An intelligent detection method and system for body paint defects based on deflection imaging

Abstract

The present invention discloses an intelligent detection method and system for defects on a vehicle body paint surface based on deflection imaging. The method includes: Step S1: Calculation of the vehicle body pose; after the vehicle body to be detected enters the detection station, the vehicle body point cloud is acquired, and the vehicle body pose deviation is estimated by registering with the reference vehicle body point cloud or CAD model to determine the vehicle body pose; Step S2: Planning of the acquisition points; based on the vehicle body pose, the acquisition point trajectory of the robot adjusted by offline planning is acquired; Step S3: Acquisition of the vehicle body paint surface image; the paint surface image data is acquired in a multi-beat manner; Step S4: Detection of the paint surface defects; the acquired vehicle body paint surface image is processed based on the defect detection algorithm to obtain the size, type and three-dimensional position of the defects. The system is used to implement the above method. The present invention has the advantages of simple principle, high degree of intelligence and high detection accuracy.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of intelligent detection of vehicle body paint defects, and particularly refers to an intelligent detection method and system for vehicle body paint defects based on deflection imaging. Background Art

[0002] Automobile painting, as a high-precision and complex technological process, is very crucial in the production and manufacturing of vehicle bodies. The quality of painting directly affects the anti-rust and corrosion resistance as well as the appearance and aesthetics of automobiles, and thus affects the market sales of automobiles and the brand image of manufacturers. Automobile painting usually sprays paint on the vehicle body surface by manual or robotic means, and then goes through processes such as drying to achieve a smooth, shiny, and beautiful effect. During this process, due to factors such as process quality and transportation rubbing, the paint surface will inevitably be soiled or damaged, such as common pits, bumps, scratches, and oil stains. To ensure the quality of automobile painting, a process of detecting vehicle body paint defects is carried out before the automobile leaves the factory.

[0003] Currently, the detection of vehicle body paint defects is mainly completed manually. In a special lighting workshop, workers detect defects by visual observation with the eyes and touch with the hands, and record the defect information. Although the above manual method can meet the production requirements, it highly depends on workers' experience and working status, and has problems such as low work efficiency, low defect detection rate (research shows that it is only between 70% and 80%), and instability. To improve the production efficiency and quality of automobiles, the industry urgently needs a high-efficiency and high-precision automated detection method.

[0004] With the in-depth development of computer vision and robotic technologies, the combination of robots and vision for defect detection in automobile painting quality monitoring has gradually become a trend. For example, some practitioners have proposed a method of using machine vision technology for paint defect detection in the vehicle body painting process. This method directly obtains the vehicle body paint image through a camera and combines it with a defect detection algorithm to achieve automatic detection. However, due to the interference of the reflective characteristics of the vehicle body paint on imaging, it is difficult to display the defects in the paint image, and thus the detection rate is relatively low.

[0005] In recent years, some other practitioners have proposed a method of combining bright and dark field light sources with a 2D camera to detect defects. Although this method weakens the paint surface reflection interference to a certain extent and can display defects in the image, the detection rate is still not high, especially for defects with relatively small scales.

[0006] Therefore, some other practitioners have proposed a method of using structured light stripe projection to obtain the reflection map of the paint surface to display tiny defects. Although this method has a relatively high detection rate for defects, a large number of stripe maps need to be projected to achieve high-quality imaging, and this inefficient characteristic is difficult to meet the on-line detection requirements of the paint surface. Summary of the Invention

[0007] The technical problem to be solved by the present invention lies in: aiming at the technical problems existing in the prior art, the present invention provides an intelligent detection method and system for body paint defects based on deflection imaging, which has a simple principle, high intelligence and high detection accuracy.

[0008] To solve the above technical problems, the present invention adopts the following technical solutions:

[0009] An intelligent detection method for body paint defects based on deflection imaging, which includes:

[0010] Step S1: Calculation of the body pose; when the body to be detected enters the detection station, the body point cloud is obtained, and the body pose deviation is estimated by registering with the reference body point cloud or CAD model to determine the body pose.

[0011] Step S2: Planning of the acquisition points; based on the body pose, the acquisition point trajectory planned offline by the robot is adjusted and collected.

[0012] Step S3: Acquisition of the body paint image; the paint image data is obtained in a multi-beat manner.

[0013] Step S4: Detection of paint defects; based on the defect detection algorithm, the collected body paint image is processed to obtain the size, type and three-dimensional position of the defects.

[0014] As a further improvement of the method of the present invention: in the step S3, the acquisition of the body paint image includes the original fringe pattern, the phase map and the point cloud map, and the paint image data is obtained by combining Fourier transform profilometry and deep learning.

[0015] As a further improvement of the method of the present invention: the process of the step S3 includes:

[0016] Step S301: Display and acquisition of the fringe pattern; after reaching the acquisition point, the phase-shifted fringe is displayed by the diffuse backlight source, and the reflection pattern of the body paint is synchronously acquired.

[0017] Step S302: Extraction of the body paint phase map; the phase map is extracted from the original fringe pattern by using Fourier transform profilometry; Fourier transform profilometry calculates the phase value of all points I j (x, y) of the original fringe pattern according to the following formula

[0018]

[0019] where Im[I j (x, y)], Re[I J (x, y)] are respectively all points I of the original fringe pattern jThe imaginary and real parts of the Fourier transform of (x, y); the grayscale value of the j-th point is Ij(x, y), where x and y are the horizontal and vertical coordinate positions of the point.

[0020] Step S303: Establishment of the body paint point cloud map; according to the calibration data, use the phase diagram to establish the point cloud map.

[0021] As a further improvement of the method of the present invention: in step S302, for the parts with complex body paint curvature, use Fourier transform profilometry combined with deep learning, input the original fringe pattern and the Fourier spectrum pattern into the trained deep learning model to obtain the result.

[0022] As a further improvement of the method of the present invention: when using the deep learning method to pre-train the model, the original fringe pattern and the Fourier spectrum pattern pass through the input layer to obtain features Figure 1 ; the features Figure 1 pass through the sampling layer to obtain features Figure 2 ; the features Figure 2 pass through the fully connected layer to obtain the result.

[0023] As a further improvement of the method of the present invention: in step S4, use the defect detection method based on multi-modal fusion, and use multi-source data for stage detection according to the image features of different defects.

[0024] As a further improvement of the method of the present invention: the process of step S4 includes:

[0025] Step S401: Defect detection of the original fringe pattern; the original fringe pattern is directly collected; first, the original fringe pattern is processed by morphology to obtain the morphological feature map; by analyzing the geometric perimeter and area of the morphological features, determine the presence and type of defects;

[0026] Step S402: Defect detection in combination with the phase diagram; the phase diagram is obtained by Fourier transform profilometry; first, the phase diagram is processed by morphology to obtain the morphological feature map; by analyzing the geometric perimeter and area of the morphological features, determine the presence and type of defects;

[0027] Step S403: Defect detection in combination with the point cloud map; the point cloud map is obtained by combining the phase diagram with the calibration data; first, the point cloud map is processed by morphology to obtain the morphological feature map; by analyzing the morphological features, determine the presence and type of defects.

[0028] As a further improvement of the method of the present invention: in step S4, when using the method of combining morphology processing with deep learning to pre-train the model, the data passes through the input layer to obtain features Figure 1 ; the features Figure 1 pass through the sampling layer to obtain features Figure 2 ; the features Figure 2The result is obtained through the fully connected layer.

[0029] The present invention further provides an intelligent detection system for body paint surface defects based on deflection imaging, which includes:

[0030] A paint surface imaging acquisition unit for acquiring body paint surface images, including an optical imaging component, an acquisition robotic arm, and an acquisition guiding component; the optical imaging component includes more than two cameras and more than one diffuse reflection light source; one set of optical imaging components is installed at the end of each acquisition robotic arm; the acquisition guiding component 1 is installed on the body production line for obtaining the body pose and guiding the positioning of the acquisition robotic arm.

[0031] A defect detection control unit for acquisition control and defect detection, including an imaging component controller, a robotic arm controller, and an image processing component; the imaging component controller is connected to the optical imaging component, the robotic arm controller is connected to the acquisition robotic arm, and the image processing component is connected to the acquisition guiding component for calculating the body pose and defect detection.

[0032] After the body is transferred to the detection station, the acquisition guiding component acquires the body point cloud, and the image processing component calculates the body pose; according to the body pose, the acquisition robotic arm guides the optical imaging component to the specified position and stepwise acquires the body paint surface images within the specified range according to the offline planned trajectory.

[0033] As a further improvement of the system of the present invention: the acquisition guiding component is installed above the detection station; after the body is transferred to the detection station, the acquisition guiding component is triggered to acquire the body point cloud, and the image processing component calculates the body pose.

[0034] Compared with the prior art, the advantages of the present invention are as follows:

[0035] The intelligent detection method and system for body paint surface defects based on deflection imaging of the present invention have a simple principle, high intelligence, and high detection accuracy. By adopting a deflection imaging method combining Fourier transform profilometry and deep learning, rapid imaging of smooth and reflective surfaces such as body paint surfaces can be achieved using a single fringe pattern, improving the acquisition speed and meeting the requirements of on-line detection of body paint surfaces. Further, the present invention utilizes a multi-modal fusion defect detection method, which uses the original fringe pattern, phase map, and point cloud map, and combines morphological methods and deep learning methods to rapidly and accurately detect defects on smooth and reflective surfaces such as body paint surfaces in stages. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of the method of the present invention.

[0037] Figure 2 is a schematic structural principle diagram of the system of the present invention.

[0038] Figure 3 They are the original fringe pattern, fringe frequency spectrum diagram, and transformed wrapped phase diagram in the implementation of Fourier transform profilometry in a specific application example of the present invention.

[0039] Figure 4 They are the original fringe defect diagram, phase defect diagram, and point cloud defect diagram in the implementation of the multi-mode detection method in a specific application example of the present invention, as well as the defect identification diagram detected by using the working method of the present invention.

[0040] Legend Explanation:

[0041] 1. Acquisition guiding component; 2. Optical imaging component; 3. Acquisition robotic arm; 4. Image processing component; 5. Body to be detected. Detailed Implementation Manner

[0042] The following will further elaborate on the present invention in detail with reference to the specification drawings and specific embodiments.

[0043] In the description of the present application, it should be understood that the orientation or positional relationship indicated by terms such as "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present application.

[0044] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present application, "a plurality" means two or more, unless otherwise specifically defined.

[0045] In the present application, unless otherwise clearly specified and defined, terms such as "assembly", "connected", "connected to", "fixed" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal communication of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to specific circumstances.

[0046] The intelligent detection method for vehicle body paint defects based on deflection imaging in the present invention is an efficient deflection imaging method based on fringe projection. The method in the present invention captures one fringe pattern, and through the combination of Fourier transform profilometry and deep learning, quickly acquires 2D and 3D data of smooth and reflective surfaces. Then, based on the acquired paint surface data, a defect detection method of multi-modal fusion is adopted, that is, by utilizing the characteristic advantages of the fringe pattern, reflection image, and 3D point cloud, combining traditional machine learning methods and deep learning to achieve precise detection of paint surface defects. As Figure 1 shown, the process of the method in the present invention may include:

[0047] Step S1: Calculation of vehicle body pose;

[0048] After the vehicle body 5 to be detected enters the detection station, the acquisition guiding component 1 acquires the vehicle body point cloud, and the image processing component 4 estimates the vehicle body pose deviation through registration with the reference vehicle body point cloud or CAD model to determine the vehicle body pose;

[0049] Step S2: Planning of acquisition points;

[0050] Based on the vehicle body pose, the acquisition robot adjusts the acquisition point trajectory planned offline;

[0051] Step S3: Acquisition of vehicle body paint surface images;

[0052] The acquisition robot guides the optical imaging component 2 to acquire paint surface image data in a multi-beat manner;

[0053] Step S4: Detection of paint surface defects;

[0054] The image processing component 4 processes the acquired vehicle body paint surface images based on the defect detection algorithm to obtain the size, type, and three-dimensional position of the defects.

[0055] In a specific application example, in the step S2, the image processing component 4 is installed on the acquisition robotic arm 3, and different acquisition robotic arms 3 can be responsible for different vehicle body areas, and the paint surface point acquisition of the entire vehicle body is completed in a multi-machine synchronous manner.

[0056] In a specific application example, in the step S3, the acquisition of vehicle body paint surface images includes the original fringe pattern, phase map, and point cloud map. The method in the present invention combines Fourier transform profilometry and deep learning. Compared with the traditional multi-image phase-shifted fringe projection method, using 1 fringe pattern can acquire vehicle body paint surface data, which can significantly improve the acquisition speed.

[0057] Specifically, in the traditional phase-shifted fringe projection method, the phase is encoded into the gray value of the fringe pattern, and the gray value I j (x,y) of the j-th point of the fringe pattern can be expressed as:

[0058]

[0059] For the j-th point, a, b, and the phase value in the formula are all unknowns. According to the condition for solving the system of equations by combining them, obtaining each phase diagram requires consuming more than 3 phase-shifted fringe patterns.

[0060] However, the Fourier transform profilometry in the embodiment of the present invention uses frequency domain filtering to replace equation solving. After transforming a single fringe pattern into the frequency domain, the frequency components of the phase are directly extracted, and then the phase value is solved by inverse Fourier transform, which can achieve obtaining the body paint data only by using one fringe pattern, and can greatly improve the acquisition speed of the body paint data.

[0061] As a preferred embodiment, the process of step S3 of the present invention may include:

[0062] Step S301: Display and acquisition of the fringe pattern;

[0063] After the acquisition robot guides the optical imaging component 2 to the acquisition point, the phase-shifted fringe is displayed through the diffuse backlight source, and the image processing component 4 (such as a camera) synchronously acquires the reflection pattern of the body paint. Figure 3 a is the acquired original fringe pattern.

[0064] Step S302: Extraction of the body paint phase diagram;

[0065] The image processing component 4 extracts the phase diagram from the original fringe pattern by using Fourier transform profilometry. The main process is as follows: First, the frequency spectrum diagram of the original fringe pattern is obtained by using Fourier transform, as shown in Figure 3 b; Then, the fringe frequency components in the frequency spectrum diagram are extracted by using frequency domain filtering. The frequency spectrum diagram containing only the components undergoes inverse Fourier transform to obtain a complex result diagram. The phase value of the corresponding point I j (x, y) of the original fringe pattern is calculated according to the following formula

[0066]

[0067] where Im[I j (x, y)], Re[I J (x, y)] are the imaginary part value and the real part value of the corresponding point of point I j (x, y) in the complex result image respectively, and the calculated result is as shown in Figure 3 c.

[0068] Step S303: Establishment of the body paint point cloud diagram;

[0069] The image processing component 4 establishes a point cloud diagram by using the phase diagram according to the camera calibration data.

[0070] Further, in a specific application example, in step S302, for parts with complex curvature of the vehicle body paint surface, using Fourier transform profilometry combined with deep learning, the original fringe pattern and the Fourier spectrum image are input into the trained deep learning model to obtain the result.

[0071] When using the deep learning method to pre-train the model, the original fringe pattern and the Fourier spectrum image pass through the input layer to obtain features Figure 1 ; The features Figure 1 pass through the sampling layer to obtain features Figure 2 ; The features Figure 2 pass through the fully connected layer to obtain the result.

[0072] In a specific application example, in step S4, the defects on the vehicle body paint surface include pinholes, dirt spots, etc., and their feature manifestations on the original fringe pattern are uneven. Using a single image data for defect detection will have risks such as missed detection and false detection. The present invention innovatively uses a defect detection method based on multi-modal fusion, and uses multi-source data for staged detection according to the image features of different defects.

[0073] As a preferred embodiment, the process of step S4 of the present invention may include:

[0074] Step S401: Defect detection of the original fringe pattern;

[0075] The original fringe pattern is directly collected by the image processing component 4.

[0076] As Figure 4 shown in a, defects such as scratches are generally more obvious in the original fringe pattern. First, morphological processing processes such as edge detection and contour extraction of the original fringe pattern are performed to obtain the corresponding set of morphological features; by analyzing the geometric perimeter of the morphological features in the set, the presence and type of defects are judged, and the coordinate positions of the defects are framed with a rectangular box, as Figure 4 shown.

[0077] Step S402: Defect detection in combination with the phase diagram;

[0078] The phase diagram is mainly obtained by Fourier transform profilometry.

[0079] As Figure 4 shown in b, defects such as impurities are generally more obvious in the phase diagram. First, morphological processing processes such as binarization and connected component extraction of the phase diagram are performed to obtain the set of morphological features; by analyzing the area of the morphological features, the presence and type of defects are judged, and the coordinate positions of the defects are framed with a rectangular box, as Figure 4 shown.

[0080] Step S403: Defect detection in combination with the point cloud map;

[0081] The point cloud map is obtained by combining the phase map with the calibration data.

[0082] As Figure 4 shown in c, defects such as depressions are generally more obvious in the point cloud map. First, the point cloud map is processed through three-dimensional filling and other processes, and then the filled point cloud is subtracted from the original point cloud to obtain a set of morphological features; by analyzing the volume of the morphological features, the presence and type of defects are judged, and the coordinate positions of the defects are framed with a rectangular box, as Figure 4 shown.

[0083] Furthermore, in the above step S401, in the parts of the vehicle body paint surface with complex curvature, a method combining morphological processing and deep learning is used to input the original stripe map into the trained deep learning model to obtain a result.

[0084] Furthermore, when pre-training the model using the method of combining morphological processing and deep learning, the original stripe map passes through the input layer to obtain features Figure 1 ; the features Figure 1 pass through the sampling layer to obtain features Figure 2 ; the features Figure 2 pass through the fully connected layer to obtain a result.

[0085] Furthermore, in the above step S402, in the parts of the vehicle body paint surface with complex curvature, a method combining morphological processing and deep learning is used to input the phase map into the trained deep learning model to obtain a result.

[0086] Furthermore, when pre-training the model using the method of combining morphological processing and deep learning, the phase map passes through the input layer to obtain features Figure 1 ; the features Figure 1 pass through the sampling layer to obtain features Figure 2 ; the features Figure 2 pass through the fully connected layer to obtain a result.

[0087] Furthermore, in the above step S403, in the parts of the vehicle body paint surface with complex curvature, a method combining morphological processing and deep learning is used to input the point cloud map into the trained deep learning model to obtain a result.

[0088] Furthermore, when pre-training the model using the method of combining morphological processing and deep learning, the point cloud map passes through the input layer to obtain features Figure 1 ; the features Figure 1 pass through the sampling layer to obtain features Figure 2 ; the features Figure 2 pass through the fully connected layer to obtain a result.

[0089] As Figure 2 shown, the present invention further provides an intelligent detection system for vehicle body paint surface defects based on deflection imaging, which includes:

[0090] The paint surface imaging acquisition unit is used for acquiring images of the vehicle body paint surface, and includes an optical imaging component 2, an acquisition robotic arm 3, and an acquisition guiding component 1. The optical imaging component 2 includes more than two cameras and more than one diffuse backlight source, and the cameras are installed on the side of the light source. A set of optical imaging component 2 is installed at the end of each acquisition robotic arm 3. The acquisition guiding component 1 is installed on the vehicle body production line and is used to obtain the vehicle body pose and guide the positioning of the acquisition robotic arm.

[0091] The defect detection control unit is used for acquisition control and defect detection, and includes an imaging component controller, a robotic arm controller, and an image processing component 4. The imaging component controller is connected to the optical imaging component, the robotic arm controller is connected to the acquisition robotic arm, and the image processing component 4 is connected to the acquisition guiding component and is used for calculating the vehicle body pose and defect detection.

[0092] After the vehicle body 5 is transferred to the detection station, the acquisition guiding component 1 acquires the vehicle body point cloud, and the image processing component 4 calculates the vehicle body pose. According to the vehicle body pose, the acquisition robotic arm 3 guides the optical imaging component 2 to reach the specified position and step by step acquires the vehicle body paint surface images within the specified range according to the offline planned trajectory. Then, the image processing component 4 calculates the paint surface defect detection result.

[0093] In this embodiment, the acquisition guiding component 1 is installed above the detection station according to the actual environment of the vehicle body production line. After the vehicle body 5 is transferred to the detection station, the acquisition guiding component 2 is triggered to acquire the vehicle body point cloud, and the image processing component 4 calculates the vehicle body pose.

[0094] In Figure 1 there is one set of acquisition robotic arm and optical imaging component. It can be understood that according to the length and width of the vehicle body and the detection requirements, the number of acquisition robotic arms and optical imaging components can be one set or multiple sets (the dotted part in the figure).

[0095] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should be regarded as within the protection scope of the present invention.

Claims

1. An intelligent detection method for body paint defects based on deflection imaging, characterized in that, Including: Step S1: Calculation of vehicle body pose; after the vehicle body to be detected enters the detection station, obtain the vehicle body point cloud, estimate the vehicle body pose deviation through registration with the reference vehicle body point cloud or CAD model, and determine the vehicle body pose. Step S2: Planning of acquisition points; based on the vehicle body pose, collect the acquisition point trajectory planned offline by the robot for adjustment. Step S3: Acquisition of vehicle body paint surface image. Obtain the paint surface image data in a multi-beat manner; including: original stripe image, phase image, and point cloud image. Step S4: Detection of paint surface defects; process the collected vehicle body paint surface image based on the defect detection algorithm to obtain the size, type, and three-dimensional position of the defects; the defect detection algorithm is a defect detection method based on multi-modal fusion, including: S401: Defect detection of the original stripe image: obtain the corresponding set of morphological features from the original stripe image; judge the presence and type of defects by analyzing the geometric perimeter of the morphological features in this set, and frame the coordinate position where the defect is located with a rectangular box. S402: Defect detection of the phase image: obtain the corresponding set of morphological features from the phase image; judge the presence and type of defects by analyzing the area of the morphological features in this set, and frame the coordinate position where the defect is located with a rectangular box. S403: Defect detection of the point cloud image: obtain the corresponding set of morphological features from the point cloud image; judge the presence and type of defects by analyzing the volume of the morphological features in this set, and frame the coordinate position where the defect is located with a rectangular box.

2. The intelligent detection method for vehicle body paint defects based on deflection imaging according to claim 1, characterized in that, In step S3, the step of obtaining the original stripe image includes: step S301: after reaching the acquisition point, display the phase-shifted stripe through a diffused backlight source and synchronously collect the reflection pattern of the vehicle body paint surface.

3. The intelligent detection method for body paint defects based on deflection imaging according to claim 2, wherein, In step S3, the step of obtaining the phase image includes: step S302: extract the phase image from the original stripe image using Fourier transform profilometry.

4. The intelligent detection method for body paint defects based on deflection imaging according to claim 3, characterized in that In step S3, the step of obtaining the point cloud image includes: step S303: establish the point cloud image using the phase image according to the calibration data.

5. The intelligent detection method for body paint defects based on deflection imaging according to claim 2, wherein, In step S302, for the parts with complex curvature of the vehicle body paint surface, use Fourier transform profilometry combined with deep learning, input the original stripe image and the Fourier spectrum image into the trained deep learning model to obtain the phase image.

6. The intelligent detection method for body paint defects based on deflection imaging according to any one of claims 1-5, characterized in that, In step S401, input the original stripe image into the trained deep learning model, obtain feature map 1 through the input layer; feature map 1 passes through the sampling layer to obtain feature map 2, and feature map 2 passes through the fully connected layer to obtain the result.

7. The intelligent detection method for body paint defects based on deflection imaging according to claim 6, wherein In step S402, input the phase image into the trained deep learning model, obtain feature map 1 through the input layer; feature map 1 passes through the sampling layer to obtain feature map 2, and feature map 2 passes through the fully connected layer to obtain the result.

8. The intelligent detection method for body paint defects based on deflection imaging according to claim 7, wherein In step S403, input the point cloud image into the trained deep learning model, obtain feature map 1 through the input layer; feature map 1 passes through the sampling layer to obtain feature map 2, and feature map 2 passes through the fully connected layer to obtain the result.

9. An intelligent detection system for vehicle body paint defects based on deflection imaging, which is used to implement the intelligent detection method for vehicle body paint defects described in any one of claims 1-8, is characterized in that, Including: The paint surface imaging acquisition unit is used for acquiring images of the vehicle body paint surface, and includes an optical imaging component, an acquisition robotic arm, and an acquisition guiding component; the optical imaging component includes more than two cameras and more than one diffuse backlight source; one set of optical imaging components is installed at the end of each acquisition robotic arm; the acquisition guiding component (1) is installed on the vehicle body production line and is used to obtain the vehicle body pose and guide the positioning of the acquisition robotic arm. The defect detection control unit is used for acquisition control and defect detection, and includes an imaging component controller, a robotic arm controller, and an image processing component; the imaging component controller is connected to the optical imaging component, the robotic arm controller is connected to the acquisition robotic arm, and the image processing component is connected to the acquisition guiding component and is used for calculating the vehicle body pose and defect detection. After the vehicle body is conveyed to the detection station, the acquisition guiding component acquires the vehicle body point cloud, and the image processing component calculates the vehicle body pose; according to the vehicle body pose, the acquisition robotic arm guides the optical imaging component to reach the specified position and step by step acquires the vehicle body paint surface images within the specified range according to the offline planned trajectory.

10. The intelligent detection system for vehicle body paint surface defects based on deflection imaging according to claim 9, wherein the acquisition guiding component is installed above the detection station; after the vehicle body is conveyed to the detection station, the acquisition guiding component is triggered to acquire the vehicle body point cloud, and the image processing component calculates the vehicle body pose.

Images