spot, including spot shape and grayscale distribution. Furthermore, this method uses a residual module to deeply extract each single scale feature, increasing the number of trainable parameters without increasing the training difficulty, and improving the nonlinear learning ability of the feature-to-spot center location method. Compared with traditional algorithms, this method can better compensate for angle measurement errors caused by aberrations in the autocollimation optical system and manufacturing and assembly errors of the optomechanical system, thereby improving autocollimation angle measurement accuracy. Experimental data shows that this method achieves autocollimation angle measurement accuracy of up to 0.04". This method solves the problem of low autocollimation angle measurement accuracy caused by the inability of the autocollimation imaging spot positioning algorithm to accurately acquire spot features for spot center location, improves positioning accuracy, and thus improves angle measurement accuracy.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] The self-collimation angle measurement method based on the multi-scale residual convolutional neural network is implemented by the following steps:
[0007] S1: normalize the collected image;
[0008] S2: The normalized image is used to extract the spot scale features using three different sizes of convolution kernels;
[0009] S3: Deeply extract the three scale features obtained in S2;
[0010] S4: Merge the three scale features after depth extraction obtained in S3 into a feature vector;
[0011] S5: Convert the feature vector obtained in S4 into the two-dimensional coordinates of the center of the spot image;
[0012] S6: Convert the two-dimensional coordinates of the center of the light spot obtained in S5 to obtain the measured two-dimensional angle value and then output the two-dimensional angle measurement value corresponding to the light spot image.
[0013] Furthermore, the multi-scale residual convolutional neural network includes an image normalization module, a multi-scale feature extraction module and an angle output module connected in sequence.
[0014] The image normalization module includes a single normalization layer for normalizing the image;
[0015] The multi-scale feature extraction module includes three groups of residual modules connected end to end, a convolution feature extraction part before each group of residual modules, and an average pooling layer after each group of residual modules, which is used to extract the spot scale features of the normalized image and then perform depth extraction to output a feature vector;
[0016] The angle output module includes multiple fully connected layers, which are used to process the feature vector and output the two-dimensional coordinates of the center of the light spot, and then convert the measured two-dimensional angle value according to the basic principle of the autocollimator.
[0017] Furthermore, the three convolution feature extraction parts of the multi-scale feature extraction module include three groups of convolution layers, normalization layers and activation function layers connected in sequence, and the convolution kernels of the three groups of convolution layers are large, medium and small convolution kernels.
[0018] Furthermore, the residual module includes two convolutional layers, between which a normalization layer and an activation function layer are inserted in sequence. After the second convolutional layer, a normalization layer is also added. The multi-channel two-dimensional features generated by the normalization layer after the second convolutional layer are connected with the residual to perform point-to-point matrix addition, and the point-to-point matrix addition is followed by an activation function layer.
[0019] Furthermore, the residual connection is formed by processing the original multi-channel two-dimensional features that have not been processed by the convolution layer in the residual module in sequence through a layer of average pooling layer and a layer of convolution layer with a special 1×1 convolution kernel.
[0020] Furthermore, the activation function of the activation function layer is ReLU.
[0021] Furthermore, the training method based on the multi-scale residual convolutional neural network is implemented by the following steps:
[0022] S1: Acquire the autocollimation sensor image and angle true value;
[0023] S2: Convert the true angle value into a two-dimensional coordinate, and pair the two-dimensional coordinate with the autocollimation sensor image one by one to form training data;
[0024] S3: Train the neural network model using the training data described in S2;
[0025] S4: Freeze some parameters and train again until the loss function value converges.
[0026] Furthermore, the acquisition preparation in S1 of the training method based on the multi-scale residual convolutional neural network includes the following steps:
[0027] S1.1: A deflection mirror frame that can be angled in two dimensions and a double-sided reflector mounted on the frame are combined to form a rotating target.
[0028] S1.2: For the same rotating target, simultaneously position the autocollimator and the standard instrument directly opposite the two reflective surfaces of the double-sided reflector in the rotating target. Adjust the autocollimator and the standard instrument until the centroid of the light spot in the autocollimator image is at the center of the image and the two-dimensional angles shown by the standard instrument are all zero. Keep the autocollimator, the standard instrument, and the deflection mirror mount base in the rotating target stationary in the subsequent steps.
[0029] S1.3: Control the deflection mirror frame in the rotating target to a series of angles, and use the autocollimation angle meter and standard instruments to measure the two reflective surfaces of the double-sided reflector.
[0030] Furthermore, the training method in S3 of the training method based on the multi-scale residual convolutional neural network is to train the network parameters using a step-by-step decreasing learning rate until the loss function value converges.
[0031] Furthermore, all neural network parameters before the fully connected layer in the neural network are frozen, and the learning rate is again reduced step by step based on S3 of the training method based on the multi-scale residual convolutional neural network. The obtained training data is put into the network training until the loss function value converges again, and the trained neural network is obtained.
[0032] The beneficial effects of the present invention are:
[0033] 1. The present invention utilizes multi-scale learning to extract features of large, medium, and small scales of the light spot, and utilizes the residual structure to significantly increase the trainable parameters, thereby improving the nonlinear learning ability of positioning from features to the center of the light spot. This can better compensate for the angle measurement errors caused by the aberrations of the autocollimation optical system and the manufacturing and assembly errors of the optical-mechanical system, thereby improving the accuracy of the autocollimation angle measurement.
[0034] 2. The present invention adds an image normalization module consisting of a single normalization layer before extracting the features of the spot image. Through the normalization operation, the generalization ability of the model is improved, overfitting during training is avoided, the training difficulty is reduced, and the positioning accuracy of the algorithm is improved.
[0035] 3. Compared with the traditional autocollimation measurement method, the present invention is insensitive to the optical and mechanical design of the autocollimator and the selection of the sensor. It does not need to change with the change of the autocollimator design and can be adapted to a variety of autocollimator structures, which significantly reduces the difficulty of applying this method to autocollimators designed for different needs and different purposes.
[0036] 4. The method of the present invention is presented in the form of a software algorithm, which is not coupled with other modules of the autocollimator, and is easy to debug and replace.
[0037] 5. Compared with the traditional autocollimation measurement method, the present invention has low requirements for the object plane aperture and can achieve higher measurement accuracy without using a complex object plane aperture. It simplifies the structure of the autocollimator and can effectively reduce the cost of the instrument.
[0038] 6. The present invention has low requirements for image processing computing power and can be run on embedded boards with low computing power and low power consumption. This expands the application scenarios of the present method and can be used in miniaturized devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 It is a schematic structural diagram of an embodiment of a neural network of the present invention;
[0040] Figure 2 1 is a schematic structural diagram of an embodiment of the residual module of the present invention;
[0041] Figure 3 It is a flow chart of the neural network training method of the present invention;
[0042] Figure 4 It is a schematic diagram of the use flow of the neural network of the present invention.
[0043] In the figure: 1. Image normalization module; 2. Multi-scale feature extraction module; 3. Extraction of small-scale features; 4. Extraction of medium-scale features; 5. Extraction of large-scale features; 6. Residual module; 7. Angle output module. DETAILED DESCRIPTION
[0044] Specific implementation method 1: Combination Figure 1-4 This embodiment describes the self-collimation angle measurement method based on a multi-scale residual convolutional neural network described in this embodiment. The self-collimation angle measurement method based on a multi-scale residual convolutional neural network is implemented by the following steps:
[0045] S1: normalize the collected image;
[0046] S2: The normalized image is used to extract the spot scale features using three different sizes of convolution kernels;
[0047] S3: The three scale features obtained in S2 are deeply extracted through the residual structure;
[0048] S4: Merge the three scale features after depth extraction obtained in S3 into a feature vector;
[0049] S5: Convert the feature vector obtained in S4 into the two-dimensional coordinates of the center of the spot image;
[0050] S6: Convert the two-dimensional coordinates of the center of the light spot obtained in S5 to obtain the measured two-dimensional angle value and then output the two-dimensional angle measurement value corresponding to the light spot image.
[0051] This method utilizes multi-scale learning to extract features of the large, medium, and small scales of the light spot, and utilizes the residual structure to significantly increase the trainable parameters, thereby improving the nonlinear learning ability of positioning from features to the center of the light spot. Compared with traditional algorithms, it can better compensate for the angle measurement errors caused by the aberrations of the autocollimation optical system and the manufacturing and assembly errors of the optical-mechanical system, thereby improving the accuracy of autocollimation angle measurement. Compared with traditional autocollimation measurement methods, this method has low requirements for the object plane aperture, and can achieve higher measurement accuracy without the use of a complex object plane aperture. It simplifies the structure of the autocollimator and can effectively reduce the cost of the instrument. Compared with traditional autocollimation measurement methods, the present invention significantly improves the accuracy of autocollimation angle measurement without increasing the difficulty of application, while reducing the cost of the instrument.
[0052] like Figure 1 As shown in the figure, after the image is input into the algorithm, it passes through the image normalization module, the multi-scale feature extraction module and the angle output module in sequence, and finally outputs the two-dimensional angle measurement value corresponding to the spot image.
[0053] The multi-scale residual convolutional neural network includes an image normalization module, a multi-scale feature extraction module and an angle output module which are connected in sequence.
[0054] The image normalization module includes a single normalization layer for normalizing the image. This module, composed of a single normalization layer, is added before feature extraction of the spot image. This normalization operation improves the model's generalization capabilities, avoids overfitting during training, reduces training difficulty, and improves angle measurement accuracy.
[0055] The multi-scale feature extraction module includes three groups of residual modules connected end to end, a convolution feature extraction part before each group of residual modules, and an average pooling layer after each group of residual modules. It is used to extract the spot scale features of the normalized image and then perform depth extraction to output a feature vector.
[0056] The three convolution feature extraction parts of the multi-scale feature extraction module include three groups of sequentially connected convolution layers, normalization layers, and activation function layers. The convolution kernels of the three groups of convolution layers are large, medium, and small convolution kernels. The three convolution feature extraction parts use large, medium, and small convolution kernels to extract image features of large, medium, and small scales in the image, respectively, to obtain multi-channel two-dimensional features. For each single-scale feature, a normalization layer and an activation function layer are added after the convolution layer, and the ReLU function is selected as the activation function. Then, each single-scale feature is sent to its corresponding multiple continuous residual modules for deep feature extraction.
[0057] Three convolution kernels of different sizes are used to extract the multi-scale features of the light spot, which greatly improves the neural network's ability to extract the various scale features of the autocollimator sensor imaging light spot (including spot shape, spot grayscale distribution, etc.).
[0058] like Figure 2 As shown, the residual module includes two convolutional layers, and a normalization layer and an activation function layer are sequentially inserted between the two convolutional layers. After the second convolutional layer, a normalization layer is also added. The multi-channel two-dimensional features generated by the normalization layer after the second convolutional layer are subjected to point-to-point matrix addition with the residual connection, and the point-to-point matrix addition is followed by an activation function layer. After the multi-channel two-dimensional features enter the residual module, they must pass through two convolutional layers. Between the two convolutional layers, a normalization layer and an activation function layer are sequentially inserted, and the ReLU function is selected as the activation function. After the second convolutional layer, a normalization layer is also added. The multi-channel two-dimensional features generated by this layer must be subjected to point-to-point matrix addition with the residual connection. The residual connection is formed by the original multi-channel two-dimensional features that have not been processed by the convolutional layer, which are sequentially processed by a layer of average pooling layer and a layer of convolutional layer with a special 1×1 convolution kernel. After the point-to-point matrix addition, it passes through an activation function layer with ReLU as the activation function, and the resulting multi-channel two-dimensional features are the results of the residual module.
[0059] Multiple residual modules are connected end to end to form a group of residual modules.
[0060] The residual module is used to deeply extract each single-scale feature, adding a large number of trainable parameters without increasing the difficulty of training, and improving the nonlinear learning ability of positioning from features to the center of the light spot.
[0061] After each single-scale feature passes through its corresponding residual module, the multi-channel two-dimensional feature formed must then pass through an average pooling layer, which averages the two-dimensional feature matrix of each channel into a single value. The features of the three scales are then merged into a feature vector, whose length is the sum of the number of channels of the features of the three scales.
[0062] The angle output module includes multiple fully connected layers, which are used to process the feature vector and output the two-dimensional coordinates of the center of the light spot. The deflection angle is then calculated based on the basic principles of the autocollimator, that is, the focal length of the collimator lens group and the light spot offset, and the measured two-dimensional angle value is converted. The feature vector output by the multi-scale feature extraction module is processed by multiple fully connected layers (Multilayer Perceptron, MLP) to obtain the two-dimensional coordinates of the center of the light spot, and then converted to the measured two-dimensional angle value based on the basic principles of the autocollimator.
[0063] like Figure 3 As shown, the training method based on the multi-scale residual convolutional neural network is implemented by the following steps:
[0064] S1: Acquire the autocollimation sensor image and the true angle value.
[0065] The collection preparation includes the following steps:
[0066] S1.1: A deflection mirror frame that can be angled in two dimensions and a double-sided reflector mounted on the frame are combined to form a rotating target.
[0067] S1.2: For the same rotating target, simultaneously position the autocollimator and the standard instrument directly opposite the two reflective surfaces of the double-sided reflector in the rotating target. Adjust the autocollimator and the standard instrument until the centroid of the light spot in the autocollimator image is at the center of the image and the two-dimensional angles shown by the standard instrument are all zero. Keep the autocollimator, the standard instrument, and the deflection mirror mount base in the rotating target stationary in the subsequent steps.
[0068] S1.3: Control the deflection mirror frame in the rotating target to a series of angles, and use the autocollimation angle meter and standard instruments to measure the two reflective surfaces of the double-sided reflector.
[0069] After the acquisition preparation, the sensor image of the autocollimation angle measuring instrument and the angle values of the pitch angle and yaw angle measured by the standard instrument are collected at the same time.
[0070] S2: Convert the true angle value into a two-dimensional coordinate, and pair the two-dimensional coordinate with the autocollimation sensor image one by one to form training data.
[0071] Combined with the known focal length of the autocollimator and according to the basic principle of the autocollimator, the angle value measured by the standard instrument is converted into a two-dimensional coordinate with the centroid of the aforementioned image light spot as the center and the side length of one pixel as the unit length, representing the position coordinates of the actual center represented by the light spot. This position is paired one by one with the aforementioned autocollimator angle measuring instrument sensor image collected at the same time to form a series of training data.
[0072] S3: Use the training data described in S2 to train the neural network model. Use the training data obtained in S2 to enter the aforementioned neural network model, and use a step-by-step decreasing learning rate to train the network parameters until the loss function value converges.
[0073] S4: Freeze some parameters and train again until the loss function converges. Freeze all neural network parameters before the fully connected layer in the neural network, and again reduce the learning rate in a stepwise manner based on S3. Put the resulting training data into the network training until the loss function, i.e., the mean squared error, converges again, resulting in a trained neural network.
[0074] like Figure 4 As shown, when specifically used, the method of use is achieved through the following steps:
[0075] S1: Acquire an autocollimation sensor image; acquire an autocollimation angle measuring instrument sensor image using the autocollimation angle measurement method based on a multi-scale residual convolutional neural network of the present invention.
[0076] S2: After cutting to a suitable size, feed it into the neural network; obtain the center of mass position of the light spot in the sensor image through the center of mass positioning method, cut a square of suitable size with the center of mass as the center, and feed the cropped image and the coordinates of the center of mass position of the light spot in the original image into the aforementioned neural network.
[0077] S3: Obtain the output coordinate value of the neural network, obtain the coordinates of the light spot position output by the neural network; obtain the angle value of the output of the neural network as the measurement value of the autocollimation angle measuring instrument for this measurement.
[0078] Compared to traditional autocollimation measurement methods, this method is insensitive to the autocollimator's optical and mechanical design and sensor selection. It does not need to change with changes in the autocollimator design and is adaptable to a variety of autocollimator structures, significantly reducing the difficulty of applying this method to autocollimators designed for different needs and applications. This method, implemented as a software algorithm, is independent of other autocollimator modules, facilitating debugging and replacement. Compared to traditional autocollimation measurement methods, this method has lower requirements for the object plane aperture, eliminating the need for a complex object plane aperture to achieve high measurement accuracy. This simplifies the autocollimator structure and effectively reduces instrument costs. This method also requires less image processing computing power and can be run on embedded boards with low computing power and low power consumption. This expands the method's application scenarios and allows its use in miniaturized devices. Furthermore, this method achieves autocollimation angle measurement accuracy of up to 0.04 inches. This method addresses the problem of low autocollimation angle measurement accuracy caused by the inability of the autocollimation imaging spot positioning algorithm to accurately capture spot characteristics and locate the spot center, improving positioning accuracy and, consequently, angle measurement accuracy.
[0079] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.
Claims
1. A self-collimation angle measurement method based on a multi-scale residual convolutional neural network, characterized by: The self-collimation angle measurement method based on the multi-scale residual convolutional neural network is achieved by the following steps: S1: normalize the collected image; S2: The normalized image is used to extract the spot scale features using three different sizes of convolution kernels; S3: The three scale features obtained in S2 are deeply extracted through the residual structure; S4: Merge the three scale features after depth extraction obtained in S3 into a feature vector; S5: Convert the feature vector obtained in S4 into the two-dimensional coordinates of the center of the spot image; S6: Convert the two-dimensional coordinates of the center of the light spot obtained in S5 to obtain the measured two-dimensional angle value and then output the two-dimensional angle measurement value corresponding to the light spot image.
2. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 1, wherein: The multi-scale residual convolutional neural network includes an image normalization module, a multi-scale feature extraction module and an angle output module connected in sequence. The image normalization module includes a single normalization layer for normalizing the image; The multi-scale feature extraction module includes three groups of residual modules connected end to end, a convolution feature extraction part before each group of residual modules, and an average pooling layer after each group of residual modules, which is used to extract the spot scale features of the normalized image and then perform depth extraction to output a feature vector; The angle output module includes multiple fully connected layers, which are used to process the feature vector and output the two-dimensional coordinates of the center of the light spot, and then convert the measured two-dimensional angle value according to the basic principle of the autocollimator.
3. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 2, wherein: The three convolution feature extraction parts of the multi-scale feature extraction module include three groups of convolution layers, normalization layers and activation function layers connected in sequence, and the convolution kernels of the three groups of convolution layers are large, medium and small convolution kernels.
4. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 3, wherein: The residual module includes two convolutional layers, between which a normalization layer and an activation function layer are inserted in sequence. After the second convolutional layer, a normalization layer is also added. The multi-channel two-dimensional features generated by the normalization layer after the second convolutional layer are connected with the residual to perform point-to-point matrix addition, and the point-to-point matrix addition is followed by an activation function layer.
5. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 4, characterized in that: The residual connection is formed by processing the original multi-channel two-dimensional features that have not been processed by the convolution layer in the residual module in sequence through a layer of average pooling layer and a layer of convolution layer with a special 1×1 convolution kernel.
6. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 5, characterized in that: The activation function of the activation function layer is ReLU.
7. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to any one of claims 1 to 6, characterized in that: The training method based on the multi-scale residual convolutional neural network is achieved by the following steps: S1: Acquire the autocollimation sensor image and angle true value; S2: Convert the true angle value into a two-dimensional coordinate, and pair the two-dimensional coordinate with the autocollimation sensor image one by one to form training data; S3: Train the neural network model using the training data described in S2; S4: Freeze some parameters and train again until the loss function value converges.
8. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 7, wherein: The acquisition preparation in S1 includes the following steps: S1.1: A deflection mirror frame that can be angled in two dimensions and a double-sided reflector mounted on the frame are combined to form a rotating target. S1.2: For the same rotating target, simultaneously position the autocollimator and the standard instrument directly opposite the two reflective surfaces of the double-sided reflector in the rotating target. Adjust the autocollimator and the standard instrument until the centroid of the light spot in the autocollimator image is at the center of the image and the two-dimensional angles shown by the standard instrument are all zero. Keep the autocollimator, the standard instrument, and the deflection mirror mount base in the rotating target stationary in the subsequent steps. S1.3: Control the deflection mirror frame in the rotating target to a series of angles, and use the autocollimation angle meter and standard instruments to measure the two reflective surfaces of the double-sided reflector.
9. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 7, wherein: The training method in S3 is to train the network parameters using a step-by-step decreasing learning rate until the loss function value converges.
10. The method for measuring autocollimation angle based on a multi-scale residual convolutional neural network according to claim 9, characterized in that: Freeze all neural network parameters before the fully connected layer in the neural network, and then reduce the learning rate in a step-by-step manner based on S3. Put the obtained training data into the network training until the loss function value converges again, and obtain the trained neural network.