CN114627106A - Weld defect detection method based on Cascade Mask R-CNN model - Google Patents

Abstract

A welding seam defect detection method based on the Cascade Mask R-CNN model, firstly obtains digital images of weld seam ray inspection including five kinds of defects, including round defects, strip defects, cracks, incomplete penetration and incomplete fusion, and their corresponding annotation files, And divided into training set, validation set and test set; then use the training set after image preprocessing to train the weld defect detection model built and optimized based on the Cascade Mask R‑CNN model, and use the validation set to obtain the optimal weight file The test set is then tested to evaluate the performance of the weld defect detection model. The invention improves the problem of low accuracy of the target detection model based on a single threshold value and the overfitting problem caused by directly increasing the threshold value, and improves the detection accuracy of welding seam defects.

Description

A Weld Defect Detection Method Based on Cascade Mask R-CNN Model

technical field

The invention relates to the field of target detection, in particular to a welding seam defect detection method based on a Cascade Mask R-CNN model.

Background technique

Welding is one of the most important technologies in the industrial production and manufacturing process. In order to ensure the welding quality, it is necessary to carry out non-destructive testing of welds. Among them, radiographic testing is one of the commonly used non-destructive testing techniques. At present, the characterization and positioning of defects in radiographic inspection mainly rely on manual evaluation, that is, manual evaluation of films. Manual film evaluation is greatly affected by subjective factors such as the professional level of the film reviewers and their own conditions, and the efficiency is low, which cannot meet the needs of automated and intelligent detection in modern industries.

In recent years, the development of deep learning has not only broken through many difficult visual problems, improved the level of image cognition, but also accelerated the progress of related technologies in the field of target detection. The automatic learning method of image features in deep learning has been applied to industrial Product weld defect detection has become the mainstream research direction. The target detection model based on deep learning is mainly divided into two categories: one-stage and two-stage. The one-stage target detection model uses a convolutional neural network to directly predict the category and position of the target, and realizes classification and regression in one step. It is characterized by fast speed but low accuracy; the two-stage target detection model first generates suggested candidate boxes that may contain targets, and then performs classification and regression through the prediction network. This type of model is characterized by slow speed and high accuracy. In addition, when calculating the intersection-over-union ratio (IoU) between the proposed candidate frame and the ground-truth frame, the common target detection model divides the proposed candidate frame into positive and negative by comparing the intersection-over-union ratio (IoU) and a single set threshold. Samples, usually positive samples are much larger than negative samples. In the testing phase, the positive and negative samples will be sampled to make the ratio between the two meet a certain ratio. However, in the testing phase, due to the lack of comparison of the real labeled boxes, the quality of the proposed candidate frame is low. , the detection accuracy is low; therefore, a model using a cascade structure is needed to break through the bottleneck of weld defect detection accuracy under a single threshold, so that the model can be closer to the real position of the weld defect after each regression, adapting to different The distribution of proposed candidate boxes.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a welding seam defect detection method based on the Cascade Mask R-CNN model, so as to solve the aforementioned problems existing in the prior art.

In order to achieve the above object, the technical scheme adopted in the present invention is as follows:

A welding seam defect detection method based on Cascade Mask R-CNN model, comprising the following steps:

S1. Acquire a digital image of the weld seam radiographic inspection containing five kinds of defects, such as round defect, strip defect, crack, lack of penetration, and lack of fusion, and its corresponding annotation file; The file is divided into training set, validation set and test set;

S2. Perform image preprocessing on the digital image of the weld seam radiographic inspection to obtain an enhanced and unified image;

S3. Build and optimize a weld defect detection model based on the Cascade Mask R-CNN model. The weld defect detection model includes a convolutional neural network for feature extraction and a prediction network for classification and regression;

S4. Use the training set to train the weld defect detection model, and use the verification set to verify the weight file obtained in each round of training, and obtain the weight file with the best performance on the verification set; including the following steps :

S41. Perform feature extraction on the digital image of the weld ray detection through the convolutional neural network to form a feature map;

S42, using the multi-scale detection algorithm FPN to improve the RPN to generate a suggested candidate frame for the feature map;

S43, the proposed candidate frame is mapped to the feature map to obtain a corresponding feature matrix, and the feature matrix is uniformly scaled to a specified size by ROI Align;

S44, passing the proposed candidate frame through a three-stage cascade detector composed of a target classifier and a bounding box regressor to obtain corresponding defect categories and bounding box regression parameters under three set thresholds; And filter out low-probability targets to get the final detection result;

S45. Set the training parameters of the weld defect detection model, use the training set to train the weld defect detection model through the steps S41-S44, and use the validation set to perform training on the weld defect detection model obtained in each round of training. The weight file is verified, and the weight file that performs best on the verification set is obtained;

S5. Use the test set to test the weight file obtained in step S45 with the best performance on the verification set, so as to evaluate the performance of the weld defect detection model.

Preferably, the method for image preprocessing in step S2 includes image enhancement and image denoising, and the image enhancement uses AHE algorithm to perform image enhancement on the digital image of the weld seam ray detection, and the digital image of the weld seam ray detection is image enhanced. The details are sharpened and the defect features are highlighted. The calculation formula of the image enhancement is as follows:

In the above formula: y _i,j represents the center pixel before transformation, Y _i,j represents the center pixel after transformation, m _i,j represents the gray mean value of the local area with y _i,j as the center point, T represents the pair The cumulative distribution transformation function of the point, k represents the adaptive function, which is calculated from the pixel characteristics of the local area;

In the image denoising, the DMB algorithm is used to denoise the digital image of the weld seam ray inspection, so as to reduce the image noise and preserve and enhance the defect features.

Preferably, the convolutional neural network in step S41 is ResNeXt-101, the convolutional neural network includes a convolutional layer, a pooling layer and an activation layer, and the convolutional layer extracts features from the input image to generate Feature map; the pooling layer is used to remove redundant information, reduce the amount of parameters, and expand the receptive field; the activation layer increases the nonlinearity of the output, and uses the activation function to convolve the results of the output layer to obtain a nonlinear mapping.

Preferably, in step S42, using FPN to improve RPN to generate suggested candidate frames for the feature map, including the following steps:

S421: Feed forward a part of ResNeXt-101 in the forward process of the convolutional neural network, and denote the output of the last residual block of each stage of ResNeXt-101 as {C1, C2, C3, C4, C5}, the first is In the bottom-up process, downsampling with a set step size is used for each level upwards, and the layers that do not change the size of the feature map are classified as a stage to form a feature pyramid;

S422: Upsampling is used from top to bottom to enlarge the small feature map of the top layer to the same size as the feature map of the previous stage;

S423: The horizontal connection fuses the upsampling result and the feature map of the same size generated from the bottom to the top, and convolves each fusion result through a 3×3 convolution kernel to obtain the final feature layer P={P2, P3,P4,P5}.

Preferably, the step S421 further includes: introducing an efficient attention module into {C3, C4, C5} of ResNeXt-101, firstly using a 1×1 convolution W _k and softmax to obtain the attention weight, and using the attention pool The global context features are obtained through a 1×1 convolution W _v1 and layer normalization, activated by the ReLU function, and after a 1×1 convolution W _v2 , the importance of each channel is obtained, and finally the use of The addition aggregates the global context features to the features of each position to form long-distance dependencies; the attention module calculation formula is as follows:

表示全局注意力池化的权重，LN代表层归一化，W_v2ReLU(LN(W_v1(·)))表示计算每个通道的重要程度。In the above formula: z _i represents the input of the attention module, Z _i represents the output of the attention module, N _p represents the number of positions in the feature map,

represents the weight of global attention pooling, LN represents layer normalization, and W _v2 ReLU(LN(W _v1 ( ))) represents the importance of calculating each channel.

Preferably, after the step S423, the method further includes: using the K-means clustering algorithm to count the area and the aspect ratio of the rectangular annotation frame in the annotation file, and set five types of areas {32 ² , 64 ² , 128 ² , 256 ² , 512 The anchors of ² } correspond to the five feature layers of {P2, P3, P4, P5, P6}, of which the P6 feature layer is obtained by downsampling the P5 feature layer, and the anchor of each area is set to seven aspect ratios {1: 10, 1: 5, 1: 2, 1: 1, 2: 1, 5: 1, 10: 1}, the generated anchor slides and traverses on the feature layer, and then generates the proposed candidate frame.

Preferably, the bounding box regressor in the three-stage cascade detector in step S44 is defined as a cascade regression problem, and the cascade regression changes the distribution of samples to be processed in different stages through resampling, and the bounding box regression The device is defined as follows:

In the above formula: x represents the sub-image block, b represents the sample distribution, f represents the bounding box regressor, the thresholds set by f ₁ , _{f 2} , and _{f 3} are 0.4, 0.5, and 0.6, respectively. Input, the output of f ₂ is used as the input of f ₃ , {f ₁ , f ₂ , f ₃ } is optimized for the resampling distribution of different stages, and acts on the training and testing stages at the same time;

The total loss function L that defines the training model includes two parts: bounding box regression loss and target classification loss. The calculation formula of the total loss function is as follows:

L(x ^t ,g)=L _cls (h _t (x ^t ),X ^t )+μ[X ^t ≥1]L _loc (f _t (x ^t ,b ^t ),g)

In the above formula: L _cls represents the loss function of target classification, L _loc represents the loss function of bounding box regression, {b ^t } represents the sample distribution of different training stages t and has b ^t =f _t-1 (x ^t-1 , b ^t-1 ), h _t represents the target classifier, f _t represents the bounding box regressor, g represents the real target box parameter corresponding to x ^t , μ represents the compromise coefficient, X ^t represents the label corresponding to x ^t , [ ] Represents an indicator function.

Preferably, the training parameter settings in step S45 include learning rate, momentum, weight decay, the value of batch_size and the total number of training rounds, and the ImageNet dataset training weight is used as the pre-training weight of the weld defect detection model, so The size of the digital image of the welding seam ray detection is uniformly scaled before entering the welding seam defect detection model, and the parameters of the welding seam defect detection model are updated by using the stochastic gradient descent method.

The beneficial effects of the present invention are as follows: the present invention discloses a welding seam defect detection method based on the Cascade Mask R-CNN model. By introducing a three-level cascade structure, the problem of low accuracy of the target detection model based on a single threshold and the direct The over-fitting problem caused by increasing the threshold; the RPN is improved by the multi-scale detection algorithm FPN, which combines low-level feature information with high resolution and high-level feature information with high semantics; by introducing Cascade Mask R-CNN model The efficient attention module maintains similar accuracy results on the basis of simplifying the computation of the Non-Local attention mechanism, and strengthens the original features by aggregating the same features in each position of the feature map. Compared with the existing two-stage target detection model and the original Cascade Mask R-CNN model, the invention significantly improves the detection accuracy of weld defects.

Description of drawings

Figure 1 is a flowchart of the weld defect detection method based on the Cascade Mask R-CNN model;

Figure 2 is the digital image defect part of the collected weld radiographic inspection;

Fig. 3 is the structural representation that utilizes FPN to improve RPN;

Figure 4 is a schematic diagram of an efficient attention module;

Figure 5 is a schematic structural diagram of the Cascade Mask R-CNN model;

Figure 6 is the correlation curve between the loss function value and the number of iterations in the training process.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

A welding seam defect detection method based on the Cascade Mask R-CNN model. First, the digital images of weld ray inspection and their corresponding annotation files containing five kinds of defects, including round defects, strip defects, cracks, lack of penetration and lack of fusion, are divided into It is a training set, a validation set and a test set; and the training set after image preprocessing is used to train the weld defect detection model built and optimized based on the Cascade Mask R-CNN model, and the validation set is used to obtain the optimal weight file. The test set was tested to evaluate the performance of the weld defect detection model. The method of welding seam ray detection based on Cascade Mask R-CNN model is shown in Figure 1, which includes the following steps:

S1. Acquire a digital image of the weld seam radiographic inspection containing five kinds of defects, such as round defect, strip defect, crack, lack of penetration, and lack of fusion, and its corresponding annotation file; The file is divided into training set, validation set and test set;

In the example, a total of 2600 digital images of metal welds were selected for radiographic inspection. The pixel size of the weld images was 400 × 3050, and there were 5 types of defects, namely round defects, strip defects, cracks, incomplete penetration and incomplete fusion. The various types of defects in the digital image of the weld seam ray detection are shown in Figure 2. All digital images of the weld seam ray detection and the corresponding annotation files are randomly divided into training set and test set according to the ratio of 9:1, and then from the training set Randomly divide 10% of the data as the validation set.

S2. Perform image preprocessing on the digital image of the welding seam ray detection to obtain an enhanced and unified image; the image preprocessing method includes image enhancement and image denoising, and the image enhancement adopts the AHE algorithm to perform the digital image of the welding seam ray detection. Image enhancement is performed to sharpen the details of the digital image of the weld ray detection to highlight the defect features. The calculation formula of the image enhancement is as follows:

In the above formula: y _i,j represents the center pixel before transformation, Y _i,j represents the center pixel after transformation, m _i,j represents the gray mean value of the local area with y _i,j as the center point, T represents the pair The cumulative distribution transformation function of the point, k represents the adaptive function, which is calculated from the pixel characteristics of the local area;

In the image denoising, the DMB algorithm is used to denoise the digital image of the weld seam ray inspection, so as to reduce the image noise and preserve and enhance the defect features.

S3. Build and optimize a weld defect detection model based on the original Cascade Mask R-CNN model. The weld defect detection model includes a convolutional neural network for feature extraction and a prediction network for classification and regression;

The convolutional neural network is ResNeXt-101, and the convolutional neural network includes a convolutional layer, a pooling layer and an activation layer, and the convolutional layer extracts features from the inputted image to generate a feature map; the pooling layer uses To remove redundant information, reduce the amount of parameters, and expand the receptive field; the activation layer increases the nonlinearity of the output, and uses the activation function to convolve the results of the output layer to obtain a nonlinear mapping.

The network structure of the ResNeXt-101 is shown in Table 1:

Table 1: Network structure of ResNeXt-101

S4. Use the training set to train the weld defect detection model, and use the verification set to verify the weight file obtained in each round of training, and obtain the weight file with the best performance on the verification set; including the following steps :

S41. Perform feature extraction on the digital image of the weld ray detection through the convolutional neural network to form a feature map;

S42, using the multi-scale detection algorithm FPN to improve the RPN to generate a suggested candidate frame for the feature map, and the horizontal connection in the FPN fuses low-level feature information with high resolution and high-level feature information with high semantics;

The embodiment utilizes the multi-scale detection algorithm FPN to improve the RPN to generate a suggested candidate frame for the feature map, including the following steps:

S421: Feed forward a part of ResNeXt-101 in the forward process of the network, and record the output of the last residual block of each stage of ResNeXt-101 as {C1, C2, C3, C4, C5}, as shown in Figure 3, first It is a bottom-up process. Each level is up-sampling with a set step size, and the layers that do not change the size of the feature map are grouped into a stage to form a feature pyramid;

In order to strengthen the original features, an efficient attention module is introduced into {C3, C4, C5} of ResNeXt-101, as shown in Figure 4, firstly, the 1×1 convolution W _k and softmax are used to obtain the attention weight, and the attention Force pooling obtains global contextual features, followed by a 1×1 convolution W _v1 and layer normalization, activated by the ReLU function, and a 1×1 convolution W _v2 to get the importance of each channel, Finally, the global context features are aggregated to the features of each position by addition to form long-distance dependencies; the calculation formula of the attention module is as follows:

表示全局注意力池化的权重，LN代表层归一化，W_v2ReLU(LN(W_v1(·)))表示计算每个通道的重要程度。In the above formula: z _i represents the input of the attention module, Z _i represents the output of the attention module, N _p represents the number of positions in the feature map,

represents the weight of global attention pooling, LN represents layer normalization, and W _v2 ReLU(LN(W _v1 ( ))) represents the importance of calculating each channel.

S422: Upsampling is used from top to bottom to enlarge the small feature map of the top layer to the same size as the feature map of the previous stage;

S423: The horizontal connection fuses the upsampling result and the feature map of the same size generated from the bottom to the top, and convolves each fusion result through a 3×3 convolution kernel to obtain the final feature layer P={P2, P3,P4,P5}.

After obtaining the feature layer P that has undergone feature fusion, use the K-means clustering algorithm to count the area and aspect ratio of the rectangular annotation box in the annotation file, and set five areas {32 ² ,64 ² ,128 ² ,256 ² ,512 ² The anchors of } correspond to the five feature layers of {P2, P3, P4, P5, P6}, of which the P6 feature layer is obtained by downsampling the P5 feature layer, and the anchor of each area is set to seven aspect ratios {1:10 ,1:5,1:2,1:1,2:1,5:1,10:1}, the generated anchor slides and traverses on the feature layer, and then generates the proposed candidate frame.

S43: Map the proposed candidate frame to the feature map to obtain a corresponding feature matrix, and uniformly scale the feature matrix to a specified size through ROI Align;

S44: Pass the proposed candidate frame through a three-stage cascade detector composed of a target classifier and a bounding box regressor, as shown in Figure 5, to obtain corresponding defect categories and bounding box regression parameters under three set thresholds; Finally, the final detection result is obtained by non-maximum suppression and filtering out low-probability targets;

The bounding box regressor in the three-stage cascade detector is defined as a cascade regression problem. The cascade regression changes the distribution of samples to be processed in different stages through resampling. The bounding box regressor is defined as follows:

In the above formula: x represents the sub-image block, b represents the sample distribution, f represents the bounding box regressor, the thresholds set by f ₁ , _{f 2} , and _{f 3} are 0.4, 0.5, and 0.6, respectively. The input, the output of f ₂ is used as the input of f ₃ , {f ₁ , f ₂ , f ₃ } is optimized for the resampling distribution of different stages, acting on both training and testing stages.

The total loss function L that defines the training model includes two parts: bounding box regression loss and target classification loss. The calculation formula of the total loss function is as follows:

L(x ^t ,g)=L _cls (h _t (x ^t ),X ^t )+μ[X ^t ≥1]L _loc (f _t (x ^t ,b ^t ),g)

In the above formula: L _cls represents the loss function of target classification, L _loc represents the loss function of bounding box regression, {b ^t } represents the sample distribution of different training stages t and has b ^t =f _t-1 (x ^t-1 , b ^t-1 ), h _t represents the target classifier, f _t represents the bounding box regressor, g represents the real target box parameter corresponding to x ^t , μ represents the compromise coefficient, X ^t represents the label corresponding to x ^t , [ ] Represents an indicator function.

S45: Set the training parameters of the weld defect detection model: the initial learning rate is set to 0.00125, the momentum is set to 0.9, the weight decay is set to 1e-4, the value of batch_size is set to 1, and the total number of training rounds is set to 40; The ImageNet dataset training weight is used as the pre-training weight of the weld defect detection model. The digital image of weld ray detection is uniformly scaled to 350×2600 before entering the weld defect detection model, and the stochastic gradient descent method is used to update the weld. parameters of the defect detection model; use the training set to train the weld defect detection model through the steps S41-S44, and use the verification set to verify the weight files obtained in each round of training, and obtain the Figure 6 shows the correlation curve between the loss function value and the number of iterations during the training process of the weld defect detection model.

S5. Use the test set to test the weight file with the best performance on the verification set obtained by applying step S45, so as to evaluate the performance of the weld defect detection model.

In this embodiment, the training set and the test set are used to train and test the weld defect detection model and other two-stage target detection models respectively, and introduce the average precision (AP) of each type of defect and the AP of each type of defect. The average value (mAP) was used as the evaluation index in this example. The experimental results and comparisons are shown in Table 2:

Table 2: Comparison of model experimental results (IoU=0.5)

From the analysis of the above experimental results, it can be concluded that compared with the traditional two-stage target detection model, the weld defect detection model described in this embodiment has the advantages of five defects detection of crack, strip defect, round defect, incomplete penetration and lack of fusion. The AP has been significantly improved, and the final mAP is also higher than other models.

By adopting the above-mentioned technical scheme disclosed by the present invention, the following beneficial effects are obtained:

The invention discloses a welding seam defect detection method based on the Cascade Mask R-CNN model. By introducing a three-level cascade structure, the problem of low accuracy of a target detection model based on a single threshold value and the overfitting caused by directly increasing the threshold value are improved The RPN is improved by the multi-scale detection algorithm FPN, which combines the low-level feature information with high resolution and the high-level feature information with high semantics; by introducing an efficient attention module into the Cascade Mask R-CNN model, it is easier to simplify Based on the calculation amount of the Non-Local attention mechanism, it maintains similar accuracy results, and strengthens the original features by aggregating the same features in each position of the feature map. Compared with the existing two-stage target detection model and the original Cascade Mask R-CNN model, the invention significantly improves the detection accuracy of weld defects.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (8)

1. A weld defect detection method based on a Cascade Mask R-CNN model is characterized by comprising the following steps:

s1, acquiring a welding seam radiographic inspection digital image containing five defects of round defect, strip defect, crack, incomplete penetration and incomplete fusion and a corresponding label file thereof; dividing the welding seam ray detection digital image and the corresponding label file into a training set, a verification set and a test set;

s2, carrying out image preprocessing on the welding line ray detection digital image to obtain an enhanced and unified image;

s3, building and optimizing a weld defect detection model based on a Cascade Mask R-CNN model, wherein the weld defect detection model comprises a convolutional neural network for feature extraction and a prediction network for classification and regression;

s4, training the weld defect detection model by using the training set, and verifying the weight file obtained by each training by using the verification set to obtain the weight file which is optimal to be expressed on the verification set; the method comprises the following steps:

s41, performing feature extraction on the welding line ray detection digital image through the convolutional neural network to form a feature map;

s42, improving the RPN by utilizing a multi-scale detection algorithm (FPN) to generate a suggested candidate box for the feature map;

s43, mapping the suggested candidate box to the feature diagram to obtain a corresponding feature matrix, and uniformly scaling the feature matrix to a specified size through ROIAlign;

s44, obtaining corresponding defect categories and boundary box regression parameters under three set thresholds through a three-level cascade detector consisting of a target classifier and a boundary box regressor by the suggested candidate box; finally, a final detection result is obtained by non-maximum value inhibition and low probability target filtering;

s45, setting training parameters of the weld defect detection model, training the weld defect detection model through the steps S41-S44 by adopting the training set, verifying the weight file obtained by each training by adopting the verification set, and acquiring the weight file which is optimal to be expressed on the verification set;

and S5, testing the weight file which is obtained by applying the step S45 and has the optimal performance on the verification set by using the test set, so as to evaluate the performance of the weld defect detection model.

2. The Cascade Mask R-CNN model-based weld defect detection method of claim 1, wherein the image preprocessing method in step S2 comprises image enhancement and image denoising, wherein the image enhancement adopts an AHE algorithm to perform image enhancement on the weld ray detection digital image, the weld ray detection digital image is subjected to detail sharpening to highlight defect features, and the image enhancement has the following calculation formula:

in the above formula: y is_i,jRepresenting the central pixel before transformation, Y_i,jRepresenting the transformed center pixel, m_i,jIs expressed as y_i,jThe gray level mean value of a local area of a central point, T represents a cumulative distribution transformation function of the point, k represents an adaptive function and is obtained by pixel characterization calculation of the local area;

the image denoising adopts a DMB algorithm to perform image denoising on the welding line ray detection digital image, so that the image noise is reduced, and the defect characteristics are reserved and enhanced.

3. The Cascade Mask R-CNN model-based weld defect detection method according to claim 1, characterized in that the convolutional neural network in step S41 is ResNeXt-101, the convolutional neural network comprises a convolutional layer, a pooling layer and an activation layer, and the convolutional layer extracts features from the input image to generate a feature map; the pooling layer is used for removing redundant information, reducing the number of parameters and expanding an acceptance domain; the activation layer increases the output nonlinearity, and the result of the output layer is convolved by using the activation function to obtain nonlinear mapping.

4. The Cascade Mask R-CNN model-based weld defect detection method according to claim 1, wherein in step S42, a proposed candidate frame is generated for the feature map by using FPN to improve RPN, and the method comprises the following steps:

s421: feeding forward a part of ResNeXt-101 in the forward process of the convolutional neural network, and recording the output of the last residual block of each stage of ResNeXt-101 as { C1, C2, C3, C4 and C5}, firstly, performing a bottom-up process, and reducing the sampling of a set step length upwards at each stage to form a stage without changing the size of a feature map so as to form a feature pyramid;

s422: the small feature map of the top layer is enlarged to the same size as the feature map of the last stage from top to bottom in an up-sampling mode;

s423: the cross-concatenation fuses the upsampled results with feature maps of the same size generated from bottom to top, and convolves each fused result with a 3 × 3 convolution kernel to obtain the final feature level P ═ P2, P3, P4, P5 }.

5. The Cascade Mask R-CNN model-based weld defect detection method according to claim 4, wherein the step S421 further comprises: a high efficiency attention module was introduced in { C3, C4, C5} of ResNeXt-101, first using a 1 × 1 convolution W_kAnd softmax obtains attention weight, acquires global context features by attention pooling, and then by a 1 × 1 convolution W_v1After normalization of the sum layer, activation by the ReLU function, and a further 1 × 1 convolution W_v2Obtaining the importance degree of each channel, and finally adding the obtained data to the original dataThe local context features are aggregated to the features of each position to form a long-distance dependency relationship; the attention module calculation formula is as follows:

in the above formula: z is a radical of_iIndicating the input of the attention module, Z_iIndicating the output of the attention module, N_pExpressed as a number of locations in the feature map,

weights representing global attention pooling, LN stands for layer normalization, W_v2ReLU(LN(W_v1(.))) represents the degree of importance of computing each channel.

6. The Cascade Mask R-CNN model-based weld defect detection method according to claim 4, further comprising, after the step S423: the area and the length-width ratio of a rectangular marking box in the marking file are counted through a K-means clustering algorithm, and five areas {32 } are set²,64²,128²,256²,512²The anchors of the areas correspond to five feature layers of { P2, P3, P4, P5 and P6} respectively, wherein the P6 feature layer is obtained by downsampling the P5 feature layer, seven aspect ratios of {1:10,1:5,1:2,1:1,2:1,5:1 and 10:1} are set for the anchors of each area, and the generated anchors are traversed on the feature layers in a sliding mode to generate the suggested candidate frames.

7. The Cascade Mask R-CNN model-based weld defect detection method of claim 1, wherein a boundary box regressor in the three-level Cascade detector in the step S44 is defined as a Cascade regression problem, the Cascade regression changes the sample distribution to be processed in different stages through resampling, and the boundary box regressor is defined as follows:

in the above formula: x denotes the subimage block, b denotes the sample distribution, f denotes the bounding box regressor, f₁、f₂、f₃The set threshold values are 0.4, 0.5, 0.6, respectively, f₁As output of f₂Input of f₂As output of f₃Input of { f }₁,f₂,f₃Optimizing the resampling distribution of different stages, and simultaneously acting on training and testing stages;

the total loss function L defining the training model includes two parts: the total loss function calculation formula comprises a bounding box regression loss and a target classification loss, wherein the total loss function calculation formula comprises the following steps:

L(x^t,g)＝L_cls(h_t(x^t),X^t)+μ[X^t≥1]L_loc(f_t(x^t,b^t),g)

in the above formula: l is_clsLoss function, L, representing the classification of the object_locLoss function representing regression of bounding box, { b^tDenotes the distribution of samples for different training phases t and has b^t＝f_t-1(x^t-1,b^t-1)，h_tRepresenting object classifier, f_tRepresenting a bounding box regressor, g representing the correspondence x^tMu represents the compromise coefficient, X^tDenotes x^tCorresponding label, [ ·]An indicator function is represented.

8. The Cascade Mask R-CNN model-based weld defect detection method according to claim 1, wherein the training parameter settings in step S45 include learning rate, momentum, weight attenuation, value of batch _ size and total number of training rounds, the training weights are used as pre-training weights of the weld defect detection model, the weld ray detection digital image is uniformly scaled before entering the weld defect detection model, and parameters of the weld defect detection model are updated by using a random gradient descent method.

Images