Intelligent super-lens light field camera circuit board detection method based on deep reinforcement learning

Abstract

The invention relates to an intelligent super-lens light field camera circuit board detection method based on deep reinforcement learning. The method comprises the following steps: constructing a circuit board detection model of an intelligent super-lens light field camera; obtaining illumination parameters of the surface of the circuit board by using the real-time global illumination model; accurately positioning the circuit board by using a dense optical flow method to obtain a refocusing parameter of the light field image; a depth parameter of a light field image is obtained by using a light field layer-by-layer scanning technology and a photoelectric intelligent frequency multiplexing modulation technology; constructing a 5D light field model of the intelligent super-lens light field camera, and obtaining a 5D light field image of the circuit board; and constructing a space-time angle deep reinforcement learning model based on multiple agents, inputting the obtained illumination parameters, the refocusing parameters, the depth parameters and the 5D light field image into the space-time angle deep reinforcement learning model, and outputting collaborative decision control of multi-agent optimization. The intelligent super-lens light field camera opens up a new idea for accurate and intelligent detection of circuit board defects.

Description

Technical Field

[0001] The present invention relates to the field of light field cameras and PCBA board defect detection, and in particular to a circuit board detection method using an intelligent superlens light field camera based on deep reinforcement learning. Background Art

[0002] With the development of electronic technology and the electronics manufacturing industry, requirements for smaller components, higher mounting density, and more precise integration have posed greater challenges to PCBA board defect detection, and the difficulty of detection has continued to increase. Traditional manual inspection has many problems. To ensure the performance of electronic products, faster, more precise, and non-contact automatic optical defect inspection technology has become a powerful tool in the electronics industry and can better adapt to the requirements of automated production lines.

[0003] Automated optical inspection (AOI) based on machine vision technology, combined with AI and deep learning, can significantly improve inspection accuracy and efficiency. Consequently, an increasing number of electronics manufacturers are adopting AOI and deep learning software to accurately detect and identify detailed surface defects on PCBAs, such as material imperfections, solder joint defects, and missing or misaligned components. AOI equipment also faces major challenges, including light pollution, large data volumes, redundant information, high-dimensional spatial, angular, and depth features, and slow autofocus. Under inappropriate lighting, some gaps and discontinuities can be difficult to detect, leading to light pollution and misjudgment. Light source selection is crucial. Intelligent lighting methods, such as ring lights or backlights, can effectively reduce the impact of reflections and shadows. In mixed production lines, mixed visual inspection systems for circuit boards of varying sizes can suffer from slow autofocus. These challenges also lead to insufficient algorithmic capabilities and limited real-time performance when extracting limited defect information from massive amounts of data. Metalens light field cameras, with their extreme depth of field, can enable achromatic light field imaging with a wide depth of field for inspection of circuit boards of varying sizes. Summary of the Invention

[0004] The purpose of this invention is to provide a circuit board detection method using an intelligent superlens light field camera based on deep reinforcement learning, to realize a circuit board imaging system for mixed production and detection of circuit boards of different sizes, with unique autofocus and super depth of field capabilities, suitable for PCB / SMT detection.

[0005] The purpose of the present invention can be achieved by the following technical solutions: The intelligent super-lens light field camera circuit board detection method based on deep reinforcement learning includes the following steps: S1. Build a PCBA board detection model for the intelligent superlens light field camera; S2, using the light source action variables to intelligently control the combined light source agent, and using the real-time global illumination model to obtain the illumination parameters of the target PCBA board surface; S3, using the platform motion variables to intelligently control the movement speed and direction of the platform agent, using the dense optical flow method to accurately locate the target PCBA board, and obtaining the refocusing parameters of the light field image; S4. Using the motion variables of the electro-optical crystal meta-lens to intelligently control the position and refractive index of the electro-optical crystal meta-lens intelligent body, and using light field layer-by-layer scanning and optoelectronic intelligent frequency multiplexing modulation technology to obtain the depth parameters of the light field image; S5. Build a 5D light field model of the intelligent superlens light field camera to obtain a 5D light field image of the target PCBA board; S6. Construct a multi-agent-based spatiotemporal angular deep reinforcement learning model, input the obtained illumination parameters, refocusing parameters, depth parameters and 5D light field images into the spatiotemporal angular deep reinforcement learning model, and output the multi-agent optimized collaborative control actions.

[0006] The step S1 comprises the following steps: S11, constructing a PCBA board detection model based on the intelligent superlens light field camera, including a loading platform intelligent agent, a target PCBA board, a combined light source intelligent agent, a main lens, an electro-optical crystal metalens intelligent agent and an image sensor; S12, the loading platform intelligent body is used to place the target PCBA board; S13, constructing the combined light source intelligent body including the RGB ring light source and the coaxial light source; S14. Construct the electro-optic crystal meta-lens intelligent body, which is equipped with a meta-lens array and an electro-optic crystal fusion body, a high-precision driving micro-motor and an external electric field.

[0007] The step S2 comprises the following steps: S21. Using light source action variables to intelligently control the illumination angle and color of the RGB ring light source and the brightness of the coaxial light source in the combined light source intelligent agent; S22, using a real-time global illumination model to obtain illumination parameters of the target PCBA board surface, wherein establishing a real-time global illumination model is expressed as: ; in, is the ambient light component, is the diffuse component, is the specular highlight component; The formula for ambient light component is: ; in, is the ambient light intensity, is the ambient light coefficient; The formula for the diffuse reflection component is expressed as: ; in, is the intensity of the light source, is the diffuse reflectance, is the surface normal vector, is the light source direction vector; The formula for the specular highlight component is: ; in, ( ) is the material's specular reflection coefficient, is the reflected light vector, which indicates the direction of the light after it is reflected on the surface of the object. is the sight direction vector, indicating the direction of the main lens, is the specular index, which indicates the glossiness of the object surface; the lighting parameter is expressed as , using the real-time global illumination model to measure and calculate the color difference between the test multiple combined light and the reference light, and obtain the illumination parameters of the combined light source intelligent body , used to characterize the quality of the color rendering of the light source.

[0008] The step S3 comprises the following steps: S31, using the platform motion variables to intelligently control the movement speed and direction of the platform agent, to obtain a multi-view sub-image array of the target PCBA board with the Mark point gradually translated; S32. Use dense optical flow method to accurately locate, jointly shift and superimpose the Mark points of the multi-view sub-image array to obtain the refocusing parameters of the light field image.

[0009] The step S4 comprises the following steps: S41. Adjust the position of the electro-optic crystal meta-lens intelligent body through the electro-optic crystal meta-lens action variables to achieve layer-by-layer focusing on components of different heights on the target PCBA board, and calculate the depth of field in real time using the light field layer-by-layer scanning technology; the focal length of the electro-optic crystal meta-lens intelligent body Secondary imaging object distance and secondary imaging distance The positional relationship between them satisfies the Gaussian formula: ; Under the Galileo type condition, the secondary imaging distance range is greater than half the focal length and less than one focal length; S42. Adjust the refractive index of the electro-optic crystal through the action variables of the electro-optic crystal meta-lens, and use the optoelectronic intelligent frequency multiplexing modulation technology to obtain the depth parameters of components of different heights on the target PCBA board.

[0010] The step S5 comprises the following steps: S51. Construct a 5D light field model of the intelligent superlens light field camera to characterize the viewpoint image resolution and angular resolution. S52. Jointly control multiple intelligent agents in real time and adopt a segmentation method with maximum inter-class variance to obtain a 5D light field image of the target PCBA board.

[0011] The step S6 comprises the following steps: S61. Construct a multi-agent-based spatiotemporal deep reinforcement learning model and establish an optimization problem for multi-agent joint actions with optimal image quality evaluation; S62. Using the actor network, extract correlation features between illumination parameters, refocusing parameters, and depth parameters and the 5D light field image in the time domain, spatial domain, and angular domain, and output action variables of each agent. S63. Input the illumination parameters, refocusing parameters, depth parameters, 5D light field images, light source motion variables, carrier platform motion variables and electro-optical crystal metalens motion variables into the centrally trained critic network of the spatiotemporal angular deep reinforcement learning model. The centrally trained critic network outputs the global value function with the optimal SSIM and MSE for image quality evaluation. The policy gradient mechanism of centralized training and decentralized execution is used to realize the detection method of circuit boards by intelligent superlens light field cameras.

[0012] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses a combined light source intelligent body to reduce light pollution, and can simultaneously achieve different illumination angles, brightness and color combinations, ensuring that the light source illumination remains consistent in different environments, providing uniform illumination of the target PCBA board, and highlighting scratches, flaws, marks and other small features on the PCBA board surface under different frequency lights; (2) The present invention uses intelligent control of the electro-optical crystal meta-lens and the loading platform to track, detect, analyze, and identify and understand the defects of the moving target PCBA board, making full use of spatial characteristics, angular characteristics, and depth characteristics, expanding the depth of field imaging and improving the vertical clarity range of the field of view. When operating in a mixed production line, it solves the problem of slow automatic focusing speed of the mixed visual inspection system for circuit boards of different sizes, and solves the problem of large depth of field synthesis of components of different heights; (3) The present invention adopts a multi-agent-based spatiotemporal-angular deep reinforcement learning model to extract the correlation characteristics of illumination parameters, refocusing parameters, depth parameters and 5D light field images in the time domain, space domain and angle domain respectively, and outputs multi-agent optimized collaborative control actions to realize the detection method of circuit boards using an intelligent super-lens light field camera. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of the method of the present invention; Figure 2 Schematic diagram of the 5D light field model of the intelligent superlens light field camera of the present invention; Figure 3 This is a schematic diagram of the PCBA board detection model of the intelligent superlens light field camera of the present invention; Figure 4 Schematic diagram of the multi-agent-based spatiotemporal deep reinforcement learning model of the present invention. DETAILED DESCRIPTION

[0014] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0015] This embodiment provides a circuit board detection method for an intelligent super-lens light field camera based on deep reinforcement learning, such as Figure 1 shown.

[0016] S1. Build a PCBA board detection model for the intelligent superlens light field camera; S11, constructing a PCBA board detection model based on the intelligent superlens light field camera, including a loading platform intelligent agent, a target PCBA board, a combined light source intelligent agent, a main lens, an electro-optical crystal superlens intelligent agent and an image sensor, such as Figure 2 As shown; S12, the loading platform intelligent body is used to place the target PCBA board, the linear guide rail of the loading platform intelligent body adopts a linear guide rail, the motor adopts a high-precision stepper motor, the movement speed and distance can be adjusted, the conveyor belt adopts a steel wire transmission belt, and the communication method adopts a SMEMA signal interface; S13. Constructing the combined light source intelligent entity including an RGB ring light source and a coaxial light source; the RGB ring light source can include red light (600-720nm), green light (510-530nm), and blue light (430-480nm), and can be controlled to have different illumination angles and color combinations to highlight the light characteristics of different frequencies of the target PCBA board; the coaxial light source provides uniform illumination of the target PCBA board, and the brightness can be adjusted to ensure that the brightness of the light source remains consistent in different environments, highlighting scratches, blemishes, marks and other small features on the surface of the target PCBA board; S14. Construct the electro-optical crystal metalens intelligent body, fuse the metalens array with the electro-optical crystal, and equip it with a high-precision driving micromotor and an external electric field; the electro-optical crystal metalens intelligent body fuses the metalens array with the electro-optical crystal, so that the electro-optical crystal is immersed between the microcolumns of the metalens array, and the edge of the electro-optical crystal is coated with a transparent electrode, and the refractive index of the electro-optical crystal can be controlled by applying an electric field to the electrode; the metalens array is equipped with a high-precision driving micromotor, and the micromotor drives the metalens array to move back and forth to achieve focal length adjustment.

[0017] S2, using the light source action variables to intelligently control the combined light source agent, and using the real-time global illumination model to obtain the illumination parameters of the target PCBA board surface; S21, using light source action variables to intelligently control the RGB ring light source illumination angle, color and brightness of the coaxial light source in the combined light source intelligent body; the light source action variables are expressed as , consisting of two action components {u}_{1}=\left [ {{u}_{11},{u}_{12}} \right ] ,in Used to control the illumination angle, red, green and blue switching of the RGB ring light source, and highlight the light characteristics of different frequencies of the target PCBA board through different illumination angles and color combinations; Used to control the brightness of the coaxial light source to ensure uniform illumination under different environments, highlighting scratches, flaws, marks and other small features on the surface of the target PCBA board; S22, using a real-time global illumination model to obtain illumination parameters of the target PCBA board surface, wherein establishing a real-time global illumination model is expressed as: ; in, is the ambient light component, is the diffuse component, is the specular highlight component; The formula for ambient light component is: ; in, is the ambient light intensity, is the ambient light coefficient; The formula for the diffuse reflection component is expressed as: ; in, is the intensity of the light source, is the diffuse reflectance, is the surface normal vector, is the light source direction vector; The formula for the specular highlight component is: ; in, ( ) is the material's specular reflection coefficient, is the reflected light vector, which indicates the direction of the light after it is reflected on the surface of the object. is the sight direction vector, indicating the direction of the main lens, is the specular index, which indicates the glossiness of the object surface; the lighting parameter is expressed as , using a real-time global illumination model to measure and calculate the color difference between the light of the combined light source agent and the reference light, and to obtain the illumination parameters of the combined light source agent , which is used to characterize the quality of the color rendering of the combined light source intelligent entity.

[0018] S3, using the platform motion variables to intelligently control the movement speed and direction of the platform agent, using the dense optical flow method to accurately locate the target PCBA board, and obtaining the refocusing parameters of the light field image; S31, using the platform action variables to intelligently control the moving speed and direction of the platform agent, to obtain a multi-view sub-image array of the target PCBA board with Mark points gradually translated; the platform action variables are expressed as , transmitted through the SMEMA signal interface of the platform intelligent body, used to intelligently control the movement speed and direction of the platform intelligent body, gradually translate the target PCBA board within the effective illumination range of the combined light source intelligent body and the light transmission range of the main lens, and obtain a multi-view sub-image array of the target PCBA board with mark points; S32, using dense optical flow method to accurately locate, jointly displace and superimpose the Mark points of the multi-view sub-image array to obtain the refocusing parameters of the light field image; the refocusing parameters are expressed as , use the dense optical flow method to accurately locate the Mark points in the multi-view sub-image array, take multiple sub-images in a row or column of the multi-view sub-image array, and perform joint displacement and superposition according to a rule. Use straight line light to pass through the same point of the object in different superimposed sub-images in the vertical direction. When the straight line is vertical, it indicates that it is in focus. When the straight line is tilted, there is different degrees of out-of-focus blur. It is necessary to combine displacement and superposition to achieve refocusing of the light field image and record the refocusing parameters of each light field image. .

[0019] S4. Using the motion variables of the electro-optical crystal meta-lens to intelligently control the position and refractive index of the electro-optical crystal meta-lens intelligent body, and using the light field layer-by-layer scanning technology and optoelectronic intelligent frequency multiplexing modulation technology to obtain the depth parameters of the light field image; S41. Adjust the position of the electro-optical crystal meta-lens intelligent body through the electro-optical crystal meta-lens action variables to achieve layer-by-layer focusing on components at different heights of the target PCBA board, and use the light field layer-by-layer scanning technology to calculate the depth of field in real time; The action variables of the electro-optical crystal metalens are expressed as {u}_{3}=\left [ {{u}_{31},{u}_{32}} \right ] ,in The high-precision drive micro-motor used to control the meta-lens array dynamically adjusts the position of the meta-lens array. The meta-lens array can move back and forth behind the focal plane of the main lens to adjust the focal length, so as to achieve layer-by-layer focusing of components of different heights on the target PCBA board. The focal length of the electro-optical crystal meta-lens intelligent body Secondary imaging object distance and secondary imaging distance The positional relationship between them satisfies the Gaussian formula: ; Under the Galileo type condition, the secondary imaging distance range is greater than half the focal length and less than one focal length; S42. Adjust the refractive index of the electro-optic crystal through the action variables of the electro-optic crystal meta-lens, and use the optoelectronic intelligent frequency multiplexing modulation technology to obtain the depth parameters of components at different heights on the target PCBA board; Electro-optical crystal metalens motion variables The amount Through the transmission of the edge electrodes of the electro-optical crystal, the frequency multiplexing modulation technology is used to perform optoelectronic intelligent frequency multiplexing modulation on the light of different frequencies before reaching the image sensor. This makes the transmittance of each focal point on the metalens array different, and realizes the intelligent control of the refractive index, phase and polarization state of the electro-optical crystal. The refractive index of the electro-optical crystal is in a power relationship with the modulated electric field, and the refractive index difference is: ; in Refractive index difference, is the electro-optic Kerr constant, is the vacuum wavelength of the incident light, To apply the electric field intensity, the optoelectronic intelligent frequency multiplexing modulation technology based on the refractive index is used to focus the light of different frequencies of the electro-optical crystal and obtain the depth parameters of the components of different heights on the target PCBA board. .

[0020] S5. Build a 5D light field model of the intelligent superlens light field camera to obtain a 5D light field image of the target PCBA board; S51. Construct a 5D light field model of the intelligent superlens light field camera to characterize the viewpoint image resolution and angular resolution. Construct a 5D light field model of the intelligent superlens light field camera. The 5D light field model is expressed as ,in Characterizes the angular resolution, indicating that viewpoint images, Represents the viewpoint image resolution, which means the resolution of a single viewpoint image is , represents the time of combined displacement and superposition, such as Figure 3 As shown; S52, jointly controlling multiple intelligent agents in real time, and adopting a segmentation method with maximum inter-class variance to obtain a 5D light field image of the target PCBA board; The 5D light field image is represented as , the combined light source agent, the moving speed and direction of the carrier platform agent, the position and refractive index of the electro-optical crystal meta-lens agent and other multiple agents are jointly controlled in real time, and the Lighting parameters at the moment , refocusing parameters and depth parameters ; also obtained on the image sensor The multi-view images of the single color channel corresponding to the time, the refocused images of the single color channel and the depth map are used to obtain the 5D light field image of the target PCBA board using the segmentation method with the maximum inter-class variance. .

[0021] S6. Build a multi-agent-based spatiotemporal angular deep reinforcement learning model, input the obtained illumination parameters, refocusing parameters, depth parameters and 5D light field images into the spatiotemporal angular deep reinforcement learning model, and output the multi-agent optimized collaborative control action; S61. Construct a multi-agent-based spatiotemporal deep reinforcement learning model and establish an optimization problem for multi-agent joint actions with optimal image quality evaluation; exist At this moment, the illumination parameters, refocusing parameters, depth parameters and 5D light field image are jointly expressed as: ; The coordinated control actions of multiple agents are expressed as: ; The multi-agent strategy gradient (COMA) algorithm is used to build a multi-agent spatiotemporal deep reinforcement learning model, such as Figure 4 As shown in Figure 2, the spatiotemporal deep reinforcement learning model consists of three actor networks and one centralized critic network. The input of the model is , the output structural similarity index ‌SSIM and root mean square error MSE for image quality evaluation are optimal, and the optimization solution process is: initialization: Moment , agent strategies and a centralized global value function ; Output: The best action (in At this moment, the optimal coordinated control action of multiple agents ); In multi-agent reinforcement learning scenarios, to promote joint control of agents, the Counterfactual Multi-Agent Policy Gradient (COMA) algorithm is used. By introducing a counterfactual baseline, the algorithm simulates the impact of different actions chosen by the remaining agent on the structural similarity index (SSIM) when the actions of two of the three agents are fixed. This allows for a more accurate measurement of the contribution of the current action and evaluates the multi-agent policy gradient update using the root mean square error (MSE). The algorithm then repeatedly fixes the actions of the other two agents in turn and calculates the contribution of the remaining agent's current action. The structural similarity index (SSIM) is expressed as: ; in, and They are and Momentary light field image and The average brightness, and They are and Momentary light field image and The standard deviation of yes and Momentary light field image and The covariance of and Is a constant used to avoid the denominator being zero, usually and ,in and , is the dynamic range of pixel values; The root mean square error MSE expression is: ; in , The values ​​are respectively in the horizontal resolution Vertical resolution Values ​​within the range; S62. Using the actor network, extract correlation features between illumination parameters, refocusing parameters, and depth parameters and the 5D light field image in the time domain, spatial domain, and angular domain, and output action variables of each agent. In the spatial domain, a spatial separation convolution module is used to extract the spatial information correlation features of the illumination parameters and the 5D light field image, which are input into the actor1 network of the spatiotemporal angle deep reinforcement learning model to output the light source action variables of the combined light source agent. ; In the angle domain, the angle separation convolution module is used to extract the refocusing parameters and the angular information correlation features of the 5D light field image, which are input into the actor2 network of the spatiotemporal angle deep reinforcement learning model to output the platform action variables. ; In the time domain, the temporal pose estimation module is used to extract the temporal information correlation features of the depth parameters and the 5D light field image, which are input into the actor3 network of the spatiotemporal angle deep reinforcement learning model to output the electro-optical crystal meta-lens action variables. ; S63. Input illumination parameters, refocusing parameters, depth parameters, 5D light field images, light source motion variables, platform motion variables, and electro-optical crystal meta-lens motion variables into a centrally trained critic network of a spatiotemporal angular deep reinforcement learning model. The centrally trained critic network outputs a global value function with optimal image quality evaluation SSIM and MSE. Utilizing a policy gradient mechanism of centralized training and decentralized execution, a method for detecting circuit boards using an intelligent super-lens light field camera is implemented. Using the Counterfactual Multi-Agent Policy Gradient (COMA) method, Time-of-day illumination parameters, refocusing parameters, depth parameters, 5D light field images and time 、 time The input is fed into the centrally trained critic network of the spatiotemporal deep reinforcement learning model, and a counterfactual benchmark is introduced to evaluate the actions taken by the combined light source agent, the platform agent, and the electro-optical crystal metalens agent. The contribution of each agent to the overall task completion is calculated based on the counterfactual benchmark formula. The global value function of the agent's contribution at this moment and policy gradient ,Will The policy gradient is input to the actor1, actor2 and actor3 networks, and the optimized joint action is gradually output. , to obtain the target PCBA board defect detection by light field camera.

[0022] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A circuit board detection method for intelligent superlens light field camera based on deep reinforcement learning, It is characterized in that The following steps are involved: S1. Build a PCBA board detection model for the smart superlens light field camera; S2, using the light source action variables to intelligently control the combined light source agent, and using the real-time global illumination model to obtain the illumination parameters of the target PCBA board surface; S3, using the platform action variables to intelligently control the moving speed and direction of the platform agent, using the dense optical flow method to accurately locate the target PCBA board, and obtaining the refocusing parameters of the light field image; S4. Using the action variables of the electro-optic crystal meta-lens to intelligently control the position and refractive index of the electro-optic crystal meta-lens intelligent body, and using the light field layer-by-layer scanning technology and the optoelectronic intelligent frequency multiplexing modulation technology to obtain the depth parameters of the light field image; S5. Build a 5D light field model of the intelligent superlens light field camera to obtain a 5D light field image of the target PCBA board; S6. Construct a multi-agent based spatiotemporal angular deep reinforcement learning model, input the obtained illumination parameters, refocusing parameters, depth parameters and 5D light field images into the spatiotemporal angular deep reinforcement learning model, and output the multi-agent optimized collaborative control actions.

2. According to the circuit board detection method of the intelligent super lens light field camera based on deep reinforcement learning in claim 1, It is characterized in that The step S1 comprises the following steps: S11, constructing a PCBA board detection model of the intelligent superlens light field camera, including a loading platform agent, a target PCBA board, a combined light source agent, a main lens, an electro-optical crystal superlens agent and an image sensor; S12, the loading platform intelligent body is used to place the target PCBA board; S13, constructing the combined light source intelligent body including the RGB ring light source and the coaxial light source; S14. Construct an electro-optic crystal meta-lens intelligent body, which is equipped with a meta-lens array and an electro-optic crystal fusion body, a high-precision driving micro-motor and an external electric field.

3. According to the circuit board detection method of the intelligent super lens light field camera based on deep reinforcement learning in claim 1, It is characterized in that The step S2 comprises the following steps: S21, using the light source action variable to intelligently control the RGB ring light source illumination angle, color and brightness of the coaxial light source in the combined light source intelligent body; S22, using a real-time global illumination model to obtain illumination parameters of the target PCBA board surface, wherein establishing a real-time global illumination model is expressed as: ； in, is the ambient light component, is the diffuse component, is the specular highlight component; The ambient light component formula is expressed as: ； in, is the ambient light intensity, is the ambient light coefficient; The diffuse reflection component formula is expressed as: ； in, is the intensity of the light source, is the diffuse reflectance, is the surface normal vector, is the light source direction vector; The formula for the specular highlight component is: ； in, ( ) is the material's specular reflection coefficient, is the reflected light vector, which indicates the direction of the light after it is reflected from the surface of the object. is the sight direction vector, indicating the direction of the main lens, is the specular index, which indicates the glossiness of the object surface; the lighting parameter is expressed as , using a real-time global illumination model to measure and calculate the color difference between the light of the combined light source agent and the reference light, and obtain the illumination parameters of the combined light source agent , which is used to characterize the quality of color rendering of the combined light source intelligent entity.

4. According to the circuit board detection method of the intelligent super lens light field camera based on deep reinforcement learning in claim 1, It is characterized in that The step S3 comprises the following steps: S31, using the platform motion variables to intelligently control the moving speed and direction of the platform agent, using the dense optical flow method for real-time detection and tracking, and obtaining a multi-view sub-image array of the target PCBA board that is gradually translated; S32. Use dense optical flow method to accurately locate, jointly displace and superimpose the Mark points of the multi-view sub-image array to obtain the refocusing parameters of the light field image.

5. According to the circuit board detection method of the intelligent super lens light field camera based on deep reinforcement learning according to claim 1, It is characterized in that The step S4 comprises the following steps: S41. Adjust the position of the electro-optic crystal meta-lens intelligent body through the electro-optic crystal meta-lens action variables to achieve layer-by-layer focusing on components of different heights on the target PCBA board, and use the light field layer-by-layer scanning technology to calculate the depth of field in real time; The focal length of the electro-optical crystal meta-lens intelligent body The object distance from the secondary image and the secondary imaging distance The positional relationship between them satisfies the Gaussian formula: ； Under the Galileo type condition, the secondary imaging distance range is greater than half the focal length and less than one focal length; S42. The refractive index of the electro-optic crystal is adjusted by the action variable of the electro-optic crystal meta-lens, and the depth parameters of components of different heights on the target PCBA board are obtained by using the optoelectronic intelligent frequency multiplexing modulation technology.

6. The circuit board detection method of the intelligent superlens light field camera based on deep reinforcement learning according to claim 1, It is characterized in that The step S5 comprises the following steps: S51. Construct a 5D light field model of the smart superlens light field camera to characterize the viewpoint image resolution and angular resolution. S52, jointly controlling multiple intelligent agents in real time, and adopting a segmentation method with maximum inter-class variance to obtain a 5D light field image of the target PCBA board.

7. The circuit board detection method of the intelligent superlens light field camera based on deep reinforcement learning according to claim 1, It is characterized in that The step S6 comprises the following steps: S61. Construct a multi-agent-based spatiotemporal deep reinforcement learning model, and establish an optimization problem for multi-agent joint actions with optimal image quality evaluation; S62, using the actor network, extracting correlation features of illumination parameters, refocusing parameters, depth parameters and 5D light field images in the time domain, space domain and angle domain respectively, and outputting action variables of each intelligent agent; S63. Input illumination parameters, refocusing parameters, depth parameters, 5D light field images, light source motion variables, carrier platform motion variables and electro-optic crystal metalens motion variables into the centrally trained Critic network of the spatiotemporal angular deep reinforcement learning model. The centrally trained Critic network outputs the global value function with the best image quality evaluation SSIM and MSE. The strategy gradient mechanism of centralized training and decentralized execution is used to realize the detection method of circuit boards by intelligent superlens light field cameras.

Images