Logo alignment method based on 3D curved surface dynamic calibration of 2D vision

Abstract

The application discloses a 3D curved surface dynamic calibration LOGO alignment method based on 2D vision, and relates to the field of computer vision recognition. The method comprises two stages of calibration and alignment. The calibration process realizes automatic calculation of an image coordinate and an angle conversion formula through equiangular interval collection of insulating vessel or vacuum vessel pictures. The alignment process calculates a rotation angle according to a located LOGO positioning reference coordinate, and realizes visual feedback alignment. The application adopts a visual mode to replace a manual alignment mode, can meet full-automatic uninterrupted production needs of a flow line in speed, greatly improves production efficiency, improves stability of insulating vessel or vacuum vessel marking precision, and can realize a full-automatic marking process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision recognition, and particularly relates to a 3D curved surface dynamic calibration LOGO alignment method based on 2D vision. BACKGROUND

[0002] Laser marking, as a modern precision machining method, has the advantages of maintenance-free, high flexibility, high reliability and the like compared with traditional machining methods such as printing, mechanical scribing and electric spark machining, and is widely used in the field of hardware products. Existing marking of heat preservation vessels or vacuum vessels is mainly manual marking. When marking, the worker clamps the bottom of the heat preservation vessel or vacuum vessel on a special jig, aligns the LOGO used as a positioning reference with the linear laser on the workbench, and then manually operates the marking machine to mark. The accuracy of manual alignment varies from person to person, and long-time operation will also cause accuracy to decrease due to personnel fatigue, and manual marking cannot realize full-automatic production with the production line.

[0003] In the prior art, a cylindrical product automatic alignment visual marking system invented by Yu Deshan et al. also uses a rotating device to drive the cylindrical product to rotate to mark multiple positions on the cylindrical surface, and uses vision to assist alignment. The rotary marking machine of Zhang Liang et al. also has a rotating mechanism to drive the sample to rotate, and the vision part mainly focuses on monitoring the marking quality during marking and removing unqualified products. The above prior art does not have the function of aligning the cylindrical surface features, and the use of a fixed jig is only effective for a small number of regular-shaped samples, and the positioning accuracy is not high. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application aims to provide a 3D curved surface dynamic calibration LOGO alignment method based on 2D vision which uses computer vision instead of manual alignment, can meet the needs of full-automatic uninterrupted production of the production line in terms of speed, and has high positioning accuracy.

[0005] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0006] The 3D curved surface dynamic calibration LOGO alignment method based on 2D vision comprises the following steps:

[0007] S1, selecting a positioning reference, training to obtain a YOLOv3-tiny positioning model;

[0008] S2, executing a calibration process to automatically position the LOGO image reference feature coordinates through the vessel pictures collected at equal angle intervals, and fitting a conversion formula of coordinates and angles based on the positioned reference feature coordinates and the corresponding angles;

[0009] S3, performing a bit flow, calculating the angle of rotation according to the located LOGO positioning reference coordinate, and realizing visual feedback alignment.

[0010] Preferably, in the above-mentioned 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method, the calibration process comprises the steps of:

[0011] S21, manually placing the vessel on the jig, acquiring the image of the current vessel through the camera and displaying it on the interface, and aligning the positioning reference to the mark position on the interface;

[0012] S22, set the mark position as the zero position of alignment, and control the vessel to rotate step by step at a set angle interval within the angle range on both sides of the zero position, while triggering the camera to collect images at corresponding angles;

[0013] S23, positioning the reference feature coordinates of the LOGO in the image collected in step S12 based on the YOLOv3-tiny positioning model;

[0014] S24, based on the set of located reference coordinates and the set of angles of the corresponding images, fitting the conversion formula of coordinates and angles, and the conversion formula is a linear model a'=kx+b, wherein k and b are coefficients to be calculated, and the formula is used to calculate the deviation sum of squares of the actual angle a of the sample i and the predicted angle a' i According to the principle of minimum deviation sum of squares, the least squares method is used to fit the coefficients k and b;

[0015] S25, calculating R 2 of the fitted linear model a'=kx+b , wherein SSE is the residual sum of squares, SST is the total sum of squares, a i ' is the predicted angle value, a i is the actual angle value, is the sample mean, and R 2 is evaluated between the intervals [0, 1], and R 2 evaluates the fitting accuracy, and when the fitting accuracy reaches the set threshold T, the calibration is ended, the calibration coefficient is saved, otherwise the calibration process is repeated until the calibration is completed.

[0016] Preferably, in the above-mentioned 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method, the alignment process comprises the steps of:

[0017] S31, placing the vessel on the fixed jig;

[0018] S32, control the vessel to rotate at a constant speed for one revolution and trigger the camera to collect an image at an equal angle interval;

[0019] S33, input the image collected in step S32 into a YOLOv3-tiny positioning model, and use the YOLOv3-tiny positioning model to coarsely position the reference character;

[0020] S34, based on the positioning frame of the YOLOv3-tiny positioning model, extract the region where the local feature serving as the positioning reference is located, and extract the ROI;

[0021] S35, perform histogram equalization and binarization processing on the ROI region, and segment out the positioning reference feature;

[0022] S36, perform projection on the binary image and filter the peak value, and position the character feature selected as the positioning reference, to further improve the positioning accuracy;

[0023] S37, output the coordinates of the positioning reference feature, and convert them into a predicted angle using a calibration conversion formula;

[0024] S38, select an angle from the images collected in step S32 that is closest to the calibration zero position, control the rotating mechanism to rotate the vessel to the specified angle and stop, and trigger image acquisition once;

[0025] S39, perform the positioning operations of steps S33 to S37 on the images collected in step S38, and control the vessel to rotate by a specified angle to complete the alignment.

[0026] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for realizing LOGO alignment based on 2D vision, in step S24, the coordinates X={x1, x2, x3…x n} and the angle set A={a1, a2, a3…a n} corresponding to the image are transmitted to a fitting algorithm, and the fitting algorithm uses a least square method to fit the linear model a′=kx+b of the coordinates x and the actual angle a.

[0027] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for realizing LOGO alignment based on 2D vision, in step S25, the coordinate set X is input into the linear model a′=kx+b, the predicted angle a′ is differentiated from the actual angle a to obtain the residual of the linear model a′=kx+b, and the R 2 of the linear model a′=kx+b is calculated according to the calculation formula 1-SSESST. 2 When the R 2 reaches a set threshold T, the calibration is completed, and the calibration coefficient is saved, otherwise the calibration process is repeated until the calibration is completed.

[0028] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for realizing LOGO alignment based on 2D vision, step S33 specifically includes the following steps:

[0029] Step S331, input the image set I = {i1, i2, i3…i n} collected in step S32 into the YOLOv3-tiny positioning model, and use the YOLOv3-tiny positioning model to perform coarse positioning on the reference character;

[0030] Step S332, perform feature extraction in a convolutional manner through the YOLOv3-tiny positioning model to obtain three kinds of feature maps of 13*13, 26*26, and 52*52 scales;

[0031] Step S333, predict 3 bounding boxes on each of the three feature maps, for a total of 9 bounding boxes;

[0032] Step S334, perform non-maximum suppression using the positions and scores of the bounding boxes, and finally obtain the bounding box with the largest probability as the positioning box of the reference character.

[0033] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for LOGO alignment based on 2D vision, in step S37, the coordinates x' of the positioning reference feature are output, the generated linear model a' = kx + b after calibration is converted into the corresponding predicted angle a', and a set of rotation angle sets A' = {a'1, a'2, a'3…a'N} corresponding to the images are obtained. n .

[0034] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for LOGO alignment based on 2D vision, in step S38, the angle a best closest to the calibration zero position is selected from the rotation angle set A' output in step S37, the controller stops the rotation of the vessel at the specified angle a best , and triggers image capture once.

[0035] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for LOGO alignment based on 2D vision, the threshold value T ≥ 0.98.

[0036] Preferably, in the above-mentioned 3D curved surface dynamic calibration method for LOGO alignment based on 2D vision, the training process of the YOLOv3-tiny positioning model includes the following steps:

[0037] S11, take pictures to collect vessel sample pictures of different styles and colors, wherein the sample pictures contain LOGO images;

[0038] S12, select a character in the LOGO as a positioning reference, and manually label the position and bounding box of the character on the image;

[0039] S13, an improved YOLOv3-tiny network is established, the basic feature extraction module of which is composed of a 3*3 convolution layer and a pooling layer, followed by a normalization layer and an activation function Leaky;

[0040] S14, the output layer of the improved YOLOv3-tiny network is improved, the original 2 feature layers are changed into 3 feature layers, the multi-scale detection performance is enhanced, and 9 bounding boxes are predicted in total, corresponding to 13*13, 26*26 and 52*52 three scales;

[0041] S15, the loss function of the network is determined by using GIOU, which is used as an evaluation index for measuring the performance of the model;

[0042] S16, the data set is input into the improved YOLOv3-tiny network for training, so as to obtain a YOLOv3-tiny positioning model for target detection of selected characters.

[0043] The present application has the advantages that: the present application uses computer vision instead of manual alignment, which can meet the needs of full-automatic uninterrupted production of the assembly line in speed, improves the stability of the marking precision of the heat preservation vessel or the vacuum vessel, realizes a full-automatic marking process, greatly improves the marking production efficiency of the heat preservation vessel or the vacuum vessel, and reduces the labor cost. In addition, the present application does not require the shell of the heat preservation vessel or the vacuum vessel to have a specific shape, and has more advantages in universality and precision than the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0044] Fig. 1 The calibration flowchart of the present application is shown in the figure;

[0045] Fig. 2 The positioning flowchart of the present application is shown in the figure;

[0046] Fig. 3 The flowchart of the positioning reference feature in the positioning flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0047] The application will be described further with reference to the drawings and specific examples in the description, it should be noted that before any embodiment of the application is explained in detail, the application is not limited to the component configuration and arrangement details set forth in the following description or shown in the following drawings in its application. The application can have other embodiments and can be practiced or implemented in various ways. Moreover, it should be understood that the language and terminology used here are for the purpose of description and should not be considered as limiting. The use of "include", "contain" or "have" and their variants herein means to cover the items listed thereafter and their equivalents and additional items. The terms "mount", "connect" and "couple" are used broadly and include direct and indirect mounting, connecting and coupling. In addition, "connect" and "couple" are not limited to physical or mechanical connections or couplings, but can include direct or indirect electrical connections or electrical couplings.

[0048] Please refer to Figs. 1 to 3 As shown, the embodiment of the application proposes a 3D curved surface dynamic calibration based on 2D vision LOGO alignment method, including steps:

[0049] S1, select positioning reference, train to obtain YOLOv3-tiny positioning model;

[0050] S2, execute calibration process, realize automatic positioning of LOGO image reference feature coordinates through equiangular interval collected vessel pictures, based on the positioned reference feature coordinates and corresponding angles, fit the conversion formula of coordinates and angles;

[0051] S3, execute alignment process, calculate the angle of rotation according to the positioned LOGO positioning reference coordinates, realize visual feedback alignment.

[0052] Specifically, in the preferred embodiment of the application, the training process of the YOLOv3-tiny positioning model includes steps:

[0053] S11, take pictures to collect vessel sample pictures of different styles and colors, wherein the sample pictures contain LOGO images;

[0054] S12, select a character in the LOGO as the positioning reference, manually mark the position and bounding box of the character on the image;

[0055] S13, establish an improved YOLOv3-tiny network, the basic feature extraction module of which is composed of 3*3 convolution layer and pooling layer, followed by normalization layer and activation function Leaky;

[0056] S14, improving the output layer of the YOLOv3-tiny network, changing the original two feature layers into three feature layers for output, enhancing the multi-scale detection performance, and predicting three bounding boxes at each scale corresponding to 13*13, 26*26 and 52*52, and a total of nine bounding boxes are predicted;

[0057] S15, using GIOU to determine the loss function of the network, so as to use it as an evaluation index for measuring the performance of the model;

[0058] S16, inputting the data set into the improved YOLOv3-tiny network for training, so as to obtain a YOLOv3-tiny positioning model for target detection of the selected characters.

[0059] Specifically, in the preferred embodiment of the application, the calibration process comprises the steps of:

[0060] S21, manually placing the vessel on the jig, acquiring the image of the current vessel through the camera and displaying it on the interface, and aligning the positioning reference to the mark position on the interface;

[0061] S22, setting the mark position as the zero position of alignment, and controlling the vessel to rotate step by step at a set angle interval within the angle range on both sides of the zero position, while triggering the camera to collect images at corresponding angles;

[0062] S23, positioning the reference feature coordinates of the LOGO in the image collected in step S12 based on the YOLOv3-tiny positioning model;

[0063] S24, fitting a conversion formula for converting coordinates and angles based on the set of positioned reference coordinates and the set of corresponding angles of the images, and the conversion formula is a linear model a'=kx+b, where k and b are coefficients to be calculated, and the formula is used to calculate the deviation sum of squares of the actual angle a of the sample i and the predicted angle a' i , and the least squares method is used to fit the coefficients k and b according to the principle of minimum deviation sum of squares;

[0064] S25, calculating R 2 of the fitted linear model a'=kx+b , wherein SSE is the residual sum of squares, SST is the total sum of squares, a' i is the predicted angle value, a i is the actual angle value, is the sample mean, and R 2 is evaluated in the interval [0, 1], and R 2 evaluates the fitting accuracy, and R 2The greater the model precision is, the higher the model precision is. When the fitting precision reaches a set threshold T, the calibration ends, the calibration coefficients are saved, otherwise the calibration process is repeated until the calibration is completed.

[0065] In step S24, the reference coordinate set X={x1, x2, x3…x n} and the angle set A={a1, a2, a3…a n} corresponding to the images are transmitted to the fitting algorithm, and the fitting algorithm uses the least square method to fit the linear model a′=kx+b converted from the coordinates x and the actual angles a. In step S25, the coordinate set X is input into the linear model a′=kx+b, the predicted angles a′ are differentiated from the actual angles a to obtain the residual of the linear model a′=kx+b, and the R 2 of the linear model a′=kx+b is calculated according to the calculation formula 1-SSE / SST. 2 When the R n reaches a set threshold T, the calibration ends, the calibration coefficients are saved, otherwise the calibration process is repeated until the calibration is completed.

[0066] Specifically, in step S24, the square sum of deviations is calculated according to the least square principle. The partial derivatives of k and b are calculated respectively, and the partial derivatives are set to 0, so that:

[0067]

[0068]

[0069] Further derivation obtains:

[0070]

[0071] According to the above formula, it is obtained that:

[0072]

[0073]

[0074] The reference coordinate set X={x1, x2, x3…x n} and the angle set A={a1, a2, a3…a n} corresponding to the images are input into the above formula to obtain the fitting coefficients k and b, and the fitting of the linear model is completed.

[0075] Specifically, in the preferred embodiment of the present application, the alignment process comprises the following steps:

[0076] S31, placing the vessel on the fixing jig;

[0077] S32, controlling the vessel to rotate at a constant speed for one revolution and triggering the camera to capture an image at an equal angle interval;

[0078] S33, input the image collected in step S32 into a YOLOv3-tiny positioning model, and use the YOLOv3-tiny positioning model to coarsely position the reference character;

[0079] S34, based on the positioning frame of the YOLOv3-tiny positioning model, extract the region where the local feature serving as the positioning reference is located, and extract the ROI;

[0080] S35, perform histogram equalization and binarization processing on the ROI region, and segment out the positioning reference feature;

[0081] S36, perform projection on the binary image and filter the peak value, and perform positioning on the character feature serving as the positioning reference, to further improve the positioning accuracy;

[0082] S37, output the coordinates of the positioning reference feature, and convert them into a predicted angle using a calibration conversion formula;

[0083] S38, select an angle from the images collected in step S32 that is closest to the calibration zero position, and control the rotating mechanism to rotate the vessel to the specified angle and stop and trigger image acquisition once;

[0084] S39, perform the positioning operations of steps S33 to S37 on the image collected in step S38, and control the vessel to rotate by the specified angle to complete the alignment.

[0085] The step S33 specifically includes the following steps:

[0086] Step S331, input the image set I = {i1, i2, i3…i n} collected in step S32 into the YOLOv3-tiny positioning model, and use the YOLOv3-tiny positioning model to coarsely position the reference character;

[0087] Step S332, perform feature extraction in a convolutional manner through the YOLOv3-tiny positioning model to obtain three kinds of feature maps with scales of 13*13, 26*26 and 52*52;

[0088] Step S333, predict 3 bounding boxes on each of the three feature maps, for a total of 9 bounding boxes;

[0089] Step S334, use the position and score of the bounding box to perform non-maximum suppression, and finally obtain the bounding box with the largest probability as the positioning frame of the reference character.

[0090] Specifically, in step S37, the coordinates x' of the positioning reference feature are output, and the generated linear model a'=kx+b after calibration is converted into the corresponding predicted angle a', to obtain a set of rotation angles A' corresponding to a set of images A'={a'1, a'2, a'3... a'n}. n In step S38, the angle a best closest to the calibration zero position is selected from the set of rotation angles A' output in step S37, the controller stops the rotation of the vessel at the specified angle a best , and triggers the image capture.

[0091] As a preferred embodiment of the present application, the threshold value T≥0.98, when R 2 reaches the set threshold value, it is considered that the fitting accuracy meets the requirements, at which point the calibration can be ended.

[0092] In the specific implementation process, it includes a fixed jig, a servo motor, a PLC controller, a marking mechanism, and a vision mechanism composed of a computer, a high-precision CCD camera, and a coaxial light source. In the positioning process, the thermal vessel or vacuum vessel is placed on the fixed jig by manual or robotic transfer, and the software communicates with the PLC to control the rotation mechanism to rotate. The vision mechanism takes pictures at fixed angles during one rotation for coarse positioning, the rotation mechanism stops at the optimal angle position, and the vision mechanism takes pictures again for fine positioning. The rotation mechanism is positioned according to the vision feedback, and the marking mechanism performs marking after positioning.

[0093] As an embodiment, in addition to using a camera for LOGO character positioning, a photoelectric sensor can also be used to position the LOGO character. Compared with the present application, the photoelectric sensor has certain advantages in cost, but is not as good as the present application in positioning accuracy and robustness.

[0094] The basic principles, main features, and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A method for LOGO alignment based on 3D curved surface dynamic calibration of 2D vision, characterized in that, The method comprises the steps of: S1, selecting a positioning reference, training a YOLOv3-tiny positioning model; S2, performing a calibration process, automatically positioning the LOGO image reference feature coordinates through the vessel pictures collected at equal angle intervals, fitting the conversion formula of coordinates and angles based on the positioned reference feature coordinates and the corresponding angles; S3, performing a positioning process, calculating the angle of rotation according to the positioned LOGO positioning reference coordinates, and realizing visual feedback positioning; The calibration process comprises the steps of: S21, manually placing the vessel on the jig, acquiring the image of the current vessel through the camera and displaying it on the interface, and aligning the positioning reference to the mark position on the interface; S22, setting the mark position as the zero position of the alignment, and controlling the vessel to rotate step by step at a set angle interval within the angle range on both sides of the zero position, while triggering the camera to collect images at the corresponding angles; S23, positioning the reference feature coordinates of the LOGO in the images collected in step S22 based on the YOLOv3-tiny positioning model; S24, based on the located reference coordinate set and the corresponding image angle set, fitting a conversion formula for coordinate and angle conversion, the conversion formula being a linear model a'=kx+b, wherein k and b are coefficients to be calculated, using the formula to calculate the sample actual angle a i and the predicted angle a i deviation sum, according to the principle of minimum deviation sum, fitting the coefficients k and b using the least squares method; S25, calculate R for the fitted linear model a' = kx + b 2 , wherein SSE is the sum of squared residuals, SST is the total sum of squares, a' i is the predicted angle value, a i is the actual angle value, is the sample mean, R 2 is evaluated between the interval [0, 1], and R 2 evaluates the fitting accuracy, and the calibration ends when the fitting accuracy reaches a set threshold T, the calibration coefficients are saved, otherwise the calibration process is repeated until the calibration is completed.

2. The 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method according to claim 1, characterized in that, The positioning process comprises the steps of: S31, placing the vessel on the fixed jig; S32, controlling the vessel to rotate at a constant speed for one revolution and triggering the camera to collect an image at an equal angle interval; S33, inputting the image collected in step S32 into the YOLOv3-tiny positioning model, and using the YOLOv3-tiny positioning model to coarsely position the reference character; S34, extracting the region of the local feature serving as the positioning reference based on the positioning box of the YOLOv3-tiny positioning model, and extracting the ROI; S35, performing histogram equalization and binarization processing on the ROI region, and segmenting out the positioning reference feature; S36, performing projection on the binary image and selecting the peak value, positioning the character feature selected as the positioning reference, and further improving the positioning accuracy; S37, outputting the coordinates of the positioning reference feature, and converting them into a predicted angle using the calibration conversion formula; S38, selecting the angle closest to the calibration zero position from the predicted angle in the image collected in step S32, controlling the rotation mechanism to rotate the vessel to the specified angle and stopping, and triggering the image collection once; S39, performing the positioning operations of steps S33 to S37 on the image collected in step S38, and controlling the vessel to rotate the specified angle to complete the positioning.

3. The 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method according to claim 1, characterized in that, In step S24, the reference coordinate set X = {x1, x2, x3…x} is located. n } and the corresponding angle set A = {a1, a2, a3…a} of the image n The data is passed to the fitting algorithm, which uses the least squares method to fit a linear model a′=kx+b converted from coordinate x and actual angle a.

4. The 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method according to claim 1, characterized in that, In step S25, the coordinate set X is input into the linear model a'=kx+b, the predicted angle a' is differentiated with the actual angle a to obtain the residual of the linear model a'=kx+b, and R of the linear model a'=kx+b is calculated according to the calculation formula 1-SSE / SST 2 When R 2 When the set threshold T is reached, the calibration ends, and the calibration coefficients are saved, otherwise the calibration process is repeated until the calibration is completed.

5. The 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method according to claim 2, characterized in that, Step S33 specifically comprises the steps of: Step S331, input the image set I = {i1, i2, i3…i n} collected in step S32 into the YOLOv3-tiny positioning model, and use the YOLOv3-tiny positioning model to coarsely position the reference character; Step S332, extracting features in a convolutional manner through the YOLOv3-tiny positioning model to obtain three kinds of feature maps with scales of 13*13, 26*26 and 52*52; Step S333, predicting 3 bounding boxes on each of the three feature maps, for a total of 9 bounding boxes; Step S334, performing non-maximum suppression using the position and score of the bounding box, and finally obtaining the bounding box with the largest probability as the positioning box of the reference character.

6. The 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method according to claim 2, characterized in that, In step S37, the coordinates x' of the positioning reference feature are output, and the generated linear model a'=kx+b after calibration is converted into the corresponding predicted angle a', to obtain a set of rotation angles A' corresponding to the image set A'={a1', a'2, a3'...a'n}. n} 7.The 3D curved surface dynamic calibration based on 2D vision to realize LOGO alignment method according to claim 2, characterized in that, In step S38, the angle a closest to the calibration zero position is selected from the set of rotation angles A' output in step S37 best The controller stops the rotation of the vessel at the position of the specified angle a best and triggers a single shot. 8.The 2D vision based 3D curved surface dynamic calibration method for LOGO alignment according to claim 1, wherein, The threshold value T is greater than or equal to 0.

98. 9.The 2D vision based 3D curved surface dynamic calibration method for LOGO alignment according to claim 1, wherein, The training process of the YOLOv3-tiny positioning model comprises the steps of: S11, photographing and collecting vessel sample pictures of different styles and colors, wherein the sample pictures contain LOGO images; S12, select a character in the LOGO as a positioning reference, manually label the position and bounding box of the character on the image; S13, an improved YOLOv3-tiny network is established, the basic feature extraction module of which is composed of a 3*3 convolution layer and a pooling layer, followed by a normalization layer and an activation function Leaky; S14, the output layer of the improved YOLOv3-tiny network is improved, the original 2 feature layers are changed to 3 feature layers, and the multi-scale detection performance is enhanced, which corresponds to 13*13, 26*26 and 52*52 three scales, 3 bounding boxes are predicted on each scale, and a total of 9 bounding boxes are predicted; S15, using GIOU to determine the loss function of the network, which is used as an evaluation index to measure the performance of the model; S16, input the data set into the improved YOLOv3-tiny network for training, so as to obtain the YOLOv3-tiny positioning model for target detection of the selected character.

Images