High-precision Spot Positioning Method for Quadrant Detector Based on Meta-Learning

Abstract

The present invention proposes a high-precision spot positioning method for a quadrant detector based on meta-learning, including: 1) generating sufficient simulated sample data according to the measurement principle of the quadrant detector; 2) constructing a spot radius estimation network, training it with the generated simulated data, then inputting the real small sample data into the trained spot radius estimation network for prediction, and using the obtained estimated radius value to generate simulated sample data closer to the real spot distribution; 3) constructing a spot position prediction network, introducing the strategy of meta-learning to pre-train it with the simulated samples closer to the real spot distribution, fixing some parameters of the network and then fine-tuning the spot position prediction network with the real small sample data, and realizing high-precision spot positioning through the fine-tuned neural network. This enables the method to obtain high-precision spot positioning ability even with only a small number of real samples.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-contact laser measurement and optoelectronic detection, and particularly relates to a high-precision spot positioning method for a four-quadrant detector based on meta-learning. Background Art

[0002] A four-quadrant detector (4QD) is an optoelectronic detection device formed by arranging four photodiodes with exactly the same performance in a Cartesian coordinate system. Due to its advantages such as small volume, high sensitivity, fast response speed, and wide dynamic range, it is often used in many fields such as free-space laser communication, laser guidance, laser collimation, biological science, aerospace, and industrial production. When a light beam is incident on the photosensitive surface of the four-quadrant detector, a light spot will be formed. According to the photovoltaic effect, the detector can convert the optical signal into four electrical signals for output. By processing these signals of different sizes through a positioning algorithm, the centroid position of the incident light spot on the photosensitive surface can be determined. The essence of the four-quadrant detector spot positioning algorithm is to calculate the centroid position of the current light spot based on the four output photocurrent signals. However, as the light spot deviates further from the center of the detector, the offset of the signals output by the four-quadrant detector shows a non-linear change with the true position of the light spot, and the resulting positioning error is also larger, leading to a decrease in its positioning accuracy. Therefore, how to quickly and accurately fit this non-linear relationship is the key to improving the positioning accuracy of the four-quadrant detector.

[0003] The existing four-quadrant detector spot positioning algorithms can be mainly divided into two categories: data-driven and model-driven. Traditional data-driven algorithms use linear interpolation or polynomials to fit the non-linear relationship between the offset of the detector output signal and the true position of the light spot. Such methods require a large amount of real sample data, and the spot positioning accuracy is limited by the order of the polynomial, and the signal processing ability of the backend hardware is required to be relatively high. In order to reduce the dependence on the amount of sample data, the model-driven method detects the light spot position by constructing a measurement model based on the four-quadrant detector. The Chinese patent document with the application number (CN201610301117.2) uses a Gaussian model to equivalently represent the light intensity distribution of the laser spot on the photosensitive surface, and proposes a spot positioning method for a four-quadrant detector based on Gaussian distribution. First, the relational expression of the light spot position coordinates is obtained according to the corresponding relationship between the light energy and the photocurrent, and then a standard normal distribution table is established, and the center coordinates of the light spot are calculated by the query method. This method improves the measurement accuracy while reducing the calculation amount. However, the above model-driven method is sometimes difficult to fit the complex light spot energy distribution in the actual scenario. In recent years, some literatures have also proposed methods to improve the neural network to improve the spot positioning accuracy of the four-quadrant detector, but the performance of this supervised learning method is poor in the case of small samples. Summary of the Invention

[0004] In view of the deficiencies of the prior art, the present invention proposes a high-precision spot positioning method for a quadrant detector based on meta-learning. In order to improve the positioning accuracy in the case of small samples, the present invention first introduces meta-learning into the spot positioning algorithm of the quadrant detector. The core idea of meta-learning is to enable the model to acquire the ability to adjust hyperparameters and quickly learn new tasks based on the acquired existing knowledge.

[0005] To achieve the above object, the basic idea of the present invention is as follows: First, a large number of simulated sample data are generated according to the measurement principle of the quadrant detector, thereby reducing the dependence of the algorithm on real samples; a spot radius estimation network is constructed and trained with the generated simulated data, and then the real small sample data is input into the trained radius estimation network for prediction, and the estimated radius value is used to generate simulated data closer to the real spot distribution; finally, a position prediction network is constructed, and a meta-learning strategy is introduced to pre-train with the simulated data closer to the real spot distribution. After fixing some parameters of the network, the real small sample data is used to fine-tune the position prediction network, and high-precision spot position prediction is achieved through the fine-tuned neural network.

[0006] The designed method steps include: 1) generating sufficient simulated sample data according to the measurement principle of the quadrant detector; 2) constructing a spot radius estimation network, training with the generated simulated data, and then inputting the real small sample data into the trained spot radius estimation network for prediction, and using the obtained estimated radius value to generate simulated sample data closer to the real spot distribution; 3) constructing a spot position prediction network, introducing a meta-learning strategy to pre-train with the simulated samples closer to the real spot distribution, fixing some parameters of the network, and then using the real small sample data to fine-tune the spot position prediction network, and achieving high-precision spot positioning through the fine-tuned neural network. This method first introduces meta-learning into the spot positioning algorithm of the quadrant detector, and uses the generated simulated sample data to improve the expression ability and non-linear fitting ability of the network, so that the method can also obtain high-precision spot positioning ability in the case of only a small number of real samples.

[0007] The technical solution adopted by the present invention to solve its technical problems is:

[0008] Step S1: Generate simulated sample data according to the measurement principle of the quadrant detector;

[0009] Step S2: Construct a spot radius estimation network, train with the generated simulated data, and then input the real small sample data into the trained spot radius estimation network for prediction, and use the obtained estimated radius value to generate simulated sample data closer to the real spot distribution;

[0010] Step S3: Construct a spot position prediction network, perform hierarchical training using meta-learning, and use the trained network to achieve accurate prediction of the spot position coordinates.

[0011] Further, the method for generating simulation sample data in step S1 is as follows: Set a spot radius R. First, define a simulation spot model, then construct a 4QD model according to the specification information of the selected quadrant detector, and then simulate the movement of the spot along the x-axis on the 4QD to generate simulation samples. According to the calculation, the following sample set G is obtained R :

[0012] G R = {(σ x (k) , X (k) )|k ∈ [1, n]} (1)

[0013] where σ x is the normalized signal offset, R represents the radius of the spot, n represents the number of simulation samples generated by the currently set radius, and X (k) and represent the spot position and signal offset of the k-th sample.

[0014] Further, the calculation formula for the normalized signal offset σ x is as follows:

[0015]

[0016] where I 1 , I 2 , I 3 , I 4 represent the current magnitudes generated in each quadrant when the laser irradiates the quadrant detector.

[0017] Further, the spot radius estimation network in step S2 uses a BP neural network as the basic framework, with a total of t hidden layers, and each hidden layer has c neurons. Use the simulation sample data G R generated in step S1 to train the spot radius estimation network, and add noise perturbation during training to prevent overfitting.

[0018] Further, in order to estimate the true spot radius more accurately, use real small sample data to input into the trained spot radius estimation network for testing, and obtain the estimated spot radius Use the estimated value to repeat step S1 to generate simulation sample data closer to the true spot energy distribution to simulate the complex spot distribution in the actual scenario.

[0019] Further, the input layer of the spot position prediction network constructed in step S3 is the true σ x , and the output layer is the predicted spot position To improve the real-time performance of the spot positioning algorithm, the number of hidden layers of the spot position prediction network is set to l layers, where l < t.

[0020] Further, the specific method for training the spot position prediction network using the idea of meta-learning is as follows: First, use the simulation sample dataset G R and train the first few hidden layers of the network. Thanks to the spot radius estimation network in step S2, the simulation sample dataset is highly similar to the distribution of real sample data, and the sufficient simulation sample data improves the non-linear fitting ability of the network; then fix the parameters of the first few layers, and then use to train the subsequent hidden layers of the network, and then fine-tune the subsequent hidden layers with m real small sample data to obtain the trained spot position prediction network.

[0021] Compared with the prior art, the present invention and its preferred solutions have the following beneficial effects:

[0022] (1) The designed spot radius estimation network can obtain the estimated value of the spot radius according to real small sample data, and can be closer to the real spot energy distribution when generating simulation sample data, thereby reducing the dependence on real sample data.

[0023] (2) For the first time, meta-learning is introduced into the spot positioning algorithm of the quadrant detector, and the generated simulation sample dataset is used to improve the expression ability and non-linear fitting ability of the network, so that the method can also obtain high-precision spot positioning ability with only a small amount of real samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present invention will be further described in detail below with reference to the drawings and specific embodiments:

[0025] Figure 1 is the algorithm flow block diagram of the embodiment of the present invention.

[0026] Figure 2 is the schematic diagram of the working principle of the quadrant detector in the embodiment of the present invention.

[0027] Figure 3 is the schematic diagram of the principle for generating simulation sample data in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] To make the features and advantages of this patent more obvious and understandable, specific embodiments are given below for detailed description as follows:

[0029] It should be noted that the following detailed description is illustrative and aims to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs.

[0030] It should be noted that the terms used herein are for the purpose of describing specific embodiments only and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly dictates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0031] As Figure 1 shown, the high-precision spot positioning method for a quadrant detector based on meta-learning provided by the embodiments of the present invention is divided into the following steps.

[0032] Step 1, generating simulation sample data:

[0033] Step 1.1, generating sample data according to the working principle of laser positioning of the quadrant detector. As Figure 2 shown, taking a square detector as an example, the four quadrants are represented by A, B, C, and D respectively, and the non-light-sensitive area between each quadrant is called the dead zone.

[0034] Assume that the light spot approximately has a specific energy distribution and the energy is symmetric. When the laser irradiates on the quadrant detector, currents corresponding to the light intensity will be generated in the four quadrants, that is, the I Figure 2 in 1 , I 2 , I 3 , I 4 . To better characterize the relationship between the current and the spot position, the normalized signal offset is defined as follows:

[0035]

[0036]

[0037] where σ x and σ y are also called the resolution values of the x-axis and y-axis, and the value range is 0 ≤ |σ x | ≤ 1, 0 ≤ |σ y | ≤ 1. When the light spot irradiates on the center of the quadrant detector, the relationship between the currents obtained is I 1 = I 2 = I 3 = I 4 , that is, σ x = σ y= 0. When the light spot moves, the magnitude of the current also changes, and σ x > 0 indicates that the light spot is on the positive half-axis of the x-axis, and σ x < 0 indicates that the light spot is on the negative half-axis of the x-axis, and the same applies to the y-axis direction. It should be noted that σ x and σ y can only qualitatively represent the position of the light spot, and the actual coordinates of the centroid of the light spot still need to be calculated using the following algorithm. Since both the quadrant detector and the light spot have symmetry, this patent elaborates in detail on how to obtain the coordinates of the light spot in the x-axis direction, and the coordinates of the y-axis can be calculated in the same way.

[0038] Step 1.2, when generating sample simulation data, assume that the energy of the light spot conforms to a Gaussian distribution, and use h(x, y) to represent the energy density function of the light spot on the photosensitive surface of the quadrant detector:

[0039]

[0040] where P 0 represents the total energy of the incident light beam, R represents the radius of the light spot, (x, y) represents the position coordinates on the surface of the quadrant detector, and (X 0 , Y 0 ) represents the position coordinates of the center point of the light spot. The photocurrent generated in each quadrant can be expressed as:

[0041]

[0042] where η represents the photoelectric response coefficient, and S i represents the area of the i-th quadrant. According to the above formula, the estimated value of the X coordinate can be derived as follows:

[0043]

[0044] where erf -1 (x) represents the error function, represents the estimated value of the radius of the light spot, which is defined as follows:

[0045]

[0046] where m represents the number of real samples (in this embodiment, m = 4), X (j) and represent the true light spot position and signal offset of the j-th sample.

[0047] Step 1.3, set a light spot radius R, and first define a simulation light spot model according to formula (3), such as Figure 3As shown, a 4QD model is constructed according to the specifications such as the photosensitive surface and dead zone size of the selected quadrant detector. Then, the movement of the light spot is simulated along the x-axis on the 4QD to generate simulation samples. According to formulas (1) to (5), the following sample set can be calculated:

[0048]

[0049] where n represents the number of simulation samples generated by the currently set radius (in this embodiment, n = 500), and X (k) and represent the light spot position and signal offset of the k-th sample. Since the size of the photosensitive surface of the selected quadrant detector is a 2.4mm×2.4mm square, the light spot radius is set between 0.4mm and 0.8mm in the embodiment.

[0050] Step 2, construct a light spot radius estimation network:

[0051] Step 2.1, select a BP neural network as the framework of the light spot radius estimation network. The structure is as shown in the middle part, with a total of t hidden layers, and each hidden layer has c neurons. In this embodiment, t = 10 and c = 20 are set. Figure 1

[0052] Step 2.2, use the simulation sample data generated in Step 1 to train the light spot radius estimation network, and add noise perturbation during training to prevent overfitting.

[0053] Step 2.3, in order to estimate the real light spot radius more accurately, input the real small sample data into the trained light spot radius estimation network for testing to obtain the estimated light spot radius

[0054] Step 3, construct a light spot position prediction network:

[0055] Step 3.1, in order to simulate the complex light spot distribution in the actual scene, use the estimated value of the real sample light spot radius obtained in Step 2 Repeat Step 1.3 to generate simulation sample data closer to the real light spot energy distribution

[0056] Figure 1 Step 3.2, construct a light spot position prediction network as shown on the right. The input layer is the real σ Figure 1 x and the output layer is the predicted light spot position In order to improve the real-time performance of the light spot positioning algorithm, in this embodiment, the number of hidden layers of the light spot position prediction network is set to l = 3, and are represented by H 1 2 、H 3 、H 3 respectively.

[0057] Step 3.3, in order to prevent the network from overfitting to small samples, the idea of meta-learning is introduced to train the spot position prediction network. First, use the simulation sample dataset G R and to train H 1 and H 2 layers. Thanks to the spot radius estimation network in Step 2, the distribution of the simulation sample dataset is highly similar to that of the real sample data. Therefore, after fixing the parameters of the H 1 and H 2 layers, in this embodiment, use to train the H 3 layer, and then use m real small sample data to fine-tune the H 3 layer to obtain the trained spot position prediction network. This training method reduces the dependence of the algorithm on the amount of real sample data on the one hand, and on the other hand, through the generation of a highly similar simulation sample dataset, improves the expression ability and non-linear fitting ability of the network, so that the method can also obtain a high-precision position prediction ability with only a small amount of real samples.

[0058] Step 3.4, after training is completed, when the quadrant detector works, input σ x , and then the high-precision spot position coordinates can be obtained through the spot position prediction network.

[0059] The above is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still belong to the protection scope of the technical solution of the present invention.

[0060] This patent is not limited to the above best implementation manner. Anyone can obtain various other forms of high-precision spot positioning methods for quadrant detectors based on meta-learning under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention shall fall within the coverage of this patent.

Claims

1. A high-precision spot positioning method for a quadrant detector based on meta-learning, Characterized in that: It includes the following steps: Step S1: Generate simulation sample data according to the measurement principle of the quadrant detector; Step S2: Construct a spot radius estimation network, train it with the generated simulation data, then input the real small sample data into the trained spot radius estimation network for prediction, and use the obtained estimated radius value to generate simulation sample data closer to the real spot distribution; Step S3: Construct a spot position prediction network, perform hierarchical training using meta-learning and use the trained network to achieve accurate prediction of the spot position coordinates; The spot radius estimation network in step S2 uses a BP neural network as the basic framework, with a total of t hidden layers, each hidden layer having c neurons, and using the simulation sample data G generated in step S1 R Train the spot radius estimation network, and add noise perturbation during training to prevent overfitting; In step S3, the specific method of training the spot position prediction network using the idea of meta-learning is as follows: First, use the simulation sample dataset G R and to train the first few hidden layers of the network. Thanks to the spot radius estimation network in step S2, the distribution of the simulation sample dataset is highly similar to that of the real sample data, and the sufficient simulation sample data improves the non-linear fitting ability of the network; then fix the parameters of the first few layers, and then use to train the subsequent hidden layers of the network, and then fine-tune the subsequent hidden layers with m real small sample data to obtain the trained spot position prediction network.

2. The high-precision spot positioning method for a quadrant detector based on meta-learning according to claim 1, Characterized in that: The method for generating simulation sample data in step S1 is as follows: Set a spot radius R. First, define a simulation spot model, then construct a 4QD model according to the specification information of the selected quadrant detector, and then simulate the movement of the spot along the x-axis on the 4QD to generate simulation samples. The following sample set G is obtained according to the calculation R : where σ x is the normalized signal offset, R represents the radius of the light spot, n represents the number of simulation samples generated for the currently set radius, X (k) and represent the light spot position and signal offset of the k-th sample.

3. The high-precision spot positioning method for a quadrant detector based on meta-learning according to claim 2, Characterized in that: The normalized signal offset σ x has the following calculation formula: Among which I 1 、I 2 、I 3 、I 4 represent the magnitudes of the currents generated in each quadrant when the laser irradiates the quadrant detector.

4. The high-precision spot positioning method for a quadrant detector based on meta-learning according to claim 1, Characterized in that: Use real small sample data to input into the trained spot radius estimation network for testing to obtain the estimated spot radius Utilize the estimated value Repeat step S1 to generate simulation sample data closer to the real spot energy distribution To simulate the complex spot distribution in the actual scenario.

5. The high-precision spot positioning method for a quadrant detector based on meta-learning according to claim 1, Characterized in that: The input layer of the spot position prediction network constructed in step S3 is the true σ x , and the output layer is the predicted spot position To improve the real-time performance of the spot positioning algorithm, the number of hidden layers of the spot position prediction network is set to l layers, where l < t.

Images