CN112037178A - Cylinder two-dimensional image generation method based on multi-view camera - Google Patents

Abstract

The invention relates to a cylinder two-dimensional image generation method based on a multi-view camera, wherein a plurality of cameras are arranged around a cylinder to be detected; the multiple cameras respectively obtain images of the cylinder to be detected from different angles; synthesizing images of the cylinder to be detected, which are shot by a plurality of cameras at the same time, into a plane image; and detecting defects according to the synthesized plane image. The method has the advantages of full-process intelligent machine detection, no need of manual intervention, high detection coverage rate, low cost, high detection efficiency and the like.

Description

A method for generating two-dimensional images of cylinders based on multi-camera

technical field

The present invention relates to the technical field of image processing, in particular to a method for generating a two-dimensional image of a cylinder based on a multi-eye camera.

Background technique

In daily life, there are many daily necessities of cylinders, such as candles, wine bottles, beverage bottles and so on. In the production process of these products, it is often necessary to inspect the finished product to check whether the pattern on the surface of the product is damaged or whether there is a pattern printing error. Candles, wine bottles and other products have labels on the outside, and these products have stricter requirements on labels or patterns. Once the label or periphery is damaged or the pattern is printed incorrectly, it will affect the user's direct sense of the product, and thus question the quality of the product. Therefore, strict inspection is required before the product leaves the factory.

The conventional inspection method is manual inspection or random inspection, and some even leave the factory without inspecting the appearance. Visually inspect each product manually to see if it matches the design pattern or design sample. This method has technical problems such as low efficiency, easy to miss detection, and high labor intensity. Therefore, it is necessary to adopt an intelligent inspection system to solve these problems.

Computer vision is a simulation of biological vision using computers and related equipment, and the captured pictures or videos are processed by cameras to obtain two-dimensional and three-dimensional information of the corresponding scene. In modern smart factories, cameras are often used to capture images of cylindrical products, and the captured images are compared with standard images through image processing to check whether there are quality problems in the products.

The Chinese patent with the application number 201910859955.5 discloses a "machine vision-based cylindrical metal surface defect detection device", which includes a base, a feeding channel, a conveying device, and a feeding channel that are sequentially installed on the base and connected to each other. ; The conveying device includes a fixing frame. The input end of the transmission device is connected with the output shaft of the servo motor fixed on the base. A camera is also installed on the base on one side of the side monitoring station. This patent only discloses the structure of the inspection device, and does not mention any algorithms and methods for detecting product defects through image processing.

None of the above solutions in the prior art can solve the problem of intelligently inspecting the shape, surface defects, packaging defects, etc. of cylindrical products through image processing.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for generating a two-dimensional image of a cylinder based on a multi-eye camera, which can realize the shape detection of the cylinder by collecting the image of the cylinder, processing the image, and comparing the images.

The technical scheme that the present invention solves the above-mentioned technical problems is as follows:

A method for generating a two-dimensional image of a cylinder based on a multi-eye camera, comprising the following steps:

Step 1: Set up multiple cameras around the cylinder to be inspected;

Step 2: the multiple cameras obtain images of the cylinder to be inspected from different angles;

Step 3: Combine the images of the cylinder to be inspected captured by multiple cameras at the same time into a plane image;

Step 4: Perform defect detection according to the synthesized planar image.

In the step 1, among the multiple cameras, the fields of view of two adjacent cameras have overlapping areas.

In the step 1, after the camera is set, the internal parameters of the camera and the external parameters of each camera home are obtained.

In the step 3, the process of synthesizing a plurality of images into one plane image includes the following steps:

Step 301: Process a plurality of images to obtain the spatial point cloud of the overlapping area of the images captured by adjacent cameras;

Step 302: Calculate the pose of the standard point cloud and the pose of the measurement point cloud through the transformation of the space coordinate system;

Step 303 : Flatten the image, and obtain a flat image of the cylinder to be inspected after being unfolded.

In the step 301, the acquisition process of the spatial point cloud of the overlapping area is as follows:

Described step 3011: firstly perform region segmentation on two images captured by two adjacent cameras;

Described step 3012: in the two images after region segmentation, confirm a plurality of feature points respectively;

Step 3013: Match the feature points of the two images to each other;

Described step 3014: Screen each pair of matching points after matching, and remove the wrong matching point pair;

Step 3015: Using the matching point pair that has been successfully matched, calculate the position coordinates of the corresponding spatial point of the pair of matching points on the cylinder to be inspected.

In the described step 302, the process of calculating the pose of the standard point cloud and the pose of the measurement point cloud includes the following steps:

Step 3021: Process the point cloud by principal component analysis, and calculate the measurement space coordinate system;

Step 3022: Convert the point in the measurement space to the point in the design space coordinate system. .

The process of flattening the image in step 303 includes the following steps

Step 3031: Calculate the size of the expanded image, and convert the size of the image to the pixel size;

Step 3032: Convert the pixel position of the expanded image into position coordinates in the design space;

Step 3033: Project the point corresponding to the position coordinate in the design space on the camera imaging plane.

In the step 4, the defect detection process includes the following steps:

Step 401: template matching is performed between the generated image and the standard image, and the corresponding relationship is found;

Step 402: Based on the corresponding relationship, perform block detection on the generated image and the standard image.

In the step 4, the defect detection process includes the following steps:

Step 401': collect a large number of image data of qualified products, and establish a training set;

Step 402': train the deep learning network through the training set;

Step 403': Use the trained network to perform defect detection on the synthesized planar image.

The beneficial effects of adopting the scheme of the present invention are:

The invention relates to a method for generating a two-dimensional image of a cylinder based on a multi-eye camera. A plurality of cameras are arranged around a cylinder to be inspected; the multiple cameras obtain images of the cylinder to be inspected from different angles respectively; The images of the cylinder to be inspected captured at the same moment are combined into a plane image; and defect detection is performed according to the combined plane image.

The control method of the present invention has the following technical features.

1. Through image vision algorithm and image comparison, it can quickly confirm whether there is a defect. If there is, it can accurately locate the defect position, and the detection efficiency is high.

2. The method adopts multi-camera detection, and there are overlapping areas between adjacent cameras, which can achieve full coverage detection, greatly improve the detection coverage rate, low cost, and detection efficiency much higher than that of manual work.

4. The control method of the present invention uses a computer vision scheme in the whole process without manual intervention.

5. The solution of the present invention has lower requirements on hardware, so the implementation cost is low, and there is no special requirement on camera characteristics.

The method for generating a two-dimensional image of a cylinder based on a multi-eye camera of the present invention has the advantages of intelligent machine detection in the whole process, no manual intervention, high detection coverage, low cost, and high detection efficiency.

Description of drawings

FIG. 1 is a schematic diagram of the camera arrangement of the present invention.

FIG. 2 is a flow chart of the camera calibration of the present invention.

FIG. 3 is a flow chart of the image synthesis of the present invention.

FIG. 4 is a flow chart of point cloud acquisition for image synthesis according to the present invention.

FIG. 5 is a flow chart of pose calculation for image synthesis according to the present invention.

FIG. 6 is a flow chart of the image flattening process of the image synthesis of the present invention.

FIG. 7 is a schematic diagram of the triangulation principle in the point cloud acquisition process of the present invention.

FIG. 8 is a schematic diagram of converting a cylinder into a plane in the image flattening process of the present invention.

Fig. 9 is the corresponding schematic diagram of the coordinate position of the cylinder and the coordinate position of the plane image in the image flattening process of the present invention.

Fig. 10 is a schematic diagram of the conversion between the spatial coordinates of the cylinder and the plane image coordinates in the image flattening process of the present invention.

In the accompanying drawings, the list of components represented by each number is as follows:

Detailed ways

The principles and features of the present invention are described below in conjunction with the accompanying drawings, and the examples are only used to explain the present invention, but not to limit the scope of the present invention.

As shown in Fig. 1-10, a method for generating a two-dimensional image of a cylinder based on a multi-eye camera of the present invention includes the following four steps: steps 1-4.

Step 1: Set up multiple cameras around the cylinder to be inspected;

Step 2: the multiple cameras obtain images of the cylinder to be inspected from different angles;

Step 3: Combine the images of the cylinder to be inspected captured by multiple cameras at the same time into a plane image;

Step 4: Perform defect detection according to the synthesized planar image.

In the step 1, among the multiple cameras, the fields of view of two adjacent cameras have overlapping areas.

In the present invention, a common fixed-focus camera can be used for the selection of the camera, and there is no need to select an industrial camera with high frame rate and high resolution. There needs to be some overlap between the fields of view of adjacent cameras, with an overlap ratio of 20%-40%. The optimal overlap ratio is 30%. In specific implementation, the higher the overlap rate, the better the detection effect, but the more computing resources are required. The lower the overlap rate, the higher the calculation speed, but the detection failure is prone to occur, that is, although the camera's field of view overlaps, it cannot guarantee that the same part of the side of the candle is captured at the same time. When the selected overlap rate is 30%, the detection success rate can be guaranteed, and the computational complexity can be kept relatively low. Figure 1 uses three cameras as an example, but it is not limited to three cameras, as long as the overlap ratio of the fields of view between the cameras is satisfied, and the entire body range is covered at the same time, several cameras can be arranged. When the present invention is finally applied to a production line, it needs to ensure that all cameras capture images at the same time, and are controlled by a control processor so that all cameras can be triggered synchronously and capture images at the same time.

In the step 1, after the camera is set, the internal parameters of the camera and the external parameters of each camera home are obtained.

After the camera is installed, the first thing to do is to debug the camera. When debugging, you need to get the internal parameters of each camera and the external parameters between cameras. Camera internal parameters include: camera focal length, optical center, radial distortion and tangential distortion parameters. The external parameters between cameras include: rotation and translation parameters between cameras.

After the camera is debugged, calibrate the camera. The camera calibration algorithm refers to Zhang Zhengyou's calibration algorithm, and the calibration process is as follows.

First, prepare a checkerboard. The number of checkerboards is odd and the checkerboard is flat enough to prepare a high-precision checkerboard. Then, for single-camera calibration, it is necessary to ensure that the shooting angle of each checkerboard is different, to improve the calibration accuracy, and to shoot at least 20 valid images. When multiple cameras are fixed on the production line, place the candles to be detected in suitable positions to quickly locate the camera angles. The checkerboard is shot in the overlapping area of the camera. In order to improve the final calibration accuracy, try to ensure that the checkerboard is completely photographed. Finally, when performing multi-camera calibration, the calibration is performed in pairs first, and finally the global optimization algorithm such as Bundle Adjustment or Levenberg-Marquardt is used to perform global optimization and improve the global calibration accuracy.

In the step 3, the process of synthesizing a plurality of images into one plane image includes the following steps:

Step 301: Process a plurality of images to obtain the spatial point cloud of the overlapping area of the images captured by adjacent cameras;

Step 302: Calculate the pose of the standard point cloud and the pose of the measurement point cloud through the transformation of the space coordinate system;

Step 303 : Flatten the image, and obtain a flat image of the cylinder to be inspected after being unfolded.

During image compositing, overlapping areas between cameras are guaranteed when cameras are assembled in the production line. First, using the basic theory of multi-view geometry in computer vision, the discrete point cloud of the overlapping area is calculated; secondly, for the spatial point cloud estimation of the non-overlapping area: the first step is to use the electronic candle to design the standard 3D model and the measured overlapping area. The point cloud is coarsely registered; in the second step, the ICP (Iterative Closest Point) algorithm is used to perform global fine-tuning. When calculating the composite plane image after the cylinder expansion, the back projection algorithm is used, combined with the bilinear interpolation algorithm to obtain the final pixel value.

In the step 301, the acquisition process of the spatial point cloud of the overlapping area is as follows:

Described step 3011: firstly perform region segmentation on two images captured by two adjacent cameras;

Described step 3012: in the two images after region segmentation, confirm a plurality of feature points respectively;

Step 3013: Match the feature points of the two images to each other;

Described step 3014: Screen each pair of matching points after matching, and remove the wrong matching point pair;

Step 3015: Using the matching point pair that has been successfully matched, calculate the position coordinates of the corresponding spatial point of the pair of matching points on the cylinder to be inspected.

As shown in Figure 4, the detailed process of point cloud acquisition is as follows.

1. First divide the effective area. In the assembly process of the production line, there is a large color deviation between the surrounding environment and the measured object, and the color value is used for rapid segmentation.

2. Feature point detection. Usually, SIFT (Scale-invariant feature transform, scale-invariant feature transform) is used to realize feature point detection. An algorithm of SIFT used in the field of image processing. This algorithm is scale invariant and can detect key feature points in the image.

3. For feature point matching, brute force matching can be used, or kd-tree can be established, combined with KNN for fast matching.

Using the camera calibration results, the fundamental matrix F (fundamental matrix, hereinafter referred to as F) between different cameras can be calculated. The calculation formula is as follows:

F=K' ^-T [t] _x RK ^-1 (1)

In formula (1), K and K′ are the internal parameter matrices of two adjacent cameras, respectively,

其中f_x和f_y分别表示单个相机的两个焦距，c_x和c_y为 这个相机的两个光心，表示相机光轴在图像坐标系中的偏移量；

where f _x and f _y represent the two focal lengths of a single camera, respectively, and c _x and _cy are the two optical centers of the camera, representing the offset of the camera's optical axis in the image coordinate system;

其中f′_x和f′_y分别表示单个相机的两个焦距，c'_x和c'_y为 这个相机的两个光心，表示相机光轴在图像坐标系中的偏移量；

where f' _x and f' _y represent the two focal lengths of a single camera, respectively, and c' _x and c' _y are the two optical centers of the camera, representing the offset of the camera's optical axis in the image coordinate system;

The distortion parameter is not inside this formula. In actual processing, the captured image is first subjected to de-distortion processing, and then subsequent processing is performed. K in formula (1) is the same internal parameter matrix as formula (9). R represents the rotation matrix between cameras; the vector t represents the translation vector between cameras; [t]x represents the operator that converts the vector t into an antisymmetric matrix; the relevant calculation parameters can be calculated in advance after the assembly calibration is completed, without the need for each inspection Repeated calculation.

4. Screen matching points to remove false matching points. Direct use of general feature point extraction and matching algorithms usually contains many unmatched point pairs, which greatly interfere with subsequent point cloud computing. Using the spatial constraints of the camera, interfering point pairs can be excluded. Generally, by calculating the pole distance d, when the pole distance d exceeds a certain set pole distance threshold, the interference point pair is eliminated.

The polar distance calculation formula (2) is as follows:

Among them, x and x' represent a matching point pair; F is the basic matrix, namely Fundamental Matrix in the accompanying drawing, referred to as F-Matrix; (Fx) ₁ and (Fx) ₂ represent the first two-dimensional data of the vector Fx.

5. Using the matching point pair, the spatial geometric relationship between the cameras is known, and the spatial position of the matching point pair can be calculated by using the triangulation principle. Figure 7 is a schematic diagram of the principle of a triangle, and the two parallelograms respectively represent the imaging surfaces of the two cameras. C and C' represent the optical center of the camera, x' and x represent a matching point pair, and point P is the position of the spatial point to be calculated.

In the described step 302, the process of calculating the pose of the standard point cloud and the pose of the measurement point cloud includes the following steps:

Step 3021: Process the point cloud by principal component analysis, and calculate the measurement space coordinate system;

Step 3022: Convert the point in the measurement space to the point in the design space coordinate system. .

As shown in Figure 5, the specific process of the above pose calculation is as follows.

表示，坐标原点表示为t^c。1. Principal components analysis (PCA) is performed on the calculated point cloud. The data is three-dimensional. The final calculation result will obtain three mutually perpendicular directions, which is the measurement space coordinate system, and the coordinate axis can be used with a rotation matrix.

is represented, and the origin of the coordinates is represented as t ^c .

与t^s；坐标轴用旋转矩阵

表示，坐标原点表示为t^s。2. During product design, there is a custom design space coordinate system, assuming that mathematical symbols are expressed as

with t ^s ; the axes use the rotation matrix

is represented, and the origin of the coordinates is represented as t ^s .

In addition, assuming that a point in the product design space is represented as x _s , then its coordinate x _c in the measurement space coordinate system can be expressed by formula (3):

Arranging the above formula (3), the conversion relationship between the two coordinate systems can be expressed by formula (4) and formula (5).

According to equation (6) below, the product design standard model can be transformed into the measurement space.

where x _s is a model point in design space; x _c is a point representing the transformation of x _s into measurement space.

Using the above formula (3) to convert the coordinate system, it is rough to obtain the coordinate system conversion between the measurement space and the design space, and further fine-tuning is required in the specific implementation. The fine-tuning generally adopts the Iterative Closest Point (ICP for short) algorithm to fine-tune the design space point cloud and the acquired point cloud in the measurement space, which can obtain a more accurate conversion relationship and improve the accuracy of the calculation.

The process of flattening the image in step 303 includes the following steps

Step 3031: Calculate the size of the expanded image, and convert the size of the image to the pixel size;

Step 3032: Convert the pixel position of the expanded image into position coordinates in the design space;

Step 3033: Project the point corresponding to the position coordinate in the design space on the camera imaging plane.

The process of flattening an image is shown in Figure 6, and the specific steps are as follows.

1. As shown in Figure 8, the size of the expanded image is first calculated, and the physical size needs to be converted into pixels. The cylinder radius R and height H are determined in the design stage.

Assuming that the width pixel of the final image is width(pixel), the calculation formula (7) of the height pixel number heirht is:

2. As shown in FIG. 9, convert the pixel position of the expanded image into a spatial coordinate position.

Assuming that the pixel position of any point P in the expanded image is (w, h), the calculation method of the coordinate point Q in the post-space coordinate system after conversion is as follows formula (8):

x=R*cos(α)

y=R*sin(α)

In formula (8), width and height are the pixel size of the expanded image, and the unit is pixel; H is the design physical height of the cylinder, R is the design radius of the cylinder, and the unit is the physical length; the position of point P in the expanded image is (w,h ), the unit is pixel; Q point is the cylinder point corresponding to P point, the coordinates of Q point are (x, y, z), through the above calculation, the unit is the physical length; α is the angle α in Fig. The angle between the projection of the line connecting the origin and point Q on the xy plane in the XYZ coordinate system of 9 and the x-axis.

3. As shown in Figure 10, it is a schematic diagram of converting the spatial point cloud to the camera plane.

In Figure 10, a point Q on the cylinder is projected to the camera imaging plane as M. The specific calculation formula (9) is as follows:

In formula (9), K is an internal parameter of camera calibration, which is the same as in formula (1). The distortion parameter is not included in K. Before entering the algorithm, the first step is to de-distort the image; the distortion processing is a recessive processing process, using a general camera lens, the distortion is relatively small, and no processing is required; however, If it is a wide-angle lens, the distortion is relatively serious, and distortion processing is required. It needs to be adjusted flexibly according to the actual situation.

R and t represent the selection matrix and translation vector respectively, which are external parameters of camera calibration and are predetermined after the camera is configured; u and v are the coordinate values of point M; x, y and z are the coordinate values of point Q; λ is A scaling factor, which does not affect the final calculation, is a preset constant.

4. Determine whether the projection point M is in the imaging plane of the camera.

A point Q on the cylinder cannot be seen by all cameras at the same time, so these pseudo-projection points need to be excluded; the exclusion method is relatively simple, that is, the M(u,v) obtained in the previous step can be within the captured image area.

The point (u, v) projected to the camera imaging surface is usually a floating point number, and the camera imaging position is a discrete integer value, so it is necessary to interpolate the projected pixel position. During the specific implementation, the bilinear interpolation method is usually used to obtain the pixel value of the projection position.

5. For the pixels in the overlapping area, perform weight interpolation according to the distance from the projection point M to the center of the camera.

6. Fill the final calculated pixel value to the position of the flattened image, that is, the position of point P in Figure 9.

In the step 4, the defect detection process includes the following steps:

Step 401: template matching is performed between the generated image and the standard image, and the corresponding relationship is found;

Step 402: Based on the corresponding relationship, perform block detection on the generated image and the standard image.

The calculation method of block detection can use the method of making the difference in the corresponding area. When the difference is greater than a certain threshold, an abnormality is reported, and it is determined that there is a defect.

In the step 4, the defect detection process includes the following steps:

Step 401': collect a large number of image data of qualified products, and establish a training set;

Step 402': train the deep learning network through the training set;

Step 403': Use the trained network to perform defect detection on the synthesized planar image.

When using a deep learning network for defect detection, it is necessary to collect a large amount of image data and label it; then train the deep learning network; and finally use the trained network for defect detection.

In the control method of the present invention, a plurality of cameras are arranged around the cylinder to be inspected; the multiple cameras obtain images of the cylinder to be inspected from different angles respectively; the images of the cylinder to be inspected captured by the multiple cameras at the same time A plane image is synthesized; defect detection is performed according to the synthesized plane image.

The control method of the present invention has the following technical features.

1. Through the image vision algorithm, the images captured by multiple cameras are spliced and integrated into a flat image.

2. The plane image can be directly matched with the original design drawing for template matching to quickly confirm whether there is a defect. If there is, it can accurately locate the defect position.

3. The method adopts multi-camera detection, and there are overlapping areas between adjacent cameras, which greatly improves the detection coverage rate, and the cost is low, and the detection efficiency is much higher than that of manual work.

4. The control method of the present invention uses a computer vision scheme in the whole process without manual intervention.

5. The solution of the present invention has lower requirements on hardware, so the implementation cost is low, and there is no special requirement on camera characteristics.

The method for generating a two-dimensional image of a cylinder based on a multi-eye camera of the present invention has the advantages of intelligent machine detection in the whole process, no manual intervention, high detection coverage, low cost, and high detection efficiency.

It will be apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, but that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments are to be regarded in all respects as illustrative and not restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description, which are therefore intended to fall within the scope of the appended claims. All changes within the meaning and scope of the equivalents of , are included in the present invention. Any reference signs in the claims shall not be construed as limiting the involved claim.

In addition, it should be understood that although this specification is described in terms of embodiments, not each embodiment only includes an independent technical solution, and this description in the specification is only for the sake of clarity, and those skilled in the art should take the specification as a whole , the technical solutions in each embodiment can also be appropriately combined to form other implementations that can be understood by those skilled in the art.

Claims (9)

1. a cylinder two-dimensional image generation method based on a multi-eye camera, is characterized in that: comprise the steps: Step 1: Set up multiple cameras around the cylinder to be inspected; Step 2: the multiple cameras obtain images of the cylinder to be inspected from different angles; Step 3: Combine the images of the cylinder to be inspected captured by multiple cameras at the same time into a plane image; Step 4: Perform defect detection according to the synthesized planar image. 2 . The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 1 , wherein in the step 1, among the multiple cameras, there is an overlapping area in the fields of view of two adjacent cameras. 3 . . 3. The method for generating a two-dimensional image of a cylinder based on a multi-camera camera according to claim 1, wherein in the step 1, after the camera is set, the internal parameters of the camera and the exterior of each camera house are obtained. parameter. 4. The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 1, wherein in the step 3, the process of synthesizing a plurality of images into one plane image comprises the following steps: Step 301: Process a plurality of images to obtain a spatial point cloud of the overlapping area of the images captured by adjacent cameras; Step 302: Calculate the pose of the standard point cloud and the pose of the measured point cloud through the transformation of the space coordinate system; Step 303 : Flatten the image, and obtain a flat image of the cylinder to be inspected after being unfolded. 5. The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 4, wherein in the step 301, the process of obtaining the spatial point cloud in the overlapping area is: Described step 3011: firstly perform region segmentation on two images captured by two adjacent cameras; Described step 3012: in the two images after region segmentation, confirm a plurality of feature points respectively; Step 3013: Match the feature points of the two images to each other; Described step 3014: Screen each pair of matching points after matching, and remove the wrong matching point pair; Step 3015: Using the matching point pair that has been successfully matched, calculate the position coordinates of the corresponding spatial point of the pair of matching points on the cylinder to be inspected. 6 . The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 4 , wherein in the step 302 , a process of calculating the pose of the standard point cloud and measuring the pose of the point cloud. 7 . It includes the following steps: Step 3021: Process the point cloud by principal component analysis, and calculate the measurement space coordinate system; Step 3022: Convert the point in the measurement space to the point in the design space coordinate system. . 7 . The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 4 , wherein the process of flattening the image in the step 303 comprises the following steps: 8 . Step 3031: Calculate the size of the expanded image, and convert the size of the image to the pixel size; Step 3032: Convert the pixel position of the expanded image into position coordinates in the design space; Step 3033: Project the point corresponding to the position coordinate in the design space on the camera imaging plane. 8. The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 1, wherein in the step 4, the defect detection process comprises the following steps: Step 401: template matching is performed between the generated image and the standard image, and the corresponding relationship is found; Step 402: Based on the corresponding relationship, perform block detection on the generated image and the standard image. 9 . The method for generating a two-dimensional image of a cylinder based on a multi-eye camera according to claim 1 , wherein in the step 4, the defect detection process comprises the following steps: 10 . Step 401': collect a large number of image data of qualified products, and establish a training set; Step 402': train the deep learning network through the training set; Step 403': Use the trained network to perform defect detection on the synthesized planar image.

Images