Automobile safety airbag defect detection method based on improved YOLOv8

Abstract

The invention belongs to the technical field of air bag defect detection, and particularly discloses an improved YOLOv8-based automobile air bag defect detection method, which comprises the following steps of: constructing a data set for different defect types of an automobile air bag, and performing enhancement processing on the data set to obtain an expanded data set; dividing the expanded data set into a training set, a test set and a verification set, inputting the data set into an improved YOLOv8 model for training to obtain an airbag defect detection model, inputting an image of a defect to be identified into the airbag defect detection model, outputting a defect detection result, and evaluating the detection result. The problems that in the traditional manual detection process, workers are prone to fatigue, so that the detection efficiency is reduced, and the quality of detected safety airbags is poor are solved. According to the invention, improvement is carried out through the YOLOv8 algorithm, the feature fusion capability is improved, the operation burden is reduced, and thus the detection efficiency and detection precision of the air bag restraint system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of airbag defect detection, and in particular to an automobile airbag defect detection method based on improved YOLOv8. Background Art

[0002] Airbags are an important part of the passive safety system of automobiles. When a car collides, they can be inflated quickly to provide cushioning protection for the driver and passengers and reduce the risk of injury. The airbag system usually consists of sensors, control units and airbag components. The sensors can detect the intensity and direction of the vehicle collision and transmit the signal to the control unit. The control unit determines whether to trigger the airbag based on the preset algorithm and issues a command to inflate the airbag quickly when necessary.

[0003] As competition in the automotive market becomes increasingly fierce, consumers are increasingly demanding on vehicle safety performance. Automakers need to continuously improve the quality and reliability of airbags to meet market demand and regulatory requirements. In the traditional airbag production process, quality inspection mainly relies on manual visual inspection and sampling testing.

[0004] However, during the traditional manual inspection process, workers are limited by their physiological conditions and are prone to fatigue. The inspection efficiency will be reduced after working for a long time, which will lead to poor quality of the inspected airbags. Summary of the invention

[0005] The purpose of the present invention is to provide a method for automobile airbag defect detection based on improved YOLOv8, so as to solve the problem that in the traditional manual inspection process, workers are prone to fatigue, thereby reducing the inspection efficiency and further leading to poor quality of the inspected airbag.

[0006] To achieve the above object, the basic solution provided by the present invention is: a method for detecting defects in automobile airbags based on improved YOLOv8, comprising the following steps:

[0007] S1. Construct a data set for different defect types of automobile airbags, and enhance the data set to obtain an expanded data set;

[0008] S2, divide the expanded data set into training set, test set and validation set. Select 780 images from the expanded data set as the test set, 820 images as the validation set, and the remaining 4200 images as the training set;

[0009] S3, input the automobile airbag data set into the improved YOLOv8 model for training to obtain an airbag defect detection model;

[0010] S4. Input the airbag image with the defect to be identified into the airbag defect detection model, output the defect detection result, and evaluate the detection result.

[0011] The beneficial effects of the present invention are as follows: by adopting the improved YOLOv8 model, the influence of manual fatigue on detection efficiency is avoided, and the model can continuously and efficiently process a large number of airbag images without considering the working time problem, thereby greatly improving the detection speed and adapting to the fast-paced needs of large-scale production in the modern automobile industry; in terms of detection accuracy, by improving the YOLOv8 algorithm, the feature fusion capability is improved, especially the detection performance of small targets is enhanced, thereby significantly improving the detection accuracy, and providing an efficient, accurate and reliable technical means.

[0012] Solution 2 is a preferred option of the basic solution. In S1, the constructed data set is expanded by flipping, scaling, cropping, Gaussian noise and brightness adjustment. Since it is difficult to obtain airbag image data and mark defects, and the sample source is insufficient, the data set is expanded by flipping, scaling, cropping, Gaussian noise and brightness adjustment.

[0013] Scheme 3 is the preferred basic scheme. In S3, the automobile airbag image is used as input, the Backbone network extracts features, the Neck network fuses the extracted features, and finally outputs them from the Head layer to train the improved YOLOv8 model.

[0014] Solution 4 is the preferred solution of the basic solution. In S3, the improved YOLOv8 model is based on the traditional YOLOv8 model. In the Backbone layer, C3 in the original model is optimized, and in the Neck layer, a weighted bidirectional feature pyramid network (Bidirectional Feature Pyramid Network, BiFPN) is introduced, and the model detection head loss function is adjusted.

[0015] Solution 5 is the preferred solution of Solution 4. In the Backbone layer, the C3 module in the original model is replaced by the CESE_C2f module. The CESE_C2f module uses Pw_Conv (Pointwise Convolution) for dimensionality increase and Dw_Conv (Depthwise Convolution) to implement deep separable convolution. The convolution layer processes the input feature map to halve the number of channels of the output feature map. The Split operation is used to split the features, and after multiple layers of CESE (Chain-like Efficient Squeeze Excitation) are stacked, the features processed by CESE are concatenated with the previous gradient flow. Compared with the C3 module, the CESE_C2f module not only increases the gradient flow, but also reduces the amount of calculation, thereby accelerating the convergence of the model.

[0016] Solution 6 is the preferred solution of Solution 4. In the Neck layer, a weighted bidirectional feature pyramid network (Bidirectional Feature Pyramid Network, BiFPN) is used to build a new Neck network structure, fully reuse the feature information extracted by the Backbone part, and introduce learnable weights to determine the importance of different input features. BiFPN realizes cross-scale feature fusion through the following fusion formula:

[0017]

[0018] Among them, Input i is the feature image of the fused input, Output is the output fusion result, w i The weight of the fusion part is input, and the ReLU activation function is used to ensure that w i ≥0; the value is 0.0001; by introducing the weighted bidirectional feature pyramid network, the linear layout of the original network is optimized, thereby improving the effect of feature fusion.

[0019] Solution 7: This is the preferred solution of Solution 4. The loss function of the model detection head is adjusted. The SIoU loss function is selected to replace the loss function in the original model. The following parts:

[0020] (1) Angle cost, defined as follows:

[0021]

[0022] Where: c h It is expressed as the height difference between the center point of the real box and the predicted box, σ is the distance between the center point of the real box and the predicted box, is equal to the angle α, and in:

[0023]

[0024] is the center coordinate of the real frame, ( b cy ) is the center coordinate of the prediction box, when α is Or 0, the angle loss is 0. During the training process, if Then minimize α, otherwise minimize β.

[0025] (2) Distance cost, defined as follows:

[0026]

[0027] in:

[0028]

[0029] Υ=2-Λ

[0030] b cx = x x coordinates of the true and predicted bounding boxes, b cy =y coordinates of the actual and predicted bounding boxes, c w 、c h The width and height of the minimum bounding rectangle of the real box and the predicted box;

[0031] (3) Shape cost, defined as follows:

[0032]

[0033] in:

[0034]

[0035] (w,h) and (w gt ,h gt ) are the width and height of the predicted box and the real box respectively, θ controls the degree of attention to shape loss, and θ is close to 4;

[0036] (4) The final SIoU loss function is defined as follows:

[0037]

[0038] Where: B∩B GT represents the intersection of the real box and the predicted box, and B∪B GTRepresents the intersection of the real box and the predicted box; SIou is calculated based on the ratio of the intersection area to the union area of ​​the two bounding boxes. It has a fast calculation speed and is suitable for scenarios with high requirements on calculation speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a diagram of a stylus bending defect in the present invention;

[0040] Figure 2 It is a buckle breakage defect diagram in the present invention;

[0041] Figure 3 It is a diagram of misalignment defects of the spring in the present invention;

[0042] Figure 4 This is an intuitive diagram of the angle loss of the SIoU loss function in the present invention;

[0043] Figure 5 It is an intuitive diagram of the distance loss of the SIoU loss function in the present invention;

[0044] Figure 6 It is an intuitive diagram of the IoU loss of the SIoU loss function in the present invention;

[0045] Figure 7 It is a schematic diagram of the CESE_C2f module structure of Yolov8 in the present invention;

[0046] Figures 8 to 11 This is a diagram showing the improvement of the novel Neck layer structure in the present invention;

[0047] Fig.12 It is a comparison chart of the detection accuracy of three airbag defects in the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described in detail below through specific embodiments:

[0049] like Figures 1 to 12 As shown: A method for detecting defects in automobile airbags based on improved YOLOv8, comprising the following steps:

[0050] S1. Construct a data set for different defect types of automobile airbags, and enhance the data set by flipping, scaling, cropping, Gaussian noise and brightness adjustment to obtain an expanded data set;

[0051] S2, divide the expanded data set into training set, test set and validation set. Select 780 images from the expanded data set as the test set, 820 images as the validation set, and the remaining 4200 images as the training set;

[0052] S3. Input the car airbag dataset into the improved YOLOv8 model for training. Take the car airbag image as input, and use the Backbone network for feature extraction. In the Backbone layer, replace the C3 module in the original model with the CESE_C2f module. The CESE_C2f module uses Pw_Conv (Pointwise Convolution) for dimensionality increase, and uses Dw_Conv (Depthwise Convolution) to implement deep separable convolution. The convolution layer processes the input feature map to halve the number of channels of the output feature map. The Split operation is used to split the features, and after multiple layers of CESE (Chain-like Efficient Squeeze Excitation) are stacked, the features processed by CESE are spliced ​​with the previous gradient flow. The Neck network performs feature fusion on the extracted features. In the Neck layer, a weighted bidirectional feature pyramid network (Bidirectional Feature Pyramid) is introduced. A new Neck network structure is constructed to fully reuse the feature information extracted by the Backbone part. BiFPN introduces learnable weights to determine the importance of different input features. BiFPN realizes cross-scale feature fusion through the following fusion formula:

[0053]

[0054] Among them, Input i is the feature image of the fused input, Output is the output fusion result, w i The weight of the fusion part is input, and the ReLU activation function is used to ensure that w i ≥0; the value is 0.0001.

[0055] The adjusted model detection head loss function chooses to use the SIoU loss function to replace the loss function in the original model, as follows:

[0056] (1) Angle cost, defined as follows:

[0057]

[0058] Where: c h It is expressed as the height difference between the center point of the real box and the predicted box, σ is the distance between the center point of the real box and the predicted box, is equal to the angle α, and in:

[0059]

[0060] is the center coordinate of the real frame, ( b cy ) is the center coordinate of the prediction box, when α is Or 0, the angle loss is 0. During the training process, if Then minimize α, otherwise minimize β.

[0061] (2) Distance cost, defined as follows:

[0062]

[0063] in:

[0064]

[0065] Υ=2-Λ

[0066] b cx = x x coordinates of the true and predicted bounding boxes, b cy =y coordinates of the actual and predicted bounding boxes, c w 、c h The width and height of the minimum bounding rectangle of the real box and the predicted box;

[0067] (3) Shape cost, defined as follows:

[0068]

[0069] in:

[0070]

[0071] (w,h) and (w gt ,h gt ) are the width and height of the predicted box and the real box respectively, θ controls the degree of attention to shape loss, and θ is close to 4;

[0072] (4) The final SIoU loss function is defined as follows:

[0073]

[0074] Where: B∩B GT represents the intersection of the real box and the predicted box, and B∪B GT It represents the intersection of the real box and the predicted box, and finally outputs it from the Head layer to train the improved YOLOv8 model, and finally obtains the airbag defect detection model;

[0075] S4. Input the airbag image with defects to be identified into the airbag defect detection model, output the defect detection result, and use the mAP value of the detection result as an evaluation.

[0076] After analyzing and comparing the experimental results with the YOLOv8 model, the improved model was analyzed according to various performance indicators. The detection accuracy values ​​AP (Average Precision) of three types of airbag defects, including stylus bending, spring assembly dislocation, and bayonet damage, were compared between the improved and unimproved models. After the model was improved, the accuracy of stylus bending was improved from 0.988 to 0.991, and the accuracy of spring misalignment defects was improved from 0.967 to 0.991. The highest improvement of a single defect category was 0.024. The performance comparison of the other models before and after improvement is shown in Table 1.

[0077] Table 1 Comparison of model improvement experimental results

[0078] Model Name mAP50 mAP50:95 P(％) R(％) Params(M) GFLOPs YOLOv8s 97.4 67 97 942 11.12 28.5 YOLOv8s+A 98.1 67.3 98 95.4 11.8 25.8 YOLOv8s+A+B 98.3 67.3 98 96 8.04 22.0

[0079] In Table 1, YOLOv8s+A and YOLOv8s+A+B represent the network structures after the improvement of the CESE_C2f module and the Neck module, respectively.

[0080] As can be seen from Table 1, after adopting improvement A, mAP50, mAP50:95, accuracy and regression rate are all improved, indicating that this improvement can improve accuracy; at the same time, the number of model parameters and the amount of calculation are reduced, and the real-time detection performance of the model is improved. After continuing to adopt improvement B, it is found that the mAP50 accuracy is further improved, and the model calculation pressure is further reduced. The operating economy of the model is improved, which is more in line with the assembly line production requirements of airbags.

[0081] Specifically, the mean average precision value mAP value after improvement increased from 0.974 to 0.983, an increase of 0.009, and the number of parameters decreased by 3.08M, from 11.12M before improvement to 8.04M, which is a significant decrease compared with the previous step; the amount of calculation decreased from the original 28.5G to the optimized 22.0G, a reduction of 6.5G, and the calculation pressure was effectively reduced, indicating that the improvement has a significant effect on reducing the number of parameters and calculations and improving the model detection effect. It can be seen that compared with the original YOLOv8s algorithm, the improved model is more in line with the requirements of the task of airbag defect detection.

[0082] The above is only an embodiment of the present invention, and the common knowledge such as the known specific structure and characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for detecting automobile airbag defects based on improved YOLOv8, characterized in that: The following steps are involved: S1. Construct a data set for different defect types of automobile airbags, and enhance the data set to obtain an expanded data set; S2, divide the expanded data set into training set, test set and validation set. Select 780 images from the expanded data set as the test set, 820 images as the validation set, and the remaining 4200 images as the training set; S3, input the automobile airbag data set into the improved YOLOv8 model for training to obtain an airbag defect detection model; S4. Input the airbag image with the defect to be identified into the airbag defect detection model, output the defect detection result, and evaluate the detection result.

2. According to claim 1, a method for detecting defects in automobile airbags based on improved YOLOv8 is characterized in that: In S1, the constructed data set is obtained by flipping, scaling, cropping, Gaussian noise and brightness adjustment to obtain an expanded data set.

3. The automobile airbag defect detection method based on improved YOLOv8 according to claim 1 is characterized in that: In S3, the car airbag image is used as input, the Backbone network extracts features, the Neck network fuses the extracted features, and finally outputs them from the Head layer to train the improved YOLOv8 model.

4. The automobile airbag defect detection method based on improved YOLOv8 according to claim 1 is characterized in that: In S3, the improved YOLOv8 model is based on the traditional YOLOv8 model. In the Backbone layer, C3 in the original model is optimized, and in the Neck layer, a weighted bidirectional feature pyramid network (BidirectionalFeaturePyramidNetwork, BiFPN) is introduced, and the model detection head loss function is adjusted.

5. The automobile airbag defect detection method based on improved YOLOv8 according to claim 4 is characterized in that: In the Backbone layer, the C3 module in the original model is replaced by the CESE_C2f module. The CESE_C2f module uses Pw_Conv (Pointwise Convolution) for dimensionality increase and Dw_Conv (Depthwise Convolution) to implement deep separable convolution. The convolution layer processes the input feature map to halve the number of channels of the output feature map. The Split operation is used to split the features, and after multiple layers of CESE (Chain-like Efficient Squeeze Excitation) are stacked, the features processed by CESE are spliced ​​with the previous gradient flow.

6. The automobile airbag defect detection method based on improved YOLOv8 according to claim 4 is characterized in that: In the Neck layer, a weighted bidirectional feature pyramid network (Bidirectional Feature Pyramid Network, BiFPN) is used to build a new Neck network structure, which fully reuses the feature information extracted by the Backbone part. BiFPN introduces learnable weights to determine the importance of different input features. BiFPN realizes cross-scale feature fusion through the following fusion formula: Among them, Input i is the feature image of the fused input, Output is the output fusion result, w i The weight of the fusion part is input, and the ReLU activation function is used to ensure that w i ≥0; the value is 0.0001.

7. The automobile airbag defect detection method based on improved YOLOv8 according to claim 4 is characterized in that: The adjusted model detection head loss function chooses to use the SIoU loss function to replace the loss function in the original model, as follows: (1) Angle cost, defined as follows: Where: c h It is expressed as the height difference between the center point of the real box and the predicted box, σ is the distance between the center point of the real box and the predicted box, is equal to the angle α, and in: is the center coordinate of the real frame, is the center coordinate of the prediction box, when α is Or 0, the angle loss is 0. During the training process, if Then minimize α, otherwise minimize β. (2) Distance cost, defined as follows: in: Υ=2-Λ b cx = x coordinates of the true and predicted bounding boxes, b cy =y coordinates of the actual and predicted bounding boxes, c w 、c h The width and height of the minimum bounding rectangle of the real box and the predicted box; (3) Shape cost, defined as follows: in: (w,h) and (w gt ,h gt ) are the width and height of the predicted box and the real box respectively, θ controls the degree of attention to shape loss, and θ is close to 4; (4) The final SIoU loss function is defined as follows: Where: B∩B GT represents the intersection of the real box and the predicted box, and B∪B GT Represents the intersection of the true box and the predicted box.

Images