An Automatic Defect Detection Method Based on Quantitative Analysis in the Fiber Optic Gyroscope Assembly Process

Abstract

This invention proposes an automatic defect detection method for fiber optic gyroscope assembly process images based on quantitative analysis. The steps are as follows: acquiring training and testing sample sets; defining the loss function of the image defect detection network model O for the fiber optic gyroscope assembly process and iteratively training the image defect detection network model; and automatically acquiring the defect detection results of the fiber optic gyroscope assembly process based on quantitative analysis. This invention calculates the loss value of the bounding box information using the EIOU loss function, which can separately calculate the width and height differences between the predicted and actual bounding box information, minimizing the differences in width and height between the predicted and actual bounding box information, thus improving detection accuracy. Furthermore, by automatically performing quantitative analysis on the defect bounding box information in the output of the trained image defect detection network model, more refined and specific detection result information is provided.

Description

Technical Field

[0001] This invention belongs to the field of electronic product manufacturing technology and relates to a defect detection method in the assembly process of fiber optic gyroscopes, specifically a method for automatic image defect detection in the assembly process of fiber optic gyroscopes based on quantitative analysis. Background Technology

[0002] Fiber optic gyroscopes are a new type of all-solid-state inertial instrument characterized by high reliability and long lifespan. However, due to the complex assembly process of the fiber optic path in fiber optic gyroscopes, fiber damage may occur. This damage affects product performance; some defects may be detected during the testing phase, while others may be stress concentration defects that gradually appear after product delivery, severely impacting the product's longevity. After the fiber is wound and glued, the optical path is closed through the splice, making it impossible to connect testing instruments within the optical path for inspection. Therefore, fiber defects can only be detected through external imaging.

[0003] In the field of image processing technology, the mainstream method for detecting defects in the fiber optic gyroscope assembly process is based on deep learning. Among them, fully supervised deep learning models are widely used due to their outstanding detection accuracy. The training samples required for fully supervised learning models are all manually labeled. Essentially, it treats 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 fully supervised learning models can be seen as an application of its related classic networks in the industrial field. Classic algorithms include the two-stage detection network Faster-RCNN, which first outputs the region of interest (ROI) of the fiber optic gyroscope assembly process defect, and then performs subsequent detection of the defect category and location; and the single-stage detection network YOLO, which directly detects the category and location of the fiber optic gyroscope assembly process defect. In the field of fiber optic gyroscope assembly process defect detection, where detection accuracy is emphasized, two-stage models are more commonly used. In the current field of fiber optic gyroscope assembly process defect detection, where detection speed is prioritized, single-stage models are more widely used.

[0004] The technical approach or basic principle of defect detection in the fiber optic gyroscope assembly process is to use existing image deep learning methods to detect defects in the images of the fiber optic gyroscope assembly process and extract defect information. For example, in their paper "Fiber Optic Winding Defect Detection Based on Improved YOLO Algorithm" (Electronic Measurement Technology, May 2023), Zhang Qianchuang et al. proposed a fiber optic winding defect detection method based on an improved image deep learning algorithm. This method uses the Shufflenetv2 network to replace the convolutional and pooling layers in the YOLO backbone network; it uses the K-means clustering algorithm to cluster the original anchor boxes to obtain prediction boxes suitable for fiber optic winding defects; it uses CIOU to calculate the localization 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 true and predicted bounding box information, while ignoring the difference in width and height between the true and predicted bounding box information. This affects the detection accuracy of the image defect detection network model. Furthermore, this invention uses the defect category, confidence level, and bounding box information as the detection result, resulting in a single and difficult-to-quantify detection result. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and propose an automatic detection method for image defects in the assembly process of fiber optic gyroscopes based on quantitative analysis. This method is used to solve the technical problems of low detection accuracy, single detection result information and difficulty in quantification in the existing technology.

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

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

[0008] N original images containing various defect categories during the fiber optic gyroscope assembly process are preprocessed. The location bounding box information and category of defects in M ​​original images and their corresponding preprocessed M×6 images are labeled. The M original images and their corresponding preprocessed M×6 images and their labels are combined to form a training sample set R. The remaining NM original images and their corresponding preprocessed (NM)×6 images are combined to form a test sample set E, where N≥50 and M≥N / 2.

[0009] (2) Define the loss function of the network model O for image defect detection during the fiber optic gyroscope assembly process:

[0010] L t =λ1L cls +λ2L obj +λ3L box

[0011] Among them, Lcls L obj L box Let λ1, λ2, and λ3 represent the loss values ​​for defect category information, confidence level information, and location box information, respectively, and let L be the loss values ​​for L. cls L obj L box The weights;

[0012] (3) Iteratively train the image defect detection network model:

[0013] The image defect detection network model O is iteratively trained using the training sample set R to obtain the trained image defect detection network model O*.

[0014] (4) Automatically acquire defect detection results during 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 level and location box information of the fiber optic gyroscope assembly process corresponding to each test sample. The location box information of the defects is automatically quantitatively analyzed. Then, the radius of curvature of the fiber winding path curve obtained by quantitative analysis, as well as the category, confidence level and location box information, are used as the final defect detection result of the fiber optic gyroscope assembly process, thus completing 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, this invention calculates the loss value of the bounding box information using the EIOU loss function. This allows for the separate calculation of the width and height differences between the predicted and true bounding box information. Furthermore, the loss value of the image defect detection network model is calculated using the category, confidence level, and the loss value of the bounding box information. This minimizes the difference in width and height between the predicted and true bounding box information, avoiding the low accuracy caused by directly calculating the aspect ratio difference between the predicted and true bounding box information using the CIOU loss function in existing technologies. This effectively improves the detection accuracy.

[0018] 2. In the process of obtaining defect detection results during the assembly of fiber optic gyroscopes, this invention provides more refined and specific detection results by automatically and quantitatively analyzing the defect location box information in the output of the trained image defect detection network model. This enables the evaluation of defects in the fiber optic gyroscope assembly process and avoids the shortcomings of existing technologies where the detection results are singular and difficult to quantify. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the present invention.

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

[0021] Figure 3 This is a diagram showing the effect of quantitative analysis of the test image and its defect location box information used in the embodiments of the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] Reference Figure 1 The present invention includes the following steps:

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

[0025] N original images containing various defect categories during the fiber optic gyroscope assembly process are preprocessed. The location bounding box information and category of defects in M ​​original images and their corresponding preprocessed M×6 images are labeled. The M original images and their corresponding preprocessed M×6 images and their labels are used to form a training sample set R. The remaining NM original images and their corresponding preprocessed (NM)×6 images are used to form a test sample set E, where N≥50, M≥N / 2, and in this embodiment N=60, M=45.

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

[0027] Each of the N fiber optic gyroscope assembly process images, magnified by an electron microscope and captured by an industrial camera at different angles and sizes, is subjected to rotational expansion, mirror expansion, cropping expansion, image brightness enhancement, image contrast enhancement, and image sharpness enhancement to obtain N×6 preprocessed fiber optic gyroscope assembly process images.

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

[0029] L t =λ1L cls +λ2L obj +λ3L box

[0030] Among them, L cls L obj L box Let λ1, λ2, and λ3 represent the loss values ​​for defect category information, confidence level information, and location box information, respectively, and let L be the loss values ​​for L. cls L obj L box The weights;

[0031] Loss value L based on category, confidence level, and bounding box information cls L obj L box The calculation formulas are as follows:

[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 T represents the true and predicted class information for each training sample, respectively. lo P lo Let P represent the ground truth and predicted bounding box information for each training sample, respectively. Let IoU(·) denote the intersection-union function. sc T represents lo With P lo The intersection and union ratio, P lo ∩T lo P lo ∪T lo T represents respectively lo With P lo The intersection and union of P, |·| represents the absolute value operation, P sc T sc w represents the true and predicted confidence levels for each training sample, respectively. c h c P lo With T lo The width and height of the minimum bounding rectangle C. P lo T lo width, P lo T lo The height, d(·) represents the Euclidean distance function, d 2 (P lo ,T lo ) represents 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] To overcome the limitations of existing CIOU (Complete Intersection over Union) loss on model detection accuracy, this invention introduces EIOU (Efficient Intersection over Union) loss. Based on CIOU loss, the aspect ratio loss is decomposed to calculate the width and height differences between the ground truth and predicted bounding boxes separately. This loss function comprises four parts: overlap loss, center distance loss, width loss, and height loss, minimizing the width and height differences between the ground truth and predicted bounding boxes. The calculation formula is as follows:

[0038]

[0039] Among them, 1-IoU(P lo ,T lo ) represents P lo With T lo The overlap loss, P represents lo With T lo The center distance loss, P lo With T lo The width and height loss.

[0040] This embodiment constructs an image defect detection network model including a cascaded Backbone network, Neck network, and Head network, such as... Figure 2 As shown; the Backbone network includes one Focus module, four Conv-CSP modules, and one 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; where:

[0041] The Backbone network consists of a Focus module containing four cascaded parallel slice layers, a stitching layer, and a Conv module; a Conv module containing a cascaded single convolutional layer, a single normalization layer, and a single nonlinear activation layer; a CSP module containing a cascaded Conv module, parallel Conv modules and residual modules, a stitching layer, and a Conv module; and an SPP module containing a cascaded Conv module, three parallel pooling layers, a stitching layer, and a Conv module.

[0042] The Neck network consists of an FPN feature pyramid network containing two Conv modules, two CSP modules, two upsampling layers, and two splicing layers; and a PAN path aggregation network containing two Conv modules, two CSP modules, and two splicing layers. Each Conv module includes a cascaded single convolutional layer, a single normalization layer, and a single nonlinear activation layer. Each CSP module includes a cascaded Conv module, parallel Conv modules and Conv-Conv modules, a splicing layer, and a Conv module.

[0043] Step 3) Iteratively train the image defect detection network model, which is implemented 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 in the t-th iteration. t The weights and bias parameters in the equation are w. t b t And let t = 1, in this embodiment T = 300.

[0045] Step 3b) The Backbone network extracts features for 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 for F N The detection is performed to obtain the bounding box information P for 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 performs convolution operations on the compressed results. It extracts features from the small-sized feature maps of each training sample, then compresses and convolves these extracted feature maps, and finally concatenates the high-level feature maps obtained from compression and convolution with the low-level feature maps obtained from convolution. This concatenation is followed by multiple pooling operations, and finally, all pooling results are concatenated to obtain the feature map F for each training sample. B ;

[0047] The FPN feature pyramid network and the Backbone network form a bottom-up side connection, which affects the feature map F extracted by the Backbone network. B Feature extraction is performed, and the extracted feature map is upsampled. The upsampled feature map is then concatenated with the feature map extracted by the Backbone network. Finally, feature extraction is performed on the concatenated feature map.

[0048] The PAN path aggregation network and the FPN feature pyramid network form a top-down side connection. Feature extraction is performed on the feature map extracted by the FPN feature pyramid network, and the extracted feature map is concatenated 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 ;

[0049] The Head network and the PAN path aggregation network form side connections for the fused feature map F N Extract the detection results to obtain the bounding box information P for each training sample. lo Confidence level P sc and category P la .

[0050] Step 3c) employs a binary classification cross-entropy loss function and an EIOU loss function, and uses the true location box information T of each training sample. lo Confidence level T sc and category T la and its corresponding predicted location 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 The detection network model O is updated to obtain the detection network model for this iteration. t The update formulas are as follows:

[0051]

[0052] in, L respectively t For the weight parameter ω t Bias parameter b t The partial derivative of w, where α represents the learning rate, w t '、b t ' represents ω t b t The updated results.

[0053] Step 3d) Determine if t≥T holds true. If yes, obtain the trained fiber optic gyroscope assembly process image defect detection network model O*. Otherwise, let t=t+1 and O=O. t And perform step 3b).

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

[0055] The test sample set E is used as input to the trained image defect detection network model O* for forward propagation to obtain the defect category, confidence level, and location box information of the fiber optic gyroscope assembly process for each test sample. The location box information of the defects is automatically quantitatively analyzed. Then, the radius of curvature of the fiber winding path curve obtained from the quantitative analysis, along with the category, confidence level, and location box information, are used as the final defect detection result of the fiber optic gyroscope assembly process, thus completing the automatic acquisition of defect detection results. The steps are as follows:

[0056] Step 4a1) Obtain the defect location box information P for the fiber optic gyroscope assembly process corresponding to each test sample. lo After grayscale processing, binarization is performed, and the average value method is selected for P. lo Grayscale processing is performed to obtain a grayscale image that removes color interference and reflects shape and texture features. 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 values ​​in the grayscale image less than the threshold and 255 to values ​​greater than the threshold. The calculation formula is:

[0057]

[0058] Where β represents the 80th percentile of the grayscale pixel value, value b The value represents the pixel value in a binary image. g This represents the pixel values ​​in a grayscale image. After binarization, a binary image B is obtained. lo ;

[0059] Step 4a2) Process the resulting binary image B lo After clustering, noise reduction is performed using the density-based DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering method on the binary image B. lo Pixels with a median value of 255 are clustered. After clustering using the DBSCAN algorithm, the binary image B... lo The pixels in the image are divided into different categories based on their density. Finally, defective pixels are preserved to achieve the denoising effect, resulting in the denoised binary image B. l ' o ;

[0060] Step 4b) Analyze the denoised binary image B l ' o Fit the defective pixels in the image to the denoised binary image B. l ' oBy scanning along the x-direction and saving the midpoint trajectory, the approximate coordinates of the fiber optic winding path can be obtained. A quadratic polynomial is then used to fit the fiber optic winding path to obtain the curve equation of the fiber optic winding path:

[0061] y = a2x 2 +a1x+a0

[0062] Where x represents the coordinates of the fiber winding path curve, 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] ②Let the equation of the curve be y = a²x 2 If a + a1x + a0, then for the known data points, we have:

[0065]

[0066] ③Let For A, For T, Let χ be the polynomial coefficients. The formula for calculating χ is:

[0067]

[0068] Based on the fitted fiber winding path curve equation, the radius of curvature of the fiber winding path curve is calculated using the following formula:

[0069]

[0070] Combining the fiber winding path curve equation y=a2x 2 From +a1x+a0, we can obtain the radius of curvature ρ of the fiber winding path curve as:

[0071]

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

[0073] The effect of quantitative analysis of the images and defect location bounding box information in the test samples in this embodiment is as follows: Figure 3 As shown, where Figure 3 (a) represents the original image of the test sample with defect location box information, where the dashed rectangle represents the defect location box information; Figure 3 (b) represents the binary image B obtained by binarizing the grayscale information of the defect location box. lo , Figure 3 (c) represents the binary image B obtained after denoising following 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) represents the calculated radii of curvature of the fiber winding path curve at the start, midpoint, and end points, marked in the defect location box information of the original image of the test sample. From Figure 3 It can be seen that automatic quantitative analysis of defect location information can obtain a detailed fiber optic winding path curve and the radius of curvature of the fiber optic winding path curve, providing detailed and specific detection result information.

[0074] The technical effects of the present invention will be further explained below with reference to simulation results:

[0075] 1. Experimental conditions and contents:

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

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

[0078] 2. Analysis of experimental results:

[0079] 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 correctly detected and undetected defects, respectively, FP represents the number of non-defects that were incorrectly detected, A represents the total number of defect categories, and AP a This indicates the defect detection accuracy for category a.

[0082] Table 1

[0083]

[0084] As can be seen from Table 1, compared with the existing fiber optic winding defect detection method based on the improved YOLO algorithm, the present invention effectively improves the average detection accuracy and can be used for subsequent practical applications.

[0085] In summary, this invention can utilize deep image learning algorithms to perform defect detection and quantitative analysis on input fiber optic gyroscope assembly process images, displaying the radius of curvature of the fiber winding path curve in the fiber optic gyroscope assembly process images, as well as defect type, confidence level, and location box information. This invention is rigorous, complete, and has high detection accuracy, providing detailed and specific detection result information.

Claims

1. An automatic image defect detection method for fiber optic gyroscope assembly process based on quantitative analysis, characterized in that, The steps include: (1) Obtain the training sample set and the test sample set: N original images containing various defect categories during the fiber optic gyroscope assembly process are preprocessed. The location bounding box information and category of defects in M ​​original images and their corresponding preprocessed M×6 images are labeled. The M original images and their corresponding preprocessed M×6 images and their labels are combined to form a training sample set R. The remaining NM original images and their corresponding preprocessed (NM)×6 images are combined to form a test sample set E, where N≥50 and M≥N / 2. (2) Define the loss function of the network model O for image defect detection during the fiber optic gyroscope assembly process: L t =λ1L cls +λ2L obj +λ3L box Among them, L cls L obj L represents the loss value calculated using the binary cross-entropy loss function, taking into account the true class, confidence level, and corresponding predicted class of each training sample, as well as the defect category information and confidence information. box This represents the loss value of the defect's bounding box information calculated using the EIOU loss function and the bounding box information of each training sample and its corresponding predicted bounding box information. λ1, λ2, and λ3 represent L, respectively. cls L obj L box The weights; (3) Iteratively train the image defect detection network model: The image defect detection network model O is iteratively trained using the training sample set R to obtain the trained image defect detection network model O*. (4) Automatically acquire defect detection results during 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 level and location box information of the fiber optic gyroscope assembly process corresponding to each test sample. The location box information of the defects is automatically quantitatively analyzed. Then, the radius of curvature of the fiber winding path curve obtained by quantitative analysis, as well as the category, confidence level and location box information, are used as the final defect detection result of the fiber optic gyroscope assembly process, thus completing 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, during the fiber optic gyroscope assembly process described in step (1) is implemented as follows: Each of the N fiber optic gyroscope assembly process images, magnified by an electron microscope and captured by an industrial camera at different angles and sizes, is subjected to rotational expansion, mirror expansion, cropping expansion, image brightness enhancement, image contrast enhancement, and image sharpness enhancement 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, Neck network, and Head network, wherein: The Backbone network consists of cascaded Focus modules, multiple Conv-CSP modules, and SPP modules; the Neck network consists of cascaded FPN feature pyramid network and PAN path aggregation network structure; the Head network consists of 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 mentioned 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 T represents the true and predicted class information for each training sample, respectively. lo P lo Let P represent the ground truth and predicted bounding box information for each training sample, respectively. Let IoU(·) denote the intersection-union function. sc T represents lo With P lo The intersection and union ratio, P lo ∩T lo P lo ∪T lo T represents respectively lo With P lo The intersection and union of P, |·| represents the absolute value operation, P sc T sc w represents the true and predicted confidence levels for each training sample, respectively. c h c P lo With T lo The width and height of the minimum bounding rectangle C. P lo T lo width, P lo T lo The height, d(·) represents the Euclidean distance function, d 2 (P lo ,T lo ) represents 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 t-th iteration. t The weights and bias parameters in the equation are w. t b t And let t = 1; (3b) The Backbone network extracts features for 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 for F N The detection is performed to obtain the bounding box information P for each training sample. lo Confidence level P sc and category P la ; (3c) The binary classification cross-entropy loss function and the EIOU loss function are used, and the true location box information T of each training sample is used. lo Confidence level T sc and category T la and its corresponding predicted location 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 The detection network model O is updated to obtain the detection network model for this iteration. t ; (3d) Determine whether t≥T holds true. If yes, obtain the trained fiber optic gyroscope assembly process image defect detection network model O*. Otherwise, let t=t+1, O=O t And perform 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 steps are as follows: The Backbone network compresses each training sample multiple times, then concatenates and performs convolution operations on the compressed results. It extracts features from the small-sized feature maps of each training sample, then compresses and convolves these extracted feature maps, and finally concatenates the high-level feature maps obtained from compression and convolution with the low-level feature maps obtained from convolution. This concatenation is followed by multiple pooling operations, and finally, all pooling results are concatenated to obtain the feature map F for each training sample. B .

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

8. The method according to claim 5, characterized in that, The step (3c) described for w t b t The update is performed using the following formulas: in, L respectively t For the weight parameter ω t Bias parameter b t The partial derivative of w, where α represents the learning rate, w t '、b t ' represents 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 box information described in step (4) is achieved through the following steps: (4a) Defect location box information P for the fiber optic gyroscope assembly process corresponding to each test sample lo After grayscale processing, binarization is performed, and the resulting binary image B is then processed. lo After clustering, denoising is performed to obtain the denoised binary image B'. lo ; (4b) The denoised binary image B' lo After fitting the defective pixels in the data, the radius of curvature ρ of the fitted fiber winding path curve is calculated and used as the defect location bounding box information P. lo The quantitative analysis results.

Images