Method for automatically detecting image defects in fiber-optic gyroscope assembly process based on quantitative analysis

Abstract

The invention provides a method for automatically detecting image defects in the assembly process of a fiber-optic gyroscope based on quantitative analysis. The method comprises the following implementation steps: acquiring a training sample set and a test sample set; defining a loss function of the image defect detection network model O in the assembly process of the fiber-optic gyroscope, and carrying out iterative training on the image defect detection network model; and automatically acquiring a defect detection result in the assembly process of the fiber-optic gyroscope based on quantitative analysis. According to the method, the loss value of the position frame information is calculated through the EIOU loss function, and the width and height differences between the predicted position frame information and the real position frame information can be calculated respectively, so that the width and height differences between the predicted position frame information and the real position frame information are minimum, and the detection precision is improved; and through automatic quantitative analysis of defect position frame information in output of the trained image defect detection network model, finer and more specific detection result information is provided.

Description

Technical Field

[0001] The invention belongs to the technical field of electronic product production, and relates to a method for detecting defects in an assembly process of a fiber optic gyroscope, and specifically to an automatic method for detecting image defects in an assembly process of a fiber optic gyroscope based on quantitative analysis. Background Art

[0002] Fiber optic gyroscope is a new type of all-solid-state inertial instrument with high reliability and long life. Due to the complex assembly process of the fiber optic optical path of the fiber optic gyroscope, it may cause fiber damage, which will affect the performance of the product. Some damage will be exposed during the test phase, and some may be stress concentration defects, which will gradually appear after the product is delivered, which is extremely unfavorable to the long life of the product. After the optical fiber is wound and glued, the optical path is closed through the fusion point, and it is impossible to connect the detection instrument in the optical path for detection. The only way is to try to detect fiber defects through external images.

[0003] In the field of image processing technology, the mainstream of fiber optic gyroscope assembly process defect detection methods is based on deep learning. Among them, the fully supervised model based on deep learning is widely used due to its outstanding detection accuracy. The training samples required by the fully supervised learning model need to be manually labeled. Its essence is to regard the fiber optic gyroscope assembly process defect detection problem as a classification task in computer vision. Therefore, the fiber optic gyroscope assembly process defect detection method based on the fully supervised learning model can be regarded as an application of its related classical network in the industrial field. The classical algorithms include the two-stage detection network Faster-RCNN, which first outputs the defect interest area of ​​the fiber optic gyroscope assembly process, and then performs subsequent fiber optic gyroscope assembly process defect detection category and position detection; the single-stage detection network YOLO, etc., which directly detects the defect detection category and position of the fiber optic gyroscope assembly process. In the field of fiber optic gyroscope assembly process defect detection that emphasizes detection accuracy, the two-stage model is widely used. In the field of fiber optic gyroscope assembly process defect detection that pursues detection speed at this stage, the single-stage model is widely used.

[0004] The technical idea or basic principle of defect detection in the FOG assembly process is to use the existing image deep learning method to detect the FOG assembly process image and extract defect information. For example, Zhang Qianchuang et al. proposed a method for defect detection in FOG based on an improved image deep learning algorithm in their paper "Fiber Winding Defect Detection Based on Improved YOLO Algorithm" (Electronic Measurement Technology, May 2023). This method uses the Shufflenetv2 network to replace the convolutional layer and pooling layer in the YOLO backbone network; uses the K-means clustering algorithm to cluster the original anchor frame to obtain a prediction frame suitable for fiber winding defects; uses CIOU to calculate the positioning loss, and uses Focal loss as the confidence and category loss function to achieve network convergence. However, this method uses the CIOU loss function to only consider the difference in aspect ratio between the real position frame information and the predicted position frame information, while ignoring the difference in width and height between the real position frame information and the predicted position frame information, which affects the detection accuracy of the image defect detection network model. In addition, this invention uses the defect category, confidence and position frame information as the detection result, and the detection result information is single and difficult to quantify. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and propose an automatic detection method for image defects in the fiber optic gyroscope assembly process based on quantitative analysis, which is used to solve the technical problems existing in the prior art of low detection accuracy, single detection result information and difficulty in quantification.

[0006] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Obtain training sample set and test sample set:

[0008] Preprocess N original images including various defect categories in the fiber optic gyroscope assembly process, and annotate the position frame information and categories of the defects in the M original images and their corresponding preprocessed M×6 images. Then, the M original images and the corresponding preprocessed M×6 images and their labels form a training sample set R, and the remaining NM original images and their corresponding preprocessed (NM)×6 images form a test sample set E, where N≥50 and M≥N / 2.

[0009] (2) Define the loss function of the image defect detection network model O in the FOG assembly process:

[0010] L t =λ1L cls +λ2L obj +λ3L box

[0011] Among them, Lcls , L obj , L box They represent the loss values ​​of defect category information, confidence information, and location box information, respectively. λ1, λ2, and λ3 represent L cls , L obj , L box The weight of

[0012] (3) Iterative training of the image defect detection network model:

[0013] Iteratively train the image defect detection network model O through the training sample set R to obtain the trained image defect detection network model O*;

[0014] (4) Automatically obtain defect detection results of the fiber optic gyroscope assembly process based on quantitative analysis:

[0015] The test sample set E is used as the input of the trained image defect detection network model O* for forward propagation to obtain the defect category, confidence and position frame information of the fiber optic gyroscope assembly process corresponding to each test sample, and the defect position frame information is automatically analyzed quantitatively. Then, the curvature radius of the fiber winding path curve calculated by the quantitative analysis, as well as the category, confidence and position frame information are used together as the final defect detection result of the fiber optic gyroscope assembly process to complete the automatic acquisition of defect detection results.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] 1. In the process of training the image defect detection network model, the present invention calculates the loss value of the position frame information through the EIOU loss function, and can respectively calculate the width and height differences between the predicted position frame information and the actual position frame information, and calculates the loss value of the image defect detection network model through the category, confidence and loss value of the position frame information, so that the difference in width and height between the predicted position frame information and the actual position frame information is minimized, avoiding the defect of low precision caused by the prior art directly calculating the aspect ratio difference between the predicted position frame information and the actual position frame information through the CIOU loss function, and effectively improving the detection accuracy.

[0018] 2. In the process of obtaining defect detection results of the fiber optic gyroscope assembly process, the present invention automatically performs quantitative analysis on the defect position box information in the output of the trained image defect detection network model, thereby providing more detailed and specific detection result information, capable of evaluating the defects of the fiber optic gyroscope assembly process, and avoiding the defects of the prior art that the detection result information is single and difficult to quantify. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is an implementation flow chart of the present invention.

[0020] Figure 2 It is a structural diagram of an image defect detection network model according to an embodiment of the present invention.

[0021] Figure 3 It is a rendering of the quantitative analysis of the test image and its defect position frame information adopted by the embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0023] Reference Figure 1 , the present invention comprises the following steps:

[0024] Step 1) Obtain training sample set and test sample set:

[0025] Preprocessing N original images including various defect categories in the fiber optic gyroscope assembly process, and marking the position frame information and categories of the defects in the M original images and the corresponding preprocessed M×6 images, the M original images and the corresponding preprocessed M×6 images and their labels form a training sample set R, and the remaining NM original images and their corresponding preprocessed (NM)×6 images form a test sample set E, where N≥50, M≥N / 2, and in this embodiment, N=60, M=45;

[0026] The implementation steps for preprocessing each original image are:

[0027] Each of the N fiber optic gyroscope assembly process images of different angles and sizes magnified by an electron microscope and taken by an industrial camera is rotated and expanded, mirrored, cropped, enhanced in brightness, enhanced in contrast, and enhanced in sharpness to obtain N×6 preprocessed fiber optic gyroscope assembly process images.

[0028] Step 2) Define the loss function of the fiber optic gyroscope assembly process image defect detection network model O:

[0029] L t =λ1L cls +λ2L obj +λ3L box

[0030] Among them, L cls , L obj , L box They represent the loss values ​​of defect category information, confidence information, and location box information, respectively. λ1, λ2, and λ3 represent L cls , L obj , L box The weight of

[0031] The loss value L of category, confidence, and location box information cls , L obj , L box The calculation formulas are:

[0032] L cls =L BCE (P la ,T la )

[0033]

[0034] L obj =L BCE (P sc ,T sc )

[0035]

[0036] Among them, L BCE (·) represents the binary cross entropy loss function, T la , P la Respectively represent the true and predicted category information of each training sample, T lo , P lo Respectively represent the true and predicted location box information of each training sample, IoU(·) represents the intersection over union function, P sc Indicates T lo With P lo The intersection-over-combination ratio, P lo ∩T lo , P lo ∪T lo Respectively represent T lo With P lo The intersection and union of |·| represents the absolute value operation, P sc 、T sc Respectively represent the true and predicted confidence of each training sample, w c 、h c Respectively represent P lo With T lo The width and height of the minimum bounding rectangle C, Respectively represent P lo 、T lo The width of Respectively represent P lo 、T lo The height of , d(·) represents the Euclidean distance function, d 2 (P lo ,T lo ) indicates T lo With P lo The square of the distance from the center point, express and The square of the distance, express and The square of the distance.

[0037] In order to overcome the limitation of CIOU (Complete Intersection over Union) loss in the prior art on the model detection accuracy, the present invention introduces EIOU (Efficient Intersection over Union) loss. On the basis of CIOU loss, the aspect ratio loss is separated to calculate the width and height differences between the real position frame information and the predicted position frame information respectively. The loss function includes four parts: overlap loss, center distance loss, width loss and height loss, which can minimize the difference in width and height between the real position frame information and the predicted position frame information. The calculation formula is:

[0038]

[0039] Among them, 1-IoU(P lo ,T lo ) indicates P lo With T lo The overlap loss, Indicates P lo With T lo The center distance loss, Respectively represent P lo With T lo Width and height loss.

[0040] This embodiment constructs an image defect detection network model including a cascaded Backbone network, a Neck network, and a Head network, such as Figure 2 As shown in the figure, the Backbone network includes a Focus module, four Conv-CSP modules and an SPP module; the Neck network includes a cascaded FPN feature pyramid network and a PAN path aggregation network structure; the Head network includes three parallel convolutional layers; among them:

[0041] Backbone network, in which the Focus module includes four cascaded parallel slice layers, a splicing layer and a Conv module; the Conv module includes a cascaded single convolution layer, a single normalization layer and a single non-linear activation layer; the CSP module includes a cascaded Conv module, a parallel Conv module and a residual module, a splicing layer and a Conv module; the SPP module includes a cascaded Conv module, three parallel pooling layers, a splicing layer and a Conv module;

[0042] Neck network, in which the FPN feature pyramid network contains two Conv modules, two CSP modules, two upsampling layers and two splicing layers; the PAN path aggregation network contains two Conv modules, two CSP modules, and two splicing layers; the Conv module includes a cascaded single convolutional layer, a single normalization layer and a single non-linear activation layer; the CSP module includes a cascaded Conv module, parallel arranged Conv modules and Conv-Conv modules, a splicing layer and a Conv module.

[0043] Step 3) Iteratively train the image defect detection network model, and the implementation steps are as follows:

[0044] Step 3a) Initialize the number of iterations to t, the maximum number of iterations to T, T ≥ 200, and the image defect detection network model O for the tth iteration t The weight and bias parameters in are w t 、b t , and let t=1, in this embodiment T=300.

[0045] Step 3b) The Backbone network extracts features from each training sample; the Neck network obtains the feature map F extracted by the Backbone network B The corresponding fusion feature map F N ; Head network to F N Perform detection to obtain the location box information P of each training sample lo , confidence level P sc and category P la ,in:

[0046] The Backbone network compresses each training sample multiple times, then concatenates and convolves the compressed results, extracts features from the small-size feature maps of each training sample, and then compresses and convolves the extracted feature maps, and then concatenates the high-level feature maps obtained by compression and convolution and the low-level feature maps obtained by convolution, and then performs multiple pooling operations. Finally, all pooling results are concatenated to obtain the feature map F of each training sample. B ;

[0047] The FPN feature pyramid network and the Backbone network form a bottom-up side connection to extract the feature map F B Perform feature extraction, upsample the extracted feature map, concatenate the upsampled feature map with the feature map extracted by the Backbone network, and finally perform feature extraction on the concatenated feature map;

[0048] The PAN path aggregation network and the FPN feature pyramid network form a top-down side connection, extract features from the feature map extracted by the FPN feature pyramid network, and splice the extracted feature map with the feature map extracted by the FPN feature pyramid network. Finally, feature extraction is performed on the spliced ​​feature map to obtain the fused feature map F N ;

[0049] The Head network and the PAN path aggregation network form a side connection to the fused feature map F N Extract the detection results and obtain the location box information P of each training sample lo , confidence level P sc and category P la .

[0050] Step 3c) Use the binary cross entropy loss function and the EIOU loss function, and use the real position box information T of each training sample lo , confidence level T sc and category T la and its corresponding predicted position box information P lo , confidence level P sc and category P la Calculate the loss value L of the image defect detection network model t , then through L t For w t 、b t Update to get the detection network model O of this iteration t , the update formulas are:

[0051]

[0052] in, Respectively represent L t For the weight parameter ω t , bias parameter b t The partial derivative of α represents the learning rate, w t ', b t ' indicates ω t 、b t The updated results.

[0053] Step 3d) Determine whether t≥T holds true. If so, obtain the trained fiber optic gyroscope assembly process image defect detection network model O*. Otherwise, set t=t+1, O=O t , and proceed to step 3b).

[0054] Step 4) Automatically obtain defect detection results of the fiber optic gyroscope assembly process based on quantitative analysis:

[0055] The test sample set E is used as the input of the trained image defect detection network model O* for forward propagation, and the defect category, confidence and position frame information of the fiber optic gyroscope assembly process corresponding to each test sample are obtained. The defect position frame information is automatically quantitatively analyzed, and then the curvature radius of the fiber winding path curve calculated by quantitative analysis, as well as the category, confidence and position frame information are used together as the final defect detection result of the fiber optic gyroscope assembly process to complete the automatic acquisition of the defect detection result. The implementation steps are as follows:

[0056] Step 4a1) Defect position frame information P of the fiber optic gyroscope assembly process corresponding to each test sample lo After grayscale processing, binary processing is performed, and the average value method is selected to calculate P lo After grayscale processing, a grayscale image with color interference removed and capable of reflecting morphological and texture features is obtained. Then, the grayscale image is binarized by setting a threshold. The specific process is as follows: 1) Select a threshold β; 2) Assign 0 to the grayscale image that is less than the threshold and 255 to the grayscale image that is greater than the threshold. The calculation formula is:

[0057]

[0058] Among them, β represents the 80% quantile of the grayscale pixel value, value b Represents the pixel value in the binary image, value g Represents the pixel value in the grayscale image, and the binary image B is obtained after binarization lo ;

[0059] Step 4a2) The processed binary image B lo After clustering, denoising is performed, using the density-based DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering method to perform denoising on the binary graph B lo The pixels with a median value of 255 are clustered. After clustering by the DBSCAN algorithm, the binary graph B lo The pixels in the image are divided into different categories according to density, and finally the defective pixels are saved to achieve the denoising effect, and the denoised binary image B is obtained. l ' o ;

[0060] Step 4b) denoise the binary image B l ' o The defective pixels in the image are fitted to the denoised binary image B l ' oScan along the x direction and save the midpoint trajectory to get the rough coordinates of the fiber winding path. Use a quadratic polynomial to fit the fiber winding path to get the fiber winding path curve equation:

[0061] y=a2x 2 +a1x+a0

[0062] Where x is the coordinate of the fiber winding path curve, and the polynomial coefficients a0, a1, and a2 are obtained by least squares fitting. The fitting derivation process is as follows.

[0063] ① Given n data points (x1,y1),(x2,y2),...,(x n ,y n )

[0064] ② Let the curve equation be y=a2x 2 +a1x+a0, then for known data points, we have

[0065]

[0066] ③Set For A, For T, is χ, and the calculation formula of the polynomial coefficient χ is:

[0067]

[0068] According to the fitted optical fiber winding path curve equation, the curvature radius of the optical fiber winding path curve is calculated. The calculation formula is:

[0069]

[0070] Combined with the fiber winding path curve equation y = a2x 2 +a1x+a0, the curvature radius ρ of the fiber winding path curve can be obtained as:

[0071]

[0072] The calculated curvature radius ρ of the fiber winding path curve is used as P lo The quantitative analysis results.

[0073] The results of quantitative analysis of the images and defect position frame information in the test sample of this embodiment are as follows: Figure 3 As shown, Figure 3 (a) represents the original image of the test sample with defect location frame information, where the dotted rectangular frame represents the defect location frame information; Figure 3 (b) shows the binary image B obtained by binarizing the defect location frame information after grayscale processing lo , Figure 3 (c) shows the binary image B obtained by denoising after clustering l ' o , Figure 3 (d) represents the actual and fitted fiber winding path curves, where the smooth curve represents the fitted fiber winding path curve. Figure 3 (e) shows the result of calculating the curvature radius of the starting point, midpoint and end point of the fiber winding path curve and marking them in the defect position frame information of the original image of the test sample. Figure 3 It can be seen that the automatic quantitative analysis of the defect position frame information can obtain a fine fiber winding path curve and the curvature radius of the fiber winding path curve, providing fine and specific detection result information.

[0074] The following is a further description of the technical effects of the present invention in conjunction with the simulation results:

[0075] 1. Experimental conditions and contents:

[0076] 1) Simulation conditions: The simulation of this embodiment adopts a 12th Gen Intel(R) Core(TM) i5-12500H 2.5GHz processor, the compiler uses PyCharm Community Edition version: 11.0.15+10-b2043.56 amd64, the compilation language uses Python 3.9.12, and the compilation environment is anaconda version: conda 4.12.0;

[0077] 2) Simulation content: Based on the image data of the FOG assembly process, the network model for image defect detection of the fiber winding path in the FOG assembly process is established. The defects of the fiber winding path are quantitatively analyzed, and the curvature radius of the fiber winding path curve is obtained. The final defect detection result is obtained by combining it with the model output result, and compared with the existing method, as shown in Table 1.

[0078] 2. Experimental results analysis:

[0079] In order to evaluate the detection performance of the two methods, the average detection accuracy mAP is calculated according to the following formula:

[0080]

[0081] Where TP and FN represent the number of defects that are correctly detected and not detected, respectively, FP represents the number of non-defects that are incorrectly detected, A represents the total number of defect categories, and AP represents the total number of defect categories. a Represents the defect detection accuracy corresponding to category a.

[0082] Table 1

[0083]

[0084] It can be found from Table 1 that, compared with the optical fiber wrapping defect detection method based on the improved YOLO algorithm in the prior art, the present invention effectively improves the average accuracy of detection and can be used for subsequent practical applications.

[0085] In summary, the present invention can use the deep image learning algorithm to perform defect detection and quantitative analysis on the input fiber optic gyroscope assembly process image, and display the curvature radius of the fiber optic winding path curve in the fiber optic gyroscope assembly process image, as well as the defect category, confidence level and position frame information. The present invention is rigorous, complete, and has high detection accuracy, and can provide detailed and specific detection result information.

Claims

1. A method for automatically detecting image defects in the assembly process of a fiber optic gyroscope based on quantitative analysis, characterized in that: The steps include: (1) Obtain training sample set and test sample set: Preprocess N original images including various defect categories in the fiber optic gyroscope assembly process, and annotate the position frame information and categories of the defects in the M original images and their corresponding preprocessed M×6 images. Then, the M original images and the corresponding preprocessed M×6 images and their labels form a training sample set R, and the remaining NM original images and their corresponding preprocessed (NM)×6 images form a test sample set E, where N≥50 and M≥N / 2. (2) Define the loss function of the image defect detection network model O in the FOG assembly process: L t =λ1L cls +λ2L obj +λ3L box Among them, L cls , L obj , L box They represent the loss values ​​of defect category information, confidence information, and location box information, respectively. λ1, λ2, and λ3 represent L cls , L obj , L box The weight of (3) Iterative training of the image defect detection network model: Iteratively train the image defect detection network model O through the training sample set R to obtain the trained image defect detection network model O*; (4) Automatically obtain defect detection results of the fiber optic gyroscope assembly process based on quantitative analysis: The test sample set E is used as the input of the trained image defect detection network model O* for forward propagation to obtain the defect category, confidence and position frame information of the fiber optic gyroscope assembly process corresponding to each test sample, and the defect position frame information is automatically analyzed quantitatively. Then, the curvature radius of the fiber winding path curve calculated by the quantitative analysis, as well as the category, confidence and position frame information are used together as the final defect detection result of the fiber optic gyroscope assembly process to complete the automatic acquisition of defect detection results.

2. The method according to claim 1, characterized in that The preprocessing of N images including various defect categories in the fiber optic gyroscope assembly process described in step (1) is implemented by: Each of the N fiber optic gyroscope assembly process images of different angles and sizes magnified by an electron microscope and taken by an industrial camera is rotated and expanded, mirrored, cropped, enhanced in brightness, enhanced in contrast, and enhanced in sharpness to obtain N×6 preprocessed fiber optic gyroscope assembly process images.

3. The method according to claim 1, characterized in that The fiber optic gyroscope assembly process image defect detection network model described in step (2) includes a cascaded Backbone network, a Neck network and a Head network, wherein: The Backbone network includes a cascaded Focus module, multiple Conv-CSP modules and an SPP module; the Neck network includes a cascaded FPN feature pyramid network and a PAN path aggregation network structure; the Head network includes multiple parallel convolutional layers.

4. The method according to claim 3, characterized in that The loss value L of the category information, confidence information, and location box information described in step (2) cls , L obj , L box , the calculation formula is: L cls =L BCE (P la ,T la ) L obj =L BCE (P sc ,T sc ) Among them, L BCE () represents the binary cross entropy loss function, T la , P la Respectively represent the true and predicted category information of each training sample, T lo , P lo Respectively represent the true and predicted location box information of each training sample, IoU(·) represents the intersection over union function, P sc Indicates T lo With P lo The intersection-over-combination ratio, P lo ∩T lo , P lo ∪T lo Respectively represent T lo With P lo The intersection and union of |·| represents the absolute value operation, P sc , T sc Respectively represent the true and predicted confidence of each training sample, w c 、h c Respectively represent P lo With T lo The width and height of the minimum bounding rectangle C, Respectively represent P lo , T lo The width of Respectively represent P lo , T lo The height of , d(·) represents the Euclidean distance function, d 2 (P lo ,T lo ) indicates T lo With P lo The square of the distance from the center point, express and The square of the distance, express and The square of the distance.

5. The method according to claim 4, characterized in that The iterative training of the image defect detection network model O described in step (3) is implemented as follows: (3a) Initialize the number of iterations to t, the maximum number of iterations to T, T ≥ 200, and the image defect detection network model O in the tth iteration t The weight and bias parameters in are w t 、b t , and let t = 1; (3b) The Backbone network extracts features from each training sample; the Neck network obtains the feature map F extracted by the Backbone network. B The corresponding fusion feature map F N ; Head network to F N Perform detection to obtain the location box information P of each training sample lo , confidence level P sc and category P la ; (3c) The binary cross entropy loss function and the EIOU loss function are used, and the real position box information T of each training sample is used lo , confidence level T sc and category T la and its corresponding predicted position box information P lo , confidence level P sc and category P la Calculate the loss value L of the image defect detection network model t , then through L t For w t 、b t Update to get the detection network model O of this iteration t ; (3d) Determine whether t ≥ T is true. If so, obtain the trained fiber optic gyroscope assembly process image defect detection network model O*. Otherwise, let t = t + 1, O = O t , and execute step (3b).

6. The method according to claim 5, characterized in that The Backbone network described in step (3b) extracts features from each training sample. The implementation steps are: The Backbone network compresses each training sample multiple times, then concatenates and convolves the compressed results, extracts features from the small-size feature maps of each training sample, and then compresses and convolves the extracted feature maps, and then concatenates the high-level feature maps obtained by compression and convolution and the low-level feature maps obtained by convolution, and then performs multiple pooling operations. Finally, all pooling results are concatenated to obtain the feature map F of each training sample. B .

7. The method according to claim 5, characterized in that The Neck network described in step (3b) obtains the feature map F extracted by the Backbone network B The corresponding fusion feature map F N , the implementation steps are: The feature map F extracted by FPN feature pyramid network on Backbone network B Perform feature extraction, upsample the extracted feature map, concatenate the upsampled feature map with the feature map extracted by the Backbone network, and finally perform feature extraction on the concatenated feature map; The PAN path aggregation network extracts features from the feature map extracted by the FPN feature pyramid network, and concatenates the extracted feature map with the feature map extracted by the FPN feature pyramid network. Finally, feature extraction is performed on the concatenated feature map to obtain the fused feature map F N .

8. The method according to claim 5, characterized in that The pair w described in step (3c) t 、b t To update, the update formulas are: in, Respectively represent L t For the weight parameter ω t , bias parameter b t The partial derivative of α represents the learning rate, w t ', b t ' indicates ω t 、b t The updated results.

9. The method according to claim 1, characterized in that: The automatic quantitative analysis of the defect location frame information described in step (4) is implemented by: (4a) The defect position frame information P of the FOG assembly process corresponding to each test sample lo After grayscale processing, binary processing is performed, and the binary image B lo After clustering, denoising is performed to obtain the denoised binary image B. l ' o ; (4b) The denoised binary image B l ' o After fitting the defect pixel points in the image, the curvature radius ρ of the fiber winding path curve obtained by fitting is calculated and used as the defect position frame information P lo Quantitative analysis results.

Images